Deep-Learning-Based Prediction of Residential Floor Plan Attributes for Elderly-Oriented Housing Assessment

Summary

This protocol describes a meta-ensemble deep learning workflow for automated analysis of residential floor plan images. The procedure includes dataset preprocessing, architectural attribute prediction using multiple convolutional neural networks, stacking-based ensemble learning, rule-based residential classification, and explainability analysis using SHapley Additive exPlanations to support residential layout assessment.

Abstract

The growing demand for smart housing solutions for aging populations has increased the need for automated methods for residential floor plan analysis. Conventional architectural evaluation approaches are often manual, time-consuming, and difficult to scale, highlighting the need for artificial intelligence-based techniques that can interpret spatial layouts and support residential design assessment. This protocol describes a meta-ensemble deep learning framework for automated analysis of residential floor plan images. The proposed CARE-MIRV-Net framework integrates four complementary convolutional neural networks—MobileNetV2, InceptionV3, ResNet101, and VGG16—to predict architectural attributes, including square footage, number of bedrooms, bathrooms, and garages. Predictions from the base models are combined through a stacking-based meta-ensemble model and subsequently used within a rule-based decision layer to categorize residential layouts as elderly care, medical care, or general residential. Experimental evaluation demonstrated robust predictive performance across multiple target variables. The proposed model achieved a mean absolute error (MAE) of 432.48 and a coefficient of determination (R²) of 0.7053 for square footage prediction. For bathroom prediction, the framework achieved R² of 0.7605, while garage prediction yielded the lowest MAE of 0.1697 and the highest R² of 0.6958. Qualitative analyses further demonstrated the ability of the framework to generate architectural attribute predictions and support application-oriented residential layout assessment. To enhance transparency and interpretability, SHapley Additive exPlanations were incorporated to quantify the contribution of each base model to the final predictions. The proposed framework provides an interpretable and scalable approach for residential floor plan analysis and decision support.

Introduction

The rapid development of smart city technologies and smart living environments has created a growing demand for automated and data-driven solutions in residential design. In particular, the increasing aging population and the demand for healthcare-oriented housing have heightened the importance of designing residential spaces that are both functional and tailored to the needs of specific user groups^1,2. Conventional architectural floor plan analysis methods are often manual, time-consuming, and dependent on expert interpretation; therefore, they are inefficient and difficult to scale. This has created a strong need for intelligent systems capable of automatically analyzing residential layouts and supporting informed decision-making³.

Recent advances in artificial intelligence (AI), particularly in deep learning (DL) and computer vision, have significantly improved the analysis of image-based data. Convolutional neural networks (CNNs) have proven highly effective for spatial feature extraction and the learning of complex visual patterns, making them suitable for floor plan image analysis^4,5,6. These models can learn implicit representations of room structures, layout organization, and spatial relationships that are essential for predicting architectural attributes^7,8. However, reliance on a single DL model may limit generalization because different architectures may focus on distinct spatial features and exhibit biases arising from architectural variability⁹. Despite these advances, existing approaches often lack an integrated framework that combines multiple models, supports multioutput prediction, and provides application-level interpretability. Furthermore, healthcare-oriented residential assessment and explainability have received relatively limited attention within floor plan analysis systems. The absence of such comprehensive frameworks limits practical applicability and interpretability, particularly in domains where transparency and decision support are important^10,11,12.

AI has also contributed substantially to the automated analysis and generation of architectural floor plans. One study investigated the use of machine learning (ML) to transform two-dimensional floor plan images into three-dimensional models¹³. The authors highlighted the limitations of conventional computer-aided design (CAD)-based methods and emphasized the importance of scalable AI-based pipelines for spatial understanding. Their findings demonstrated that DL can effectively extract structural information from floor plans, although challenges related to generalization and data variability remain. Another study proposed a multiagent deep reinforcement learning framework for automated floor plan generation¹⁴. Although the approach offers considerable flexibility and customization for architectural design, it is primarily focused on design generation rather than predictive analysis of existing floor plans.

A comprehensive floor plan retrieval survey examined semantic, spatial, shape-based, and texture-based features for floor plan understanding¹⁵. The survey emphasized the importance of DL models, including CNNs and transformer-based architectures, for learning meaningful representations. However, it also identified challenges related to noise, distortion, and feature discrimination in complex layouts. Similarly, an ML-based approach incorporated topological and visual relationships for floor plan recognition⁸. The study demonstrated that visual accessibility and spatial connectivity are important for understanding architectural layouts, but it did not address multioutput prediction or ensemble-learning strategies. Generative AI has also been explored in architectural applications. A review of generative AI in architecture examined the use of generative adversarial networks (GANs), variational autoencoders (VAEs), and diffusion models for design automation¹⁶. The authors highlighted the increasing adoption of AI tools in practical architectural workflows. However, the focus of that work was design generation rather than predictive modeling or decision-support applications. Explainable AI has become increasingly important for interpreting DL models. A study proposed a SHapley Additive exPlanations (SHAP)-based framework for interpreting DL models in urban simulation applications¹⁷. The results demonstrated that SHAP can identify influential features and improve interpretability at both the global and local levels. Although the study supports the value of SHAP in spatial applications, it did not investigate its integration with ensemble-learning frameworks. Recent developments in DL architectures have further improved performance in image-based tasks. A study presented a comprehensive review of contemporary DL models including CNN-based and transformer-based architectures¹⁸. The study highlighted the importance of model selection and hybrid architectures for improving generalization performance. However, it did not specifically address floor plan analysis applications. Large-scale datasets have also facilitated advances in architectural analysis. For example, a study introduced a benchmark dataset containing thousands of complex floor plans¹⁹. The study demonstrated that existing models continue to face challenges when analyzing large-scale and multiapartment layouts, highlighting the need for more robust and generalizable approaches.

DL has also been extensively studied in ensemble learning. A recent study on deep ensemble methods indicated that combining multiple models can significantly improve predictive accuracy and robustness by reducing bias and variance²⁰. Despite their effectiveness, these methods have not been widely applied to architectural floor plan analysis, particularly in multioutput regression tasks. Furthermore, recent studies in automated floor plan analysis have highlighted the importance of integrated frameworks that combine prediction, classification, and interpretability^21,22,23. Existing research typically focuses on individual functions, such as feature extraction, layout generation, or classification. Consequently, a clear research gap remains in the development of a unified framework that integrates multimodel learning, residential-layout assessment, and explainable AI. Although considerable advances have been made in DL and architectural analysis, several research gaps remain. Most existing studies focus on floor plan recognition, feature extraction, or layout generation without providing a unified framework that supports both multioutput prediction and application-level decision-making. In addition, many approaches rely on a single DL model, which may limit their ability to capture diverse spatial characteristics and may introduce model-specific biases. Ensemble learning for floor plan analysis remains relatively underexplored, particularly regarding the integration of heterogeneous models to improve robustness. Furthermore, limited attention has been given to residential classification with an emphasis on elderly-oriented housing considerations and interpretability. Since many DL models function as black boxes, understanding the basis of their predictions remains challenging. These limitations highlight the need for a comprehensive framework that integrates multimodel learning, accurate prediction, application-oriented residential classification, and explainable AI.

This study presents an automated framework for the analysis of residential floor plan images, with an emphasis on spatial awareness and residential-layout assessment, to address these challenges. The proposed framework introduces a meta-ensemble DL model, CARE-MIRV-Net, which integrates four complementary CNNs—MobileNetV2, InceptionV3, ResNet101, and VGG16—within a stacking-based architecture^24,25,26,27. Each model independently predicts key architectural attributes, including square footage, number of bedrooms, bathrooms, and garages. These predictions are subsequently combined by a meta-learner to generate refined predictions. In addition, a rule-based decision layer is used to categorize residential layouts as elderly care, medical care, or general residential based on the predicted attributes.

The contributions of this work are as follows. First, a meta-ensemble DL framework is proposed for the multioutput prediction of architectural attributes from residential floor plan images. Second, the CARE-MIRV-Net model integrates four CNN architectures to improve prediction accuracy and generalization. Third, a rule-based decision layer is incorporated to support application-oriented residential layout categorization based on predicted architectural attributes. Fourth, SHAP-based explainable AI is integrated to enhance model transparency and interpretability. Finally, extensive experiments are conducted to evaluate the proposed framework using mean absolute error (MAE) and the coefficient of determination (R²) as performance metrics.

Protocol

This study used a publicly available dataset from the Kaggle platform (Floor Plan Images and Their Details, available at: https://www.kaggle.com/datasets/adilmohammed/floor-plan-images-and-their-details; accessed April 16, 2026). It did not involve human participants, animals, or personally identifiable data; therefore, ethical approval and informed consent were not required.

This protocol presents an AI-driven framework for automated analysis of residential floor plan images, with a focus on spatially aware and healthcare-oriented design. The complete experimental workflow is presented in Figure 1. The protocol is designed to bridge the gap between low-level architectural feature prediction and high-level application-oriented decision-making. It integrates DL, ensemble modeling, and explainable AI to provide an accurate, robust, and interpretable solution for analyzing residential layouts. The framework consists of multiple stages, beginning with data acquisition and preprocessing, followed by feature learning using deep CNNs, meta-ensemble prediction, and finally, application-level classification. Initially, floor plan images and their corresponding architectural attributes are preprocessed through resizing, normalization, and data augmentation to ensure consistency and improve generalization. All floor-plan images were resized to 224 × 224 pixels and normalized to the range [0, 1]. Data augmentation was applied during training using random horizontal flipping, rotation, zooming, and width/height shifting to improve model generalization and reduce overfitting. No vertical flipping was applied. Complete augmentation parameters and implementation details are provided in Supplementary File 1. The processed data were then used to train multiple DL models for feature extraction and prediction. All base models and the CARE-MIRV-Net meta-learner were trained using a consistent configuration, including data augmentation, early stopping, and adaptive learning-rate scheduling. Early stopping monitored validation loss (val_loss) with a patience value of 5 epochs, and the best model weights were restored automatically after training. Adaptive learning-rate scheduling was implemented using the ReduceLROnPlateau callback, which monitored validation loss and reduced the learning rate by a factor of 0.3 after 2 epochs without improvement, with a minimum learning rate of 1 × 10⁻⁷. Complete implementation details and parameter settings are provided in Supplementary File 1. The dataset was randomly divided into training and testing sets using an 80:20 split ratio. A fixed random seed (42) was used to ensure reproducibility. Complete implementation details are provided in Supplementary File 1.

Figure 1. Workflow of the CARE-MIRV-Net framework for residential floor plan analysis. Schematic overview of the proposed protocol. The workflow includes (1) dataset acquisition and preprocessing, (2) feature learning and prediction using MobileNetV2, InceptionV3, ResNet101, and VGG16, (3) stacking-based meta-ensemble prediction using CARE-MIRV-Net, (4) rule-based residential layout classification, (5) SHapley Additive exPlanations (SHAP)-based interpretability analysis, and (6) performance evaluation. Please click here to view a larger version of this figure.

In the second stage, four pretrained CNNs, MobileNetV2, InceptionV3, ResNet101, and VGG16, are employed as base learners. Each model is fine-tuned using transfer learning to predict key architectural attributes, including square footage, number of bedrooms, bathrooms, and garages. For each transfer-learning model, the lower pretrained layers were kept frozen, while the upper layers and the custom regression head were fine-tuned during training. Detailed model-specific fine-tuning configurations are provided in Supplementary File 1. These models capture different aspects of spatial information: MobileNetV2 provides efficient feature extraction, InceptionV3 captures multiscale spatial patterns, ResNet101 learns deep hierarchical representations, and VGG16 ensures stable and consistent feature learning. The diversity among these models enhances the overall representational capability of the framework.

To enhance the accuracy and robustness of prediction further, a stacking-based meta-ensemble model, denoted as CARE-MIRV-Net, is proposed. At this step, predictions from all base models are pooled together to create a higher-dimensional feature space, which is input to a meta-learner. Base-model predictions are first generated independently using the trained MobileNetV2, InceptionV3, ResNet101, and VGG16 models. These predictions are then concatenated to form the meta-feature vector used by the meta-learner. To prevent data leakage, the training and testing datasets are strictly separated, and the meta-learner is trained only on predictions generated from the training data. The meta-model, which is a fully connected neural network, is trained to learn the optimal combination of base-model outputs to produce refined predictions. The meta-model consists of two fully connected layers with 512 and 256 neurons using ReLU activation, followed by dropout layers (0.4 and 0.3) and a linear output layer. The model is trained using the Adam optimizer (learning rate = 0.0001), a batch size of 32, and up to 15 epochs. Early stopping was implemented by monitoring validation loss (val_loss) with a patience value of 5 epochs. The minimum improvement criterion (min_delta) was set to the TensorFlow default value of 0, and the best model weights were restored automatically after training. Adaptive learning-rate scheduling was implemented using the ReduceLROnPlateau callback, which monitored validation loss and reduced the learning rate by a factor of 0.3 after 2 consecutive epochs without improvement. The minimum learning rate was set to 1 × 10⁻⁷, and the cooldown parameter used the default TensorFlow value of 0. This ensemble method reduces individual model biases and improves generalization through the complementary strengths of multiple architectures. Detailed implementation information is provided in Supplementary File 1.

After prediction, an interpretation layer is added to convert the numerical outputs to application-level decisions. Residential layouts can be classified into three classes based on the predictions of architectural characteristics: elderly care, healthcare-integrated (medical care), and general residential use. This classification is performed using predefined threshold-based decision rules based on the predicted square footage, number of bedrooms, and number of bathrooms. The category with the highest score is assigned as the final classification, whereas ties are resolved by selecting the category satisfying the greater number of decision criteria. The complete scoring rules and classification thresholds are provided in Supplementary File 2 (Algorithm 1). This action allows the framework to provide practical information for residential housing design.

The framework uses explainable AI with SHAP to provide transparency and interpretability. This component examines the contribution of the predictions from each base model to the output of the meta-model. SHAP analysis was performed using SHAP KernelExplainer (SHAP version 0.51.0) with 100 randomly selected training samples as the background dataset and 50 test samples for explanation generation. The complete SHAP configuration, including background sample selection and implementation settings, is provided in Supplementary File 1. The explainability layer helps to understand the decision-making process by measuring feature importance and displaying contribution patterns, thereby improving model transparency and reliability. The proposed framework is evaluated using MAE and R² across all target variables. For multioutput prediction, MAE was computed as the average absolute difference between actual and predicted values across all target variables, whereas R² was calculated by comparing the explained variance of the predictions with the corresponding ground-truth values for each output. The quantitative and qualitative analyses indicate that the proposed CARE-MIRV-Net improves prediction accuracy and supports rule-based residential layout classification. Classification reliability was assessed based on the accuracy of the underlying regression predictions and the consistency of the rule-based decision layer. The framework is a suitable and scalable solution for intelligent floor plan analysis and can be applied to smart housing, elderly-care planning, and healthcare-oriented residential design. A complete list of datasets, software packages, libraries, hardware resources, and computational tools used in this study is provided in the Table of Materials. Source code, environment configuration information, software dependency specifications, and implementation scripts required to reproduce the reported workflow are provided in Supplementary File 1.

Algorithm for the Protocol

The proposed framework takes a multistage approach to processing image-based floor plans of residential houses and producing predictive and decision-level outputs, as depicted in Algorithm 1 in Supplementary File 2. This process begins with data preprocessing, whereby the input images are resized to a standard resolution and normalized to provide uniformity throughout the dataset. All floor-plan images were resized to 224 × 224 pixels and normalized to the range [0, 1]. This step optimizes the images for learning by deep neural networks and enhances overall model convergence. The second phase involves using several DL models as base learners, including MobileNetV2, InceptionV3, ResNet101, and VGG16. Independently, each model takes the input images and estimates important architectural features such as square footage and the number of bedrooms, bathrooms, and garages. The dataset was randomly divided into training and testing sets using an 80:20 split ratio, and a fixed random seed was used to ensure reproducibility. These models capture different spatial properties of the floor plans and provide complementary predictions. The feature representation is then constructed by stacking the outputs of each base model to create a single feature representation. This joint output is subsequently fed into a meta-ensemble model, CARE-MIRV-Net, which is trained to combine the predictions of all base models. Base-model predictions were generated independently and concatenated to form the meta-feature vector. Data leakage was prevented through strict separation of the training and testing datasets. The meta-model refines these predictions and produces the final architectural attribute forecasts. The meta-model consisted of fully connected layers with 512 and 256 neurons, ReLU activation functions, dropout rates of 0.4 and 0.3, the Adam optimizer (learning rate = 0.0001), a batch size of 32, and up to 15 training epochs with early stopping. After prediction, a decision layer is used to contextualize the results. According to the estimated values, each residential layout is categorized as elderly care, medical care, or general residential use. Residential categories were assigned using predefined threshold-based scoring rules derived from the predicted floor-plan attributes. The category with the highest score was selected as the final classification. Complete classification thresholds and decision rules are provided in Algorithm 1 (Supplementary File 2). Lastly, the framework incorporates an explainability component based on SHAP to assess the impact of each base model on the final prediction. SHAP analysis was performed as described previously and in Supplementary File 1. This component increases transparency and aids in understanding the decision-making process. The effectiveness and viability of the proposed approach were assessed using standard regression performance measures. Performance was evaluated using a single experimental run with a fixed random seed.

Evaluation Models

In this section, the DL models used in the study are presented, including both the individual base models and the proposed CARE-MIRV-Net. The benchmark models were selected to represent diverse and widely adopted DL architectures. MobileNetV2 was included as a lightweight and computationally efficient network, and InceptionV3 was selected for its ability to capture multiscale spatial features. ResNet101 was chosen due to its deep residual learning capability for hierarchical feature extraction, and VGG16 was included as a well-established convolutional architecture. All benchmark models were trained using identical preprocessing procedures, data augmentation settings, dataset partitions, optimization parameters, batch size, and training configuration to ensure a fair comparative evaluation.

MobileNetV2:

MobileNetV2 is a small and compact CNN system that is engineered to be fast, high-performing, and with less computational complexity. It presents some of the major innovations like inverted residual connections and linear bottlenecks, which facilitate efficient feature extraction and still have a high representational power. MobileNetV2 can use depthwise separable convolutions to greatly minimize the total number of parameters and the computational cost of traditional convolutional networks, and it is thus well adapted to large image-based learning problems.

MobileNetV2 is used in the suggested framework as one of the foundation DL models to predict the architectural attributes of residential floor plan images. The model is able to learn spatial patterns like room distribution, layout density, and structural organization that occur in floor plans effectively because of its efficient design. Such learned representations play an important role in estimating important features such as square footage, number of bedrooms, bathrooms, and garages accurately. MobileNetV2 is a lightweight model, which is a good candidate to be included in the proposed meta-ensemble framework. It adds some complementary feature representations to other DL models, and it enhances the overall robustness and generalization of the system. This model learning is very important in terms of improving the accuracy of predictions and in assisting in the classification of residential designs based on healthcare.

The MobileNetV2 model is trained using a transfer-learning strategy with pretrained weights from the ImageNet dataset. The input images are scaled to a fixed size of 224 × 224 × 3, which is compatible with the architecture. The images were resized to 224 × 224 pixels, normalized to the range [0, 1], and augmented using random flipping, rotation, zooming, and shifting. A batch size of 32 was used during training. The initial classification head of MobileNetV2 is removed and replaced with a regression head. The regression head consisted of a Global Average Pooling layer followed by fully connected layers with 512 and 256 neurons, batch normalization, dropout layers (0.4 and 0.3), and a four-neuron linear output layer. To facilitate domain adaptation, the lower pretrained layers were frozen, whereas the upper layers beginning from block_13_expand and the custom regression head were fine-tuned. This selective training approach enables the model to retain general visual features while learning domain-specific spatial attributes of floor plan images. The extracted feature maps are processed through a Global Average Pooling layer and fully connected layers with 512 and 256 neurons using ReLU activation. To enhance generalization and reduce overfitting, batch normalization and dropout layers (0.4 and 0.3) are incorporated. The final output layer contains four neurons with linear activation corresponding to the predicted architectural attributes. The Adam optimizer was used with a learning rate of 0.0001 and TensorFlow/Keras default parameters (β₁ = 0.9, β₂ = 0.999, ε = 1 × 10⁻⁷, amsgrad = False). For multioutput prediction, MAE and R² were computed by comparing the predicted and actual values for each target variable and then averaging the results across all outputs. Early stopping was applied based on validation loss, and adaptive learning-rate reduction was used to improve convergence. Training was performed for up to 15 epochs.

InceptionV3:

InceptionV3 is a deep CNN architecture that is capable of capturing multiscale features with the use of parallel convolutional operations. It is founded on the Inception module, which uses a variety of filter sizes in the same layer to both extract fine-grained and coarse features at the same time. This type of architecture design allows the network to learn intricate spatial hierarchies and be computationally efficient. On top of this, InceptionV3 uses factorized convolutions and dimensionality reduction, which enhance performance and minimize computational cost.

InceptionV3 is applied in the proposed framework as an effective feature extractor to analyze residential floor plan images. Floor plans consist of a variety of spatial patterns, such as room sizes, structural plans, and connectivity, necessitating multiscale feature learning. Inception architecture is especially better suited to this task by the fact that it is able to represent local specifics (e.g., small rooms) and global structures (e.g., the organization of the overall layout) at the same time. InceptionV3 provides complementary representations to other models like MobileNetV2 in the context of the meta-ensemble framework. The diversity of predictions is increased by its capability to model complex spatial relationships, and this is important in enhancing ensemble performance. This eventually assists in better prediction of architectural characteristics and reinforces downstream classification of elderly-friendly and healthcare-integrated housing design.

The InceptionV3 model is trained using a transfer-learning approach with pretrained weights from the ImageNet dataset. The input images are scaled to 224 × 224 × 3 to ensure compatibility with the network architecture and consistency across all models within the framework. The images were resized to 224 × 224 pixels, normalized to the range [0, 1], and augmented using random flipping, rotation, zooming, and shifting. A batch size of 32 was used during training.

The original classification layers of InceptionV3 are removed, and a custom regression head is introduced to enable multioutput prediction. The custom regression head consisted of a Global Average Pooling layer followed by fully connected layers with 512 and 256 neurons, batch normalization, dropout layers (0.4 and 0.3), and a four-neuron linear output layer. The convolutional base is partially fine-tuned, with the lower pretrained layers frozen and the upper layers together with the custom regression head fine-tuned to learn domain-specific floor plan features while retaining general visual representations acquired during pretraining. After the convolutional backbone, a Global Average Pooling layer transforms the feature maps into a compact feature representation. This is followed by fully connected layers containing 512 and 256 neurons with ReLU activation functions. To minimize overfitting and improve generalization, dropout rates of 0.4 and 0.3 are applied. The final output layer consists of four neurons with linear activation corresponding to the prediction of square footage, number of bedrooms, number of bathrooms, and number of garages. The model is trained using the Adam optimizer with a learning rate of 0.0001, a batch size of 32, and up to 15 training epochs with early stopping and adaptive learning-rate reduction to ensure stable convergence.

ResNet101:

ResNet101 is a deep structure of the CNN that is constructed on the principle of residual learning that makes it possible to train extremely deep networks by overcoming the vanishing gradient problem. It adds residual connections, or so-called skip connections, which enable the network to learn identity mappings and to retain information across layers. This architecture greatly enhances the stability of training and allows the model to learn intricate hierarchical characteristics. ResNet101 has a deep representational capacity of 101 layers and is, therefore, very effective in tasks that require extra spatial features.

ResNet101 will be used in the proposed framework as a high-capacity feature extractor, which will analyze the images of residential floor plans. The model can be used to represent complex spatial relationships, including room connectivity, layout complexity, and structural variations, because of its deep architecture. These features are critical in making precise predictions on architectural features such as square footage, bedrooms, bathrooms, and garages. In the meta-ensemble design, ResNet101 is a significant component and provides deep and abstract feature representations. ResNet101 is more concerned with fine-grained structural features and thus more diverse among base learners compared to lightweight models such as MobileNetV2. This heterogeneity plays an essential role in enhancing the performance of an ensemble and robust predictions, especially when dealing with healthcare-based residential classification.

ResNet101 is trained using a transfer-learning approach with pretrained weights from the ImageNet dataset. The input images are scaled to 224 × 224 × 3, which ensures architectural compatibility and consistency across all models within the framework. The images were resized to 224 × 224 pixels, normalized to the range [0, 1], and augmented using random flipping, rotation, zooming, and shifting. A batch size of 32 was used during training. The original classification head of ResNet101 is removed (include_top=False), and a custom regression head is introduced to enable multioutput prediction. The regression head consisted of a Global Average Pooling layer followed by fully connected layers with 512 and 256 neurons, batch normalization, dropout layers (0.4 and 0.3), and a four-neuron linear output layer. To balance feature reuse and domain adaptation, the lower pretrained layers were frozen, whereas the deeper layers beginning from conv4_block1_1_conv and the custom regression head were fine-tuned. This selective training strategy enables the model to retain generic visual features while adapting higher-level representations to floor plan images. The output of the convolutional base is passed through a Global Average Pooling layer to generate a compact feature representation. This is followed by fully connected layers with 512 and 256 neurons using ReLU activation functions. Batch normalization is applied to stabilize learning, and dropout layers with rates of 0.4 and 0.3 are incorporated to reduce overfitting and improve generalization. The final output layer contains four neurons with linear activation corresponding to square footage, number of bedrooms, bathrooms, and garages. The Adam optimizer was used with a learning rate of 0.0001, as described in the training configuration. Data augmentation included random flipping, rotation, zooming, and shifting to improve model generalization. Training was performed for up to 15 epochs with early stopping based on validation loss and adaptive learning-rate reduction.

VGG16:

VGG16 is a simple neural network architecture of the deep convolutional type that is effective and simple in visual recognition tasks. It is made up of 16 layers with a regular or homogeneous design incorporating small 3 × 3 convolutional filters, which facilitate learning hierarchical feature representations by the network. The architecture is depth-focused and simplistic, which means it is capable of capturing low-level details like edges and textures and high-level semantic structures. VGG16, with its regular design and good performance, is now a popular choice of backbone when performing transfer learning in image-related tasks.

In the suggested model, VGG16 is used as a baseline DL model to derive spatial features of images of residential floor plans. The fact that it is organized enables it to be very useful in capturing the finer details of layout patterns like room boundaries, structural alignments, and spatial organization. The features are critical in effectively forecasting the attributes of architecture such as square feet, bedrooms, bathrooms, and garages. In the meta-ensemble system, VGG16 provides stable and well-generalized feature representations. VGG16 offers a more balanced feature extraction approach compared to deeper architectures, such as ResNet101, and multiscale architectures, such as InceptionV3, increasing base learner diversity. This heterogeneity is critical in enhancing the performance of the ensemble and reliable forecasts of healthcare-oriented residential classification.

The VGG16 model is trained using a transfer-learning approach with pretrained weights from the ImageNet dataset. The input images are scaled to 224 × 224 × 3 to ensure compatibility with the architecture and consistency across all models within the framework. The images were resized to 224 × 224 pixels, normalized to the range [0, 1], and augmented using random flipping, rotation, zooming, and shifting. A batch size of 32 was used during training. The original classification layers are removed (include_top=False), and a custom regression head is added to adapt the model for multioutput prediction. The regression head consisted of a Global Average Pooling layer followed by fully connected layers with 512 and 256 neurons, batch normalization, dropout layers (0.4 and 0.3), and a four-neuron linear output layer. The lower pretrained layers were frozen, whereas the layers beginning from block4_conv1 and the custom regression head were fine-tuned. This strategy enables the model to retain general visual characteristics while learning domain-specific representations relevant to floor plan analysis. The output features of the convolutional base are passed through a Global Average Pooling layer to generate a compact feature vector. This is followed by fully connected layers with 512 and 256 neurons using ReLU activation functions. Batch normalization and dropout layers with rates of 0.4 and 0.3 are incorporated to stabilize training, reduce overfitting, and improve generalization. The final output layer contains four neurons with linear activation corresponding to square footage, number of bedrooms, bathrooms, and garage predictions. The Adam optimizer was used with a learning rate of 0.0001, as described in the training configuration. Data augmentation included random flipping, rotation, zooming, and shifting to improve model generalization. Training was performed for up to 15 epochs with early stopping based on validation loss and adaptive learning-rate reduction.

CARE-MIRV-Net (Proposed Meta-Ensemble DL Model):

The suggested CARE-MIRV-Net (Context-Aware Residential Ensemble utilising MobileNetV2, InceptionV3, ResNet101, and VGG16 Network) is a stacking-based meta-ensemble DL framework that is aimed at improving the estimate of architectural features on the basis of residential floor plan pictures. The architecture combines various heterogeneous CNNs by exploiting their synergistic advantages in the extraction of features. The proposed model is based on the combination of lightweight, deep, and multiscale architecture that enhances predictive performance and provides robustness in a wide range of residential layouts. In contrast to traditional single-model-based techniques, CARE-MIRV-Net follows a hierarchical learning approach, where base models provide initial predictions, which are then improved by a meta-learner.

When it comes to residential design that is spatially conscious and health-conscious, the proper interpretation of floor plan layouts is of paramount importance. Single DL models tend to be insufficient in generalizing different architectural patterns. The proposed framework will overcome this weakness by using four different architectures: MobileNetV2, InceptionV3, ResNet101, and VGG16, which bring forth different representational abilities of their own. MobileNetV2 can be used to extract features efficiently and is lightweight; InceptionV3 learns spatial patterns across multiple scales; ResNet101 learns deep hierarchical representations; and VGG16 learns features consistently and reliably. A combination of these models facilitates the framework to embrace both the local and global spatial attributes, which enhances the accuracy and reliability of the predicted features like square footage, number of bedrooms, bathrooms, and garages. The above predictions can also be used to predict further classification of residential layouts to include elderly-friendly, healthcare-integrated, and general residential.

CARE-MIRV-Net is implemented using a two-stage stacking-based ensemble strategy. In the first stage, the four base models independently generate predictions for the target variables. Meta-learner inputs were generated by concatenating the prediction outputs produced by the trained base models on the training data. These concatenated predictions formed the 16-dimensional meta-feature vectors used for meta-model training. Information leakage was prevented through strict separation of the training and testing datasets, and no testing samples were used during either base-model or meta-model training. The same training and validation datasets used for the base models were employed during the stacking process. During both training and inference, the 16-dimensional meta-feature vector was constructed by concatenating the four prediction outputs generated by MobileNetV2, InceptionV3, ResNet101, and VGG16. Since each model generates a four-dimensional output vector, the concatenated representation produces a 16-dimensional input to the meta-learner. This transformation combines the predictions from all base models and serves as the input to the second stage of the framework. The second stage involves constructing a fully connected neural network meta-learner to learn the optimal combination of base-model predictions. The meta-learner consisted of fully connected layers with 512 and 256 neurons, ReLU activation, batch normalization, dropout rates of 0.4 and 0.3, and a four-neuron linear output layer. Dropout layers with rates of 0.4 and 0.3 were applied after the dense layers to reduce overfitting and improve generalization. The meta-model was trained using the Adam optimizer with a learning rate of 0.0001, a batch size of 32, and up to 15 epochs with early stopping and adaptive learning-rate reduction. Validation data were used to monitor model performance during training and to select the best-performing model. After training, the meta-ensemble model generated final predictions by combining outputs from all base learners. MAE and R² were used to evaluate the performance of CARE-MIRV-Net. For multioutput prediction, MAE and R² were computed by comparing the predicted and actual values for each target variable and then averaging the results across all outputs. Results were obtained from a single experimental run using a fixed random seed. The proposed framework improves predictive ability by combining several DL structures through a stacking mechanism and providing interpretable analysis of residential floor plans. This renders CARE-MIRV-Net applicable in the context of geriatric-focused and healthcare-oriented residential design. The source code, environment configuration information, software dependency specifications, and implementation documentation are provided in Supplementary File 1.

Performance Evaluation

Performance analysis of the proposed framework is carried out to determine its predictive power, robustness, and generalization performance across different architectural characteristics. Quantitative and qualitative analyses are conducted to confirm the usefulness of the proposed model. The assessment procedure involves both standard regression measures and controlled experimental environments to enable fair comparison with the baseline models. Experiments were conducted using a single experimental run with a fixed random seed to ensure reproducibility and consistency across all benchmark and proposed models.

Performance Parameters:

Two well-known regression measures, MAE and R², are used to assess the performance of the proposed framework. MAE measures the average magnitude of prediction errors and provides a clear indication of the proximity between predicted and actual values. Conversely, the R² value is used to assess the proportion of variance in the target variables explained by the model, reflecting its overall predictive power. A combination of these metrics provides a comprehensive evaluation of the accuracy and generalization performance of all predicted architectural features. For multioutput prediction, MAE and R² were computed by comparing the predicted and actual values for each target variable and then averaging the results across all outputs.

Experimental Setup:

The experiments were conducted in a cloud-computing environment using Python (version 3.12.12). The proposed protocol was implemented using TensorFlow (version 2.19.0), Keras (version 3.13.2), NumPy (version 2.4.6), Pandas (version 2.3.3), Scikit-learn (version 1.6.1), Matplotlib (version 3.9), and SHAP (version 0.51.0). The computational environment utilized an Intel Xeon processor operating at 2.20 GHz with approximately 31 GB of system memory. The experiments were conducted on an Ubuntu 22.04 LTS operating system utilizing NVIDIA Tesla T4 × 2 GPUs. The dataset contains 2,640 residential floor plan images that were preprocessed and subdivided into training and testing datasets. The dataset was divided using an 80:20 split ratio, resulting in 2,112 training samples and 528 testing samples. MobileNetV2, InceptionV3, ResNet101, and VGG16 are four DL models that were trained using a fine-tuning approach with transfer learning. A stacking-based meta-ensemble model, CARE-MIRV-Net, was then constructed to combine predictions from these base models. Each model was trained under uniform conditions, including image resizing to 224 × 224 pixels, data augmentation, and early stopping to reduce overfitting. Data augmentation included random flipping, rotation, zooming, and shifting. A batch size of 32 and a maximum of 15 training epochs were used, with early stopping based on validation loss and adaptive learning-rate reduction. The final analysis was carried out on a hidden test set to provide an unbiased and reliable assessment of performance. The hidden test set was generated during the initial dataset partitioning and remained completely isolated throughout model training and development to ensure unbiased performance evaluation.

Dataset Description:

The dataset used to support the findings of this study is publicly available on the Kaggle platform with the title “Floor Plan Images and Their Details”²⁸. The dataset is available at: https://www.kaggle.com/datasets/adilmohammed/floor-plan-images-and-their-details (accessed April 16, 2026). The data consist of 2,640 residential floor plan images and organized architectural metadata. Every sample is a distinct housing plan, which provides a wide range of residential layouts with different sizes, configurations, and spatial arrangements. The dataset size is sufficient for DL-based training and contains a variety of architectural patterns. The dataset consists of a two-dimensional floor plan image and related numerical attributes describing the structural characteristics of the residence. These features include total square footage, number of bedrooms, number of bathrooms, and number of garages. Moreover, every record contains an image path, which allows direct mapping of visual inputs to labels. The dataset variables include floor plan image, image path, square footage, number of bedrooms, number of bathrooms, and number of garages. The floor plan image serves as the input feature, while the architectural attributes are used as prediction targets. The dataset is therefore suitable for supervised learning, with images serving as input features and architectural attributes serving as regression targets.

The dataset contains a broad variety of residential layouts, ranging from compact layouts with fewer rooms to large multiroom layouts. This variability enables the learning of meaningful spatial representations, including room distribution, layout density, and structural complexity. This diversity is especially important for the proposed framework, as it supports the categorization of floor plans into elderly-friendly, healthcare-integrated, and general residential layouts. Before model training, the dataset is processed through a sequence of preprocessing procedures to ensure consistency and compatibility with DL models. All images were resized to 224 × 224 pixels and normalized to the range [0, 1] before training. The dataset is then organized by matching image paths with their corresponding metadata, after which it is separated into training and testing subsets to facilitate performance analysis. The dataset was randomly divided into training and testing sets using an 80:20 split ratio. Records containing complete image and metadata information were included in the analysis, whereas incomplete or invalid records were excluded. A fixed random seed was used during dataset partitioning to ensure reproducibility. The dataset provides a detailed basis for AI-driven analysis of residential floor plans. The combination of 2,640 annotated images and diverse architectural attributes supports multioutput regression modeling and enables the integration of spatial intelligence with healthcare-oriented residential design decision-making.

Results

A set of experiments was performed to evaluate the proposed framework using four DL models, MobileNetV2, InceptionV3, ResNet101, and VGG16, together with the proposed CARE-MIRV-Net meta-ensemble model.

Results of Dataset Acquisition and Preprocessing

The preprocessing phase was designed to prepare the raw floor plan images and related metadata for training and evaluation using DL models. The dataset was obtained from the Kaggle repository and contained 2,640 samples. Each sample consisted of a floor plan image and its corresponding architectural attributes. Only the relevant columns, including image path, square footage, bedrooms, baths, and garages, were retained for further analysis, and incomplete or blank records were removed to maintain data consistency and quality. To facilitate model training and evaluation, the dataset was divided into training and testing subsets at an 80:20 ratio, resulting in 2,112 training samples and 528 testing samples. A fixed random seed was used during dataset partitioning to ensure reproducibility. This division ensured that the models were trained on a sufficiently large dataset while preserving an independent test set for objective performance assessment. The image paths were standardized by mapping each record to its corresponding location within the dataset directory. The four architectural attributes (square feet, bedrooms, bathrooms, and garages) served as the target variables for the multioutput regression task. Although bedrooms, bathrooms, and garages are discrete variables, they were modeled using a multioutput regression framework to enable the simultaneous prediction of all architectural attributes within a unified learning architecture. This approach allows the model to capture shared relationships among the target variables and exploit correlations between continuous and discrete architectural characteristics. Predicted values for discrete attributes were converted to the nearest valid integer during post-processing before interpretation and downstream residential-layout classification.

Table 1 outlines the architectural features predicted by the proposed model, representing key spatial and functional characteristics of residential floor plans. Exploratory data analysis was performed to better understand the distribution of the target variables using histograms and count-based visualizations. The distribution of square footage is continuous with varying densities, whereas the distributions of bedrooms, bathrooms, and garages highlight the frequency of different residential configurations. These distributions are depicted in Figure 2. The distribution of square footage is shown in Figure 2A, whereas the distributions of bedrooms, bathrooms, and garages are shown in Figure 2B–2D, respectively. Together, these plots illustrate the variability and distribution of the target variables within the dataset.

Feature Name	Description
Square Feet	Total built-up area of the residential unit
Beds	Number of bedrooms in the layout
Baths	Number of bathrooms in the layout
Garages	Number of garage spaces available

Table 1: Architectural attributes used as prediction targets. Definitions of the architectural attributes extracted from the residential floor plan dataset and used as target variables for model training and evaluation. The attributes include total square footage, number of bedrooms, number of bathrooms, and number of garage spaces associated with each residential floor plan.

Figure 2. Distribution of architectural attributes in the residential floor plan dataset. Histograms showing the distributions of (A) square footage, (B) number of bedrooms, (C) number of bathrooms, and (D) number of garages across the dataset used for model development and evaluation. Please click here to view a larger version of this figure.

The image preprocessing stage transformed all floor plan images into a standardized format suitable for DL models. All images were loaded, converted to RGB color space, resized to 224 × 224 pixels, and normalized to the range of [0, 1]. This process ensured consistent input dimensions and improved model convergence during training. Preprocessed images were randomly sampled, and labels were verified to confirm proper preprocessing and label alignment. Following preprocessing, each image was converted into a NumPy array and paired with its corresponding target values to create the input-output dataset. The resulting datasets confirmed successful preparation of the data, with training images having dimensions of (2112, 224, 224, 3) and testing images having dimensions of (528, 224, 224, 3). The corresponding label matrices had dimensions of (2112, 4) and (528, 4), respectively. The target variables were converted to floating-point values to support regression-based learning.

To further improve model generalization and reduce overfitting, data augmentation techniques were applied using an image data generator. These techniques included random horizontal flipping, rotation, zooming, and width/height shifting to increase training-data variability and improve model generalization. These transformations increased variability within the training data and enabled the models to learn more robust and invariant spatial features. The preprocessing pipeline produced clean, standardized, and well-structured image and tabular data, providing a suitable foundation for training the proposed meta-ensemble deep learning framework.

Results of Base Models and the Proposed Model

The preprocessed dataset (2,640 floor plan images) was used to train all models using training and test splits. The dataset was divided using an 80:20 split ratio, resulting in 2,112 training samples and 528 testing samples. A fixed random seed was used to ensure reproducibility. Pretrained weights were used for transfer learning and fine-tuning to adapt the models to the domain of architectural floor plan analysis. Data augmentation methods were employed during training to enhance generalization, and early stopping was used to reduce overfitting by monitoring validation performance. The models were trained with a batch size of 32 for up to 15 epochs, with early stopping applied to prevent overfitting and promote efficient convergence. Early stopping monitored the validation loss and terminated training when no further improvement was observed, while adaptive learning-rate reduction was applied to improve convergence. Both loss and MAE were used to monitor the training process and evaluate convergence behavior and stability. All models were trained under the same experimental conditions to ensure a fair comparison. Figures 3–7 show the training and validation loss and MAE curves for all models, providing insight into their learning dynamics and generalization behavior.

Figure 3. Training performance of MobileNetV2. (A) Training and validation loss across training epochs. (B) Training and validation mean absolute error (MAE) across training epochs. Please click here to view a larger version of this figure.

Figure 4. Training performance of InceptionV3. (A) Training and validation loss across training epochs. (B) Training and validation mean absolute error (MAE) across training epochs. Please click here to view a larger version of this figure.

Figure 5. Training performance of ResNet101. (A) Training and validation loss across training epochs. (B) Training and validation mean absolute error (MAE) across training epochs. Please click here to view a larger version of this figure.

Figure 6. Training performance of VGG16. (A) Training and validation loss across training epochs. (B) Training and validation mean absolute error (MAE) across training epochs. Please click here to view a larger version of this figure.

Figure 7. Training performance of CARE-MIRV-Net. (A) Training and validation loss across training epochs. (B) Training and validation mean absolute error (MAE) across training epochs. Please click here to view a larger version of this figure.

The training performance of MobileNetV2 is shown in Figure 3. The training and validation loss curves are presented in Figure 3A, whereas the training and validation MAE curves are shown in Figure 3B. Training loss and MAE gradually decreased with increasing epochs, indicating stable learning behavior. The validation curves exhibited greater variation and remained relatively higher than the training curves, suggesting reduced generalization performance. This trend is consistent with the quantitative results, in which MobileNetV2 demonstrated comparatively lower predictive performance, particularly for square footage prediction. The training performance of InceptionV3 is shown in Figure 4. The training and validation loss curves are presented in Figure 4A, whereas the training and validation MAE curves are shown in Figure 4B. The model exhibited a steady reduction in both training and validation loss and MAE, with only a small difference between the curves. This behavior suggests effective learning and relatively good generalization. The stable validation curves indicate that the model captured multiscale spatial features effectively, contributing to improved performance compared with MobileNetV2. The training behavior of ResNet101 is shown in Figure 5. The training and validation loss curves are presented in Figure 5A, whereas the training and validation MAE curves are shown in Figure 5B. The model demonstrated a gradual decline in training loss and MAE, while the validation curves followed a similar trend with minor fluctuations. Although some oscillation was observed, overall convergence remained stable. These observations are consistent with the model’s ability to learn hierarchical spatial representations through residual learning. The training performance of VGG16 is shown in Figure 6. The training and validation loss curves are presented in Figure 6A, whereas the training and validation MAE curves are shown in Figure 6B.

The training curves of VGG16 showed a continuous decrease throughout training, indicating effective learning. The validation curves exhibited oscillations and a slight separation from the training curves, suggesting mild overfitting. Nevertheless, the model remained competitive, particularly for square footage and garage prediction, indicating that its learned feature representations remained effective and stable.

The training performance of CARE-MIRV-Net is shown in Figure 7. The training and validation loss curves are presented in Figure 7A, whereas the training and validation MAE curves are shown in Figure 7B. The training and validation loss and MAE curves decreased smoothly and consistently, with only a small difference between them. This behavior indicates strong convergence and good generalization. Compared with the individual models, the meta-ensemble framework produced more stable learning behavior and reduced variability. Overall, comparison of the training curves suggests that although the individual models demonstrated different convergence and generalization characteristics, CARE-MIRV-Net exhibited the most consistent training behavior among the evaluated models.

A detailed comparison of the predictive performance of MobileNetV2, InceptionV3, ResNet101, VGG16, and the proposed CARE-MIRV-Net model is provided in Table 2. Performance was evaluated using MAE and the R² across four target variables: square footage, number of bedrooms, number of bathrooms, and number of garages. These metrics provide complementary measures of prediction accuracy and explanatory performance.

Model	Square Feet		Beds		Baths		Garages
	MAE	R²	MAE	R²	MAE	R²	MAE	R²
MobileNetV2	881.653	0.1327	0.6873	0.3154	0.6552	0.3457	0.3977	0.1133
InceptionV3	464.64	0.6615	0.4353	0.6708	0.3627	0.7459	0.23	0.6351
ResNet101	611.919	0.5872	0.4287	0.6546	0.3838	0.725	0.2298	0.6489
VGG16	504.435	0.7256	0.698	0.3355	0.4995	0.6436	0.197	0.6757
CARE-MIRV-Net (Proposed)	432.488	0.7053	0.4343	0.6837	0.3689	0.7605	0.1697	0.6958

Table 2: Performance comparison of baseline models and CARE-MIRV-Net. Comparison of prediction performance for MobileNetV2, InceptionV3, ResNet101, VGG16, and the proposed CARE-MIRV-Net model across four architectural attributes: square footage, number of bedrooms, number of bathrooms, and number of garages. Performance is reported using mean absolute error (MAE) and the coefficient of determination (R²). Lower MAE values indicate lower prediction error, whereas higher R² values indicate stronger agreement between predicted and actual values.

The individual models demonstrated varying levels of performance across the target variables. MobileNetV2 showed the lowest performance overall, particularly for square footage prediction, with an MAE of 881.6530 and an R² value of 0.1327. Performance for bathrooms and garages was also comparatively lower, indicating limitations in capturing complex spatial relationships when used as a standalone model.

InceptionV3 demonstrated strong performance across most target variables, particularly bathrooms and bedrooms, with R² values of 0.7459 and 0.6708, respectively. These results suggest effective capture of multiscale spatial characteristics. ResNet101 also demonstrated competitive performance through its deep residual architecture, which facilitated learning of hierarchical spatial representations. However, its performance remained slightly lower than that of InceptionV3 for some variables, including square footage prediction. VGG16 provided relatively balanced performance across all target variables and achieved the highest square footage R² value (0.7256) among the individual models.

The proposed CARE-MIRV-Net framework outperformed the individual models in most evaluated tasks. CARE-MIRV-Net achieved the lowest square footage MAE (432.488) while maintaining an R² value of 0.7053. For bedroom prediction, the model achieved an R² value of 0.6837. For bathroom prediction, CARE-MIRV-Net achieved the highest R² value (0.7605) among all evaluated models. Similarly, for garage prediction, the model achieved the lowest MAE (0.1697) and highest R² value (0.6958). These results indicate that the meta-ensemble framework effectively combined the complementary strengths of the individual architectures. The results further suggest that integrating predictions from multiple models reduced model-specific limitations and improved overall predictive performance. The findings demonstrate that CARE-MIRV-Net provided balanced predictive performance across all evaluated target variables. The healthcare-oriented residential assessment presented in this study should be interpreted as an application-oriented categorization derived from the rule-based decision framework and available floor-plan attributes, rather than as a clinically validated healthcare assessment.

To further evaluate predictive performance, residual error analysis was performed for all target variables, including square footage, bedrooms, bathrooms, and garages. Residual plots provide insight into prediction error distributions by showing the differences between predicted and observed values. Ideally, residuals should be randomly distributed around zero, indicating limited systematic bias and good generalization.

Figure 8 shows the residual plot for square footage prediction. Most residuals were clustered around zero, indicating relatively accurate predictions for the majority of samples. However, greater dispersion and several outliers were observed at larger predicted square footage values. These observations suggest that prediction variability increased for larger residential layouts.

Figure 8. Residual analysis for square footage prediction. Scatter plot showing residual errors as a function of predicted square footage values. The horizontal line indicates zero residual error. Please click here to view a larger version of this figure.

The distribution of residuals for bedroom prediction is shown in Figure 9. Residuals were generally centered around zero, with the diagonal pattern reflecting the discrete nature of the bedroom variable. The relatively narrow residual range indicates consistent predictive performance, with no obvious systematic bias observed. Figure 10 presents the residual plot for bathroom prediction. Residuals remained concentrated around zero but exhibited slightly greater dispersion than those observed for bedroom prediction. The visible banding pattern is consistent with the discrete nature of the target variable. Several outliers were also present at higher predicted values.

Figure 9. Residual analysis for bedroom prediction. Scatter plot showing residual errors as a function of predicted bedroom values. The horizontal line indicates zero residual error. Please click here to view a larger version of this figure.

Figure 10. Residual analysis for bathroom prediction. Scatter plot showing residual errors as a function of predicted bathroom values. The horizontal line indicates zero residual error. Please click here to view a larger version of this figure.

The residual analysis for garage prediction is shown in Figure 11. Residual values remained concentrated around zero and exhibited relatively low variance compared with the other target variables. The discrete and limited range of garage values produced distinct linear patterns within the plot. The narrow residual distribution suggests relatively consistent predictive performance for this variable.

Figure 11. Residual analysis for garage prediction. Scatter plot showing residual errors as a function of predicted garage values. The horizontal line indicates zero residual error. Please click here to view a larger version of this figure.

Overall, the residual analyses indicate that CARE-MIRV-Net generated predictions that were generally centered around zero across all target variables. Although some variability and outliers were observed, particularly for square footage prediction, the overall residual distributions support the consistency of the model’s predictive performance across the evaluated dataset.

Results of Explainable AI Analysis Using SHAP

To improve the interpretability of the proposed CARE-MIRV-Net framework, an explainable AI approach was incorporated using SHAP. SHAP is a game-theoretic method that explains model predictions by assigning feature-importance values based on Shapley values. When applied to the proposed meta-ensemble model, SHAP was used to examine the contribution of each base model to the final output of the meta-learner. The input to the meta-model consisted of concatenated predictions from four base models—InceptionV3, MobileNetV2, ResNet101, and VGG16—resulting in a 16-dimensional feature space. Each feature represented a predicted architectural attribute (square footage, bedrooms, bathrooms, or garages) generated by a specific base model. Interpretation of these features was performed by selecting a randomly chosen subset of the meta-training data as the background distribution and a smaller subset of test samples for explanation generation to improve computational efficiency.

Single-output versions of the meta-model were created to isolate and analyze feature contributions for each target variable independently. The SHAP KernelExplainer was applied as described previously and in Supplementary File 1 to estimate feature contributions and interpret the predictions of the proposed framework. These values were used to quantify the extent to which each base-model prediction contributed positively or negatively to the final output. Based on the SHAP analysis, CARE-MIRV-Net did not rely exclusively on any single base learner but instead assigned varying levels of importance to different models depending on the target variable and input characteristics. These observations indicate that the stacking-based ensemble approach utilized complementary information from multiple base models.

Figures 12–15 present SHAP summary plots for the four target variables: square footage, bedrooms, bathrooms, and garages. These plots provide insight into how the CARE-MIRV-Net meta-model used predictions from the individual base learners. In each plot, the horizontal axis represents SHAP values (feature contributions), whereas the color scale represents feature magnitude from low to high values.

Figure 12. SHAP summary plot for square footage prediction. SHAP summary plot showing the contribution of base-model predictions to square footage estimation by the CARE-MIRV-Net meta-learner. Features are ranked according to importance. Point color indicates feature value, and horizontal position indicates SHAP contribution to the prediction. SHAP, SHapley Additive exPlanations. Please click here to view a larger version of this figure.

Figure 13. SHAP summary plot for bedroom prediction. SHAP summary plot showing the contribution of base-model predictions to bedroom estimation by the CARE-MIRV-Net meta-learner. Features are ranked according to importance. Point color indicates feature value, and horizontal position indicates SHAP contribution to the prediction. SHAP, SHapley Additive exPlanations. Please click here to view a larger version of this figure.

Figure 14. SHAP summary plot for bathroom prediction. SHAP summary plot showing the contribution of base-model predictions to bathroom estimation by the CARE-MIRV-Net meta-learner. Features are ranked according to importance. Point color indicates feature value, and horizontal position indicates SHAP contribution to the prediction. SHAP, SHapley Additive exPlanations. Please click here to view a larger version of this figure.

Figure 15. SHAP summary plot for garage prediction. SHAP summary plot showing the contribution of base-model predictions to garage estimation by the CARE-MIRV-Net meta-learner. Features are ranked according to importance. Point color indicates feature value, and horizontal position indicates SHAP contribution to the prediction. SHAP, SHapley Additive exPlanations. Please click here to view a larger version of this figure.

Figure 12 shows the SHAP summary for square footage prediction. Features such as Mob_SF, Res_SF, and Inc_SF exhibited comparatively larger SHAP contributions than other features, indicating that predictions generated by MobileNetV2, ResNet101, and InceptionV3 contributed substantially to square footage estimation. Higher feature values (red) were generally associated with increased predicted square footage, whereas lower feature values (blue) were associated with lower predictions. Bedroom- and bathroom-related features exhibited comparatively smaller contributions, indicating that square footage prediction was influenced primarily by area-related predictions generated by the base models.

Figure 13 presents the SHAP analysis for bedroom prediction. Features such as Res_Beds, VGG_Beds, and Inc_Beds exhibited comparatively higher contributions than other features, indicating that predictions from ResNet101, VGG16, and InceptionV3 were influential in determining the final bedroom prediction. SHAP values were distributed around zero, suggesting that the meta-model integrated information from multiple sources rather than relying on a single dominant predictor. Features related to garages and square footage exhibited comparatively smaller contributions.

Figure 14 presents the SHAP summary for bathroom prediction. Features including VGG_Baths, Inc_Baths, and Mob_Baths showed comparatively high contributions, indicating that predictions from VGG16, InceptionV3, and MobileNetV2 contributed substantially to bathroom estimation. The distribution of SHAP values suggests that increased predictions from these base models were associated with greater influence on the final prediction. Contributions appeared more broadly distributed across models than for some other target variables.

Figure 15 presents the SHAP analysis for garage prediction. Features such as Mob_Garages, Inc_Garages, and VGG_Garages exhibited relatively strong contributions to the final prediction. The plot demonstrates that garage-related predictions from multiple base models contributed to the final output, with distinct separation between lower- and higher-valued feature contributions. The wider spread of SHAP values suggests increased sensitivity of the meta-model to garage-related prediction features.

Overall, the SHAP analysis demonstrated that the proposed CARE-MIRV-Net framework integrated contributions from multiple base learners when generating final predictions. The results provide an interpretable view of the relative influence of individual model predictions across the four architectural target variables and support understanding of the meta-ensemble decision-making process.

Results of Residential Layout Classification

Representative examples of the residential layout classification results are shown in Figures 16–18. These examples illustrate the application of the rule-based decision layer to the architectural attributes predicted by the CARE-MIRV-Net framework. Figure 16 presents an example classified as a medical care layout, Figure 17 presents an example classified as a general residential layout, and Figure 18 presents an example classified as an elderly care layout. The examples demonstrate how predicted architectural attributes are translated into application-level residential categories through the predefined decision rules described in the Protocol section. These examples should be interpreted as illustrative outcomes of the rule-based classification framework derived from the predicted floor-plan attributes and predefined decision criteria, rather than as independently validated clinical or healthcare classifications.

Figure 16. Representative example classified as a medical-care residential layout. Example residential floor plan image evaluated by the proposed framework and assigned to the medical-care category based on the predicted architectural attributes and rule-based classification criteria. Please click here to view a larger version of this figure.

Figure 17. Representative example classified as a general residential layout. Example residential floor plan image evaluated by the proposed framework and assigned to the general residential category based on the predicted architectural attributes and rule-based classification criteria. Please click here to view a larger version of this figure.

Figure 18. Representative example classified as an elderly-care residential layout. Example residential floor plan image evaluated by the proposed framework and assigned to the elderly-care category based on the predicted architectural attributes and rule-based classification criteria. Please click here to view a larger version of this figure.

A representative example shown in Figure 16 demonstrates that the model predicted a square footage of 2233.16 compared with the actual value of 2268.00, indicating a relatively small prediction error. Similarly, the predicted values for bedrooms (3.23 vs. 3), bathrooms (2.65 vs. 2), and garages (1.11 vs. 1) were close to the corresponding ground-truth values. Based on these predicted attributes, the rule-based decision layer categorized the layout as medical care. This example illustrates how the framework applies predicted architectural attributes to generate application-level residential classifications.

A second example shown in Figure 17 demonstrates a layout classified as general residential use. The predicted square footage (1692.76) was lower than the actual value (1879.00), whereas the predicted numbers of bedrooms, bathrooms, and garages remained close to the corresponding ground-truth values. Based on the predicted architectural attributes, the decision layer assigned the layout to the general residential category. This example demonstrates the application of the classification framework to a moderately sized residential layout despite minor deviations in individual predictions.

Figure 18 presents a compact residential layout with an actual square footage of 981.00 and a predicted square footage of 1076.17. The predicted values for bedrooms (1.99 vs. 2), bathrooms (1.75 vs. 2), and garages (0.98 vs. 1) were also close to the corresponding ground-truth values. Based on these predicted attributes, the decision layer categorized the layout as elderly care. This example illustrates the application of the predefined classification rules to layouts characterized by smaller floor areas and fewer rooms.

Overall, the representative examples demonstrate the integration of architectural attribute prediction and rule-based residential layout categorization within the CARE-MIRV-Net framework. Across the examples shown in Figures 16–18, the predicted architectural attributes were generally consistent with the corresponding ground-truth values, supporting the use of the proposed framework for automated residential floor plan analysis and application-level classification.

Data Availability:

The dataset used in this study is publicly available through the Kaggle repository²⁸, entitled “Floor Plan Images and Their Details,” and can be accessed at: https://www.kaggle.com/datasets/adilmohammed/floor-plan-images-and-their-details. All analyses reported in this study were conducted using data obtained from this publicly accessible source. The source code, implementation documentation, and environment configuration information required to reproduce the reported results are provided in Supplementary File 1.

Supplementary File 1. CARE-MIRV-Net implementation package. This file contains the source code, README documentation, software dependency specifications, environment information, and implementation scripts required to reproduce the CARE-MIRV-Net workflow.Please click here to download this file.

Supplementary File 2. Algorithm for CARE-MIRV-Net residential floor plan analysis. This file provides Algorithm 1, including the data preprocessing, base-model prediction, meta-feature concatenation, meta-ensemble prediction, decision-layer classification, SHAP-based explainability, and evaluation-metric workflow used in the study.Please click here to download this file.

Discussion

This study presents CARE-MIRV-Net, a stacking-based meta-ensemble DL framework for the automated analysis of residential floor plan images. The framework integrates four complementary CNN architectures—MobileNetV2, InceptionV3, ResNet101, and VGG16—to predict key architectural attributes, including square footage, number of bedrooms, bathrooms, and garages. These predictions are subsequently combined through a meta-learner and used within a rule-based decision layer to categorize residential layouts into elderly care, medical care, or general residential categories. The integration of predictive modeling, ensemble learning, and explainable AI provides a unified framework for residential floor plan analysis^3,7,17,20.

The representative results demonstrate that CARE-MIRV-Net achieved improved predictive performance compared with the individual base models across multiple architectural attributes. The ensemble framework generally reduced prediction errors and improved the R², suggesting that combining heterogeneous CNN architectures can improve predictive robustness^12,20. In addition, the incorporation of SHAP enabled interpretation of the relative contributions of the individual base models to the final predictions, thereby increasing model transparency^11,17. The framework may therefore provide a useful approach for studies requiring both predictive performance and model interpretability.

The proposed framework may contribute to the advancement of research involving automated architectural analysis, residential design assessment, and data-driven decision-support systems. By integrating multioutput prediction with application-level categorization, the framework extends beyond traditional floor plan recognition or feature extraction approaches. The methodology may be applicable in studies involving residential planning, smart housing environments, and spatial-layout assessment^1,2,8,22. However, several limitations should be considered. First, the framework was developed and evaluated using a single publicly available dataset, and external validation on independent datasets was not performed. No external validation was conducted in the present study, and model evaluation was performed using the designated test subset derived from the publicly available dataset. Second, the residential categories generated by the framework are derived from predefined rule-based criteria rather than independently validated healthcare-design labels. The elderly care, medical care, and general residential categories generated by the framework are derived from predefined rule-based criteria and should not be interpreted as clinically validated healthcare-design assessments. Third, the framework relies primarily on architectural attributes derived from floor plan images and does not incorporate healthcare-specific design standards, accessibility criteria, or occupant-centered outcome measures. No formal accessibility, healthcare-design, or residential-design standards were used to define the classification rules. Instead, the categorization framework was developed using the available floor-plan attributes and application-oriented decision criteria. Future work may include external validation on independent datasets, incorporation of formal accessibility and healthcare-design standards, and evaluation using additional design and occupant-centered outcome measures to further enhance the applicability and generalizability of the framework.

Alternative approaches may also be used to investigate the same hypothesis. For example, transformer-based vision models, graph-based spatial representations, or hybrid DL architectures could be applied to capture spatial relationships within residential layouts^13,15,19. Similarly, classification frameworks based on expert-labeled datasets or healthcare-design standards may provide alternative methods for residential layout assessment^8,22. Although these methods may offer complementary advantages, the proposed CARE-MIRV-Net framework demonstrates the feasibility of combining multiple CNN architectures within a stacking-based ensemble framework for multioutput prediction and residential layout categorization.

Future work may focus on extending the framework to support more advanced spatial reasoning and broader real-world applicability. Potential directions include the incorporation of graph-based floor plan representations to model room connectivity, the evaluation of vision transformer (ViT)-based architectures for feature extraction, and the development of multiclass or multilabel classification strategies^13,15,18. In addition, future studies may incorporate healthcare-specific design attributes, accessibility standards, and independent validation datasets to further assess the applicability of the framework in healthcare-oriented residential environments^8,22. The framework may also be adapted for integration within smart-city platforms and decision-support systems for residential planning^1,2.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
CARE-MIRV-Net implementation package	Supplementary File 1	Version 1.0	N/A
Figure and table regeneration scripts	Supplementary File 1	Version 1.0	N/A
Floor Plan Images and Their Details Dataset	Kaggle	Public Dataset	N/A
GPU	NVIDIA Corporation	Tesla T4 × 2	N/A
Intel Xeon Processor	Intel Corporation	2.20 GHz	N/A
Keras	Keras Team	Version 3.13.2	N/A
Matplotlib	Matplotlib Development Team	Version 3.9	RRID:SCR_008624
NumPy	NumPy Developers	Version 2.4.6	RRID:SCR_008633
Operating System	Linux	Ubuntu 22.04 LTS	N/A
Pandas	Pandas Development Team	Version 2.3.3	RRID:SCR_018214
Python	Python Software Foundation	Version 3.12.12	RRID:SCR_008394
Scikit-learn	Scikit-learn Developers	Version 1.6.1	RRID:SCR_002577
SHAP	SHAP Development Team	Version 0.51.0	N/A
System Memory (RAM)	Cloud Computing Environment	31 GB	N/A
TensorFlow	Google	Version 2.19.0	RRID:SCR_016345
Training environment configuration file	Supplementary File 1	Version 1.0	N/A