Interactive Multiscale Attention Fusion with Sparse Feature Convolution Network for Satellite Image Change Detection

Summary

This study presents an Interactive Multiscale Attention fusion model with Sparse Feature Convolution Networks to improve satellite image change detection. The approach takes advantage of multiscale feature extraction, the selective focus on attention, and sparse convolutions to effectively detect and localize change in bi-temporal satellite images by reducing computational complexity.

Abstract

Change detection using bi-temporal satellite imagery is an important problem in earth monitoring that aims at identifying and localizing surface changes between a temporal acquisition and distinguishing actual pseudo-changes due to seasonal cycles, atmospheric factors, or man-made developments. Current deep learning methods typically face a one-way bias in their temporal modeling, which restricts the usability of large-scale spatial relationships to empower the proper characterization of change patterns. Although recent transformer-based techniques are highly accurate, their computational requirements are also quite large, thus limiting their use on resource-constrained tasks. In response to these limitations, an Interactive Multiscale Attention Fusion (IMAF) network with Sparse Feature Convolution (SFC) based on Ghost modules to attain effective multi-scale features extraction is proposed in this paper. The network utilizes advanced encoder-decoder architecture supplemented with interactive attention modules in different scales. The network works with complementary forward and backward attention streams: the former one records progressive temporal variation and the latter one verifies a consistency of the changes by reverse temporal analysis. This interactive method allows effective representations of the duplex model of temporal relationships without being computationally prohibitive, enhancing far better approaches to distinguish legitimate land cover moves, versus noise absent noise-induced variation. Two benchmark datasets are experimented with: the Onera Satellite Change Detection dataset (OSCD), and the SZTAKI AirChange datasets. The significant experiments made on OSCD datasets reveal that proposed approach attain competitive F1 rates of 58.14% (13 channels) and 50.68% (3 channels) with just 2.13M parameters and 19.6G FLOPS, 19 times few parameters, and 5.6x inference turnover as compared to ChangeFormer. Competitive performance on Szada/1 and Tiszadob/3 test set samples of the SZTAKI dataset, F1-score of 74.29% and 92.86% respectively makes it possible to implement in the most edge devices, real-time analytics, and extensive mapping systems where operational energy directly depends on computational energy.

Introduction

Unprecedented levels of urbanization in the global environment, environmental resource loss, and alterations of the landscape due to climate have directly triggered the necessity to provide an exact, timely and automatic system of monitoring that can identify and measure changes over the surface of the Earth. Satellite remote sensing has grown to be the basic building block of mass environmental surveillance since they provide consistent temporal and spatial coverage in addition to being vital in the complete examination of the change. Bi-temporal Satellite imagery change detection is the characterization of surface changes between two sequential acquisitions of the same geographical region. It is a complicated process, though, because of the numerous confounding variables, including the variation of the seasons, the weather conditions, sensor properties, geometrical distortions, and phenomenological alterations that may obscure or resemble the real in land cover. The manual image interpretation of satellite imagery currently in use is time-consuming, subjective, and not suitable to cope with the continually growing quantities of earth observation information generated by the current satellite constellations. This is the basic limitation that has resulted in the creation of automated change detection solutions. Conventional pixel-based¹ and object-based change detection² methods tend to be poorly suited to operate within the limits of these complexities and have a low standard of robustness to apply to a large range of geographic units and temporal scales. Deep learning architectures, particularly convolutional neural networks and variants, enable the extraction of hierarchical features that represent low-level spectral detail and high-level semantic patterns relevant for distinguishing changes. Deep learning change detection models have proven to be highly promising in capturing latent spatio-temporal changes in remote sensing data, and in identifying the true land cover versus pseudo changes, introduced by extraneous factors. These end-to-end learning approaches take advantage of the fact that features can be extracted and classification can be performed optimally at once. Despite substantial progress in change detection accuracy, most existing approaches remain difficult to deploy in real-world operational environments. Transformer-based approaches³, though efficient in the global context, have a high parameter count, quadratic computational complexity, inference latency, and high energy consumption. Siamese CNN-based architectures, on the other hand, are superior in efficiency, yet have small receptive fields and lack global contextual modelling in large part, thus significantly worsening performance. These constraints are further exacerbated in the large-scale and resource-constrained environments, where scalability and operational cost are critical considerations. Scalability, energy efficiency, and the speed at which inferences are made are frequently vital in an emergency response setting, and there is a need to create solutions that balance performance and deployability. Therefore, a gap remains between accuracy-focused computationally-intensive models and the practical requirements of real-world deployment.

To overcome such issues, an Interactive Multiscale Attention Fusion (IMAF) with Sparse Feature Convolution (SFC) Network is proposed, which focuses on the design of an efficient and deployable change detection framework as opposed to accuracy optimization alone. The proposed solution focuses on optical satellite image pairs, usually RGB or high-resolution multispectral images (0.5-30 m resolution), acquired by satellites, e.g., Landsat, Sentinel-2, or WorldView. The technique presupposes precise geometric co-registration of bi-temporal pairs, because registration errors can also have a considerable effect on detection. It deals with binary change detection (as opposed to multi-class semantic change) and cannot distinguish between various change types. More than this, the method uses radiometrically corrected imagery to reduce the effects of seasonal, atmospheric, and illumination changes, which would be mistaken as land cover change. As opposed to the traditional Siamese CNNs⁴ that separate bi-temporal images with the aid of a simple concatenation, the proposed IMAF module allows multiscale contextual fusion of both global and local features. Recent transformer-based approaches, such as BIT⁵ and ChangeFormer³ are getting very good results, but their self-attention mechanisms have the O(N²) complexity. The proposed framework captures subtle changes with bidirectional attention, reducing computational complexity and inference latency. In addition to that, the proposed design is sparsity-aware, which is superior to dense multiscale fusion designs such as UNet++^6,7 and SNUNet⁸. The sparse feature convolution with the ghost module reduces FLOPs by approximately 50%, preserving representation quality. Thus, the proposed framework favors an accuracy-efficiency tradeoff for reliable deployment on resource-constrained environments.

The main contributions of this study are: (i) A lightweight change detector that attains a competitive F1-score using 19 times fewer parameters compared with transformer-based counterparts. (ii) An Interactive Multiscale Attention Fusion module that obtains global contextual information without the quadratic computational cost of the traditional self-attention mechanisms. (iii) A convolution scheme with feature sparsity based on Ghosts modules, which minimizes the floating-point operations by 49.4 percent with no quality degradation in the feature representation. (iv) Extensive experimental validation on a large set of benchmark datasets, with better efficiency-accuracy trade-offs and reasonable practicability of actual application in the real world.

The remainder of this manuscript comprises the following sections: Section 2 gives an extensive overview of the recent work in the area of binary change detection methodologies, with special attention given to deep learning solutions and their implementation on standard benchmarks. Section 3 gives the theoretical background and structural design of the proposed solution, mathematical representation of the temporal fusion mechanism and the attention elements, the characteristics of the experimental setup, datasets, evaluation metrics, and implementation details required to produce reproducible research. Section 4 illustrates detailed experimental findings that comprise ablation, comparisons with the latest binary change detection algorithms that explain the capacity in various geographical areas in the OSCD and SZTAKI Airchange datasets. Section 5 determines the outcome of findings, limitations of a current approach, and possibilities of future research in change detection.

The development of change detection algorithms in remote sensing has been greatly changed from pixel-based methods to deep learning frameworks. Early change detection procedures were mainly based on algebraic processes like image differencing, image rationing, and change vector analysis, which worked directly on pixel values to detect temporal changes^9,10. Alternative methods involving post-classification comparison were developed, in which each temporal image was independently classified and then compared with cross-tabulation¹¹. Object-based change detection techniques represented an important step of progress with the introduction of spatial context as well as mitigation of pixel-level noise^2,12.

The advent of deep learning techniques has revolutionized the space of remote sensing change detection and paved the way for overcoming several constraints in traditional techniques. Recently, Convolution Neural Networks (CNN's) have become the standard architecture for automated feature extraction and change classification^13,14. Siamese NestedUNet for Change Detection¹⁴and Siamese Swin-Unet¹⁵ have emerged as a significant innovation owing to densely connected Siamese network topology, which helps to effectively retain information of localization across the hierarchy of the network.

The problem of bi-temporal change detection in high-resolution optical remote sensing is still difficult to resolve because of a variety of confounding factors, and the attention-based multi-scale transformer networks with the elaborate features fusion strategies have been proposed¹⁶. The EGMT-CD framework proposed edge-guided multimodal transformers of heterogeneous image pairs, where the homologous images are not accessible because of cloud cover¹⁷. Architectures like DASUNet break through manual deep learning bottlenecks in the full-scale feature fusion, with 0.85% higher F1 index improvement over SNUNet-24 on the CDD dataset using improved gradient flow¹⁸. Hybrid CNN-Transformer models deal with the issue of false negatives at a higher cost by providing an additional on-cost on the frequency components to feature learning¹⁹. The hybrid U-shaped transformer networks²⁰, however, are still difficult to provide local-global features integration and changes in regions that periodically occur, even on a small scale. The use of multitask semantic change detection, and precise location and class identification is more complicated compared to binary tasks^21,22. Bi-temporal correlations can be successfully modeled by cross-attention without significant architectural changes^14,23. The perspective-language paradigm expression is also a recent frontier with CLIP-based designs combining textual explanations into change correspondence²⁴ and semi-supervised patterns eliminating reliance on labelled data²⁵. Change Mamba presents the state space models of spatio-temporal models^26,27. Different applications of specialized attention methods are demonstrated: addressing the use of hyperspectral anomaly detection²⁸, multiarea target classification²⁹, or limited-label segmentation³⁰, and demonstrate the flexibility of attention mechanisms in remote sensing analytics.

Authors	Method/Architecture	Technical Specifications	Key Innovation & Findings / Dataset Performance
Singh, A. [8]	Traditional Methods	Pixel-based differencing, CVA	Digital change detection framework, Multiple datasets, Accuracy: 65-80%
Mas, J.F. [10]	Post-classification	Independent temporal classification	Comparative methodology analysis, Tropical imagery, OA: 72-85%
Lu et al. [9]	Traditional Methods	Algebraic operations, CVA, PCA	Comprehensive technique comparison, Multiple sources, Accuracy: 70-88%
Benedek & Szirányi [23]	Mixed Markov Model	Multilayer conditional framework	Probabilistic change modeling, Sztaki AirChange, DA: 92.1%
Blaschke, T. [2]	Object-based	Segmentation-based analysis	Spatial context integration, Various HR imagery, SA: 85-92%
Chen et al. [11]	Object-based	Multi-scale segmentation	Hierarchical object analysis, HR imagery, OA: 87.3%
Zhan et al. [12]	Siamese CNN	5-layer CNN, kernel size 3×3	Deep siamese architecture for CD, Aerial images, F1: 0.847, IoU: 0.735
Daudt et al. [4]	Siamese FCN	FC-EF, FC-Siam-diff, FC-Siam-conc	Early/late fusion strategies, OSCD, F1: 0.431, IoU: 0.276
Peng et al. [26]	UNet++	Nested U-Net, dense skip connections	Multi-scale feature aggregation, HR satellite, F1: 0.753, IoU: 0.604
Chen & Shi [25]	Spatial-Temporal Attention	Attention blocks, temporal modeling	Spatio-temporal feature learning, LEVIR-CD, F1: 0.887, IoU: 0.797
Fang et al. [7]	SNUNet-CD	Siamese NestedUNet, dense connections	Information transmission optimization, LEVIR: F1: 0.897, WHU: F1: 0.904
Chen et al. [5]	BIT-CD	ResNet18 + Transformer encoder	Binary temporal feature learning, OSCD: F1: 0.635, LEVIR: F1: 0.919
Bandara & Patel [3]	ChangeFormer	Hierarchical transformer encoder	Multi-scale transformer attention, OSCD: F1: 0.654, LEVIR: F1: 0.905
Song et al. [27]	ACANet	Axial attention + CNN backbone	Efficient attention computation, LEVIR: F1: 0.901, WHU: F1: 0.913
Shi et al. [28]	Review Paper	Comprehensive AI methods survey	Systematic analysis of AI approaches, 50+ datasets analyzed
Li et al. [14]	AMST-Net	Multiscale transformer, pyramid attention	Adaptive multiscale feature extraction, HR optical images, mIoU: 0.823
Xiang et al. [15]	EGMT-CD	Edge-guided multimodal architecture	Cross-modal feature alignment, Heterogeneous pairs, F1: 0.756
Miao et al. [16]	DASUNet	Dense supervision, fullscale fusion	Deep supervision at all decoder levels, CDD: 0.85% F1 improvement over SNUNet
Tang et al. [29]	Siamese Swin-UNet	Swin Transformer + Unet fusion	Hierarchical window attention, LEVIR: F1: 0.923, WHU: F1: 0.925
Wu et al. [18]	Hybrid U-Transformer	UNet++ backbone + Transformer blocks	Global-local feature integration, HR images, IoU: 0.851, F1: 0.919
Dong et al. [20]	ChangeCLIP	CLIP-based vision-language	Multimodal semantic understanding, RS pairs, semantic accuracy: 94.1%
Li et al. [21]	SemiCD-VL	Semi-supervised visionlanguage	Reduced annotation requirements, Limited labels, F1: 0.889
Chen et al. [22]	ChangeMamba	State Space Models (SSMs)	Linear complexity temporal modeling, RS images, efficiency: 3x faster

Table 1: Summary of Change Detection Methods in Remote Sensing Please click here to download this Table.

A summary of prior studies is presented in Table 1. Traditional methods have handcrafted features that are not suitable to complex spatial-temporal variations³¹, and deep learning architecture (UNet, FPN, PSPNet) involves computationally expensive dense convolutions. The current models do not have efficient multi-scale incorporation and dynamic attention systems. In order to fill these gaps, the current study is proposing Interactive Multi-scale Attention Fusion with Sparse Feature Convolution Network, which aims at capturing the subtle changes across scales while being computationally efficient in remote sensing, particularly on urban and man-made structural change.

Protocol

1. Ethics statement

1. Both Onera Satellite Change Detection (OSCD) and SZTAKI AirChange datasets used in this study consist of publicly available RGB and multispectral images covering diverse geographical characteristics. These datasets do not contain any restricted or sensitive data. Therefore, no ethical consent is required for this work.

2. Materials and equipment

1. Perform testing on a Windows 11 computer containing an Intel Core i7 10870H processor, an NVIDIA GeForce GTX 1650 Ti chip with 4GB of VRAM, and 32GB of RAM.
   NOTE: The programming environment is Python 3.8, and the programming environment depends on the PyTorch framework (version 1.12.1) and Torchvision (version 0.13.1) to implement deep learning.
2. Download and install libraries such as scikit-learn (version 1.0.2), containing machine learning utility, and NumPy (version 1.21.0), containing numerical operations.
3. Make sure to install the CUDA Toolkit version 11.3 and cuDNN version 8.2 provided by NVIDIA Corporation to be in a position to utilize the acceleration of the graphics card.
4. Train and evaluate performance using the publicly available OSCD and SZTAKI Airchange dataset.

3. Dataset preparation

1. Select OSCD⁴, which is composed of optical RGB satellite aerial images and multispectral images, each of 600 x 600 pixels with 10 m resolution and the SZTAKI AirChange³² dataset, which is composed of 13 optical aerial image pairs, each 952 x 640 pixels with a resolution of 1.5 m/pixel as benchmark datasets.
2. Partition the 24 cities of the OSCD dataset into fixed, predefined 14 cities for training and 10 cities for testing.
3. Assign fixed non-randomized Szada/1 and Tiszadob/3 image pairs from SZTAKI AirChange dataset to the test set and the remaining image pairs as the training set as a conventional benchmark.
4. Perform change detection on the test set and quantify performance on a single run using annotated image pairs from each dataset.

4. Preprocessing

1. Apply patch-based preprocessing.
   1. Extract image patches independently per temporal image at identical locations on-the-fly to manage high-resolution satellite images efficiently.
   2. Divide each image into overlapping patches of 128 × 128 pixels using a stride of 64 pixels.
   3. Align overlapping regions between adjacent patches to maintain boundary consistency and retain edge information.
2. Perform class-wise data augmentation.
   1. Apply differential augmentation during training on both patches in the pair based on the proportion of change pixels to handle class imbalance between "change" and "no-change" regions.
   2. Augment change patch pair (those with > 1% change pixels) six times using the following transformations: Horizontal and vertical flips, 90° rotations, Color jitter (brightness, contrast, and saturation ± 30%), Affine transformations (rotation ± 15°, translation ± 10 pixels, scaling 95-105%).
   3. Augment the no-change patch pair only once to prevent dataset imbalance.
      NOTE: This augmentation strategy, based on classes, provides balanced training in the change and no-change regions.
3. Normalize and enhance image quality.
   1. Normalize all patches using ImageNet mean and standard deviation: Mean = (0.485, 0.456, 0.406), Std = (0.229, 0.224, 0.225).
      1. Use Contrast Limited Adaptive Histogram Equalization (CLAHE) with clip limit = 2.0, tile grid size=8 x 8, in LAB Color Space (L Channel Only) to brighten up the contrasts of the areas with low contrast so that the features can be distinguished on the image, with intensities normalized to [0,255].
      2. Clear any patches that are smaller than 128 x 128 pixels and pad the smaller patches so that they are of the same size as the larger patches and maintain their spatial integrity during model training.

5. Model construction

1. Define input and output.
   1. The proposed framework takes as input a pair of co-registered satellite images I₁, I₂∈ R^H×W×C acquired at different temporal instances and produces a binary change map M ∈ R^H×W×2 that accurately delineates changed and unchanged regions.
2. Proposed framework
   1. Process the proposed framework as three stages processing pipeline. In the first stage, perform temporal attention modelling on input sequences to capture change over time.
   2. Use interactive cross-attention to capture symmetric temporal dependencies among bi-temporal features in the temporal attention modeling stage.
   3. Extract hierarchical representations of relevant features from a temporally modeled image using a sparse feature-based encoder architecture in the second stage, where spatial resolution gradually decreases while the feature dimension is increased.
   4. Use a skip-connected decoder with adaptive upsampling to produce high-resolution change maps in the reconstruction stage, while retaining lower-scale spatial details.
      NOTE: Figure 1 shows the overall data flow from input bi-temporal satellite images to the shared encoder, middle cross-attention processing blocks and decoder, which all together result in the change detection map.
3. Sparse feature encoder
   1. Construct a lightweight encoder built from Ghost convolutional blocks with ratio=2 with batch normalization and ReLU activation function applied sequentially to reduce computational complexity while maintaining representational quality.
   2. Assemble the encoder sequentially from the following three types of blocks.
      Block₁: Ghost(in_channels → 32) → BN → ReLU → Ghost(32 → 64) → BN → ReLU → MaxPool(2×2)
      Block₂: Ghost(64 → 96) → BN → ReLU → Ghost(96 → 128) → BN → ReLU → MaxPool(2×2)
      Block₃: Ghost(128 → 128) → BN → ReLU
   3. Connect the blocks to maintain smooth feature flow through the encoder and preserve spatial resolution for subsequent stages.
      NOTE: The core innovation of this approach resides in the interactive cross-attention mechanism that enables symmetric feature interaction between bi-temporal images. For each image pair, deep features F₁,F₂∈R^B×C×H'×W' are extracted where H' = H/4, W'=W/4, C=128 represents the feature dimension. These spatial features are reshaped to a sequence format  represents the spatial sequence length, and D=C represents the feature dimension. 
4. Temporal attention modelling
   1. Implement cross-attention operation using standard scaled dot-product attention formulation as illustrated in Figure 2
      
      where Q, K, and V represent query, key, and value matrices respectively, and d_k denotes the dimension of key vectors.
      NOTE: Multi-head attention with h = 8 heads to capture different types of temporal relationships simultaneously is employed independently at each scale. Each attention head processes a different subspace of the feature representations, enabling the model to focus on various aspects of temporal changes.
   2. Compute the multi-head attention output as:
      MultiHead (Q,K,V) = Concat (head₁,…,head_h ) W^O
      where each  are learned projection matrices.
5. Symmetric interactive attention
   1. The interactive attention strategy captures symmetric change information through forward and backward operations (shown in Figure 3). Forward attention allows features from the first temporal image to attend to features from the second temporal image, computed as:
      
      where the first temporal features serve as queries while the second temporal features provide keys and values.
   2. Conversely, backward attention enables features from the second temporal image to attend to features from the first temporal image, formulated as:
      
      The attended features from both directions are concatenated and projected through a linear transformation to produce the final attended representation:
      
      ​NOTE: This interactive mechanism ensures that temporal relationships are captured symmetrically, making the network invariant to the order of input images, which is crucial for change detection applications where the temporal sequence should not bias the detection results.
6. Decoder and change map reconstruction
   1. Merge the attended features with the encoder outputs through skip connections to retain spatial details.
   2. Apply bilinear interpolation with identical parameters across all decoder stages for adaptive resizing to ensure perfect alignment across skip connections.
   3. In first stage, reduce channels from 256 to 96 pixels via Ghost convolutions at H/4×W/4 and upsample to H/2×W/2 and fuse with corresponding encoder features.
   4. In the second stage, process features with Ghost convolutions to reduce channels from 160 to 64 pixels and upsample to full resolution H×W and integrate with encoder-level skip features.
   5. Lastly, apply Ghost convolutions to reduce channels to 32 and use a final 1 × 1 convolution to generate the two-channel binary change map.
   6. Summarize the complete processing pipeline through the following algorithmic framework, which integrates all architectural components into a cohesive change detection system.
      NOTE: Notation: E^k_i denotes encoder features from image i ∈ {1,2} at level k; F_i is the final encoded feature;  is its flattened version; is the attention-enhanced feature; A is the attention matrix; D_j are decoder outputs; ϕ_k,ψ_j  are encoder/decoder blocks; [·;·] denotes channel-wise concatenation; W_out is the output layer kernel.
      See Supplementary file Algorithm 1 for "Interactive Multiscale Attention Fusion based Change Detection

Figure 1: Proposed Methodology Please click here to view a larger version of this figure.

Figure 2: Cross Attention Architecture Please click here to view a larger version of this figure.

Figure 3: Interactive Cross Attention Architecture Please click here to view a larger version of this figure.

6. Training procedure

1. Loss function
   NOTE: The training process employs a focal tversky loss function to address the class imbalance commonly encountered in change detection datasets.
   1. Formulate the loss function as follows: Let p_i ∈ [0,1] be the predicted probability for pixel i, and g_i ∊ {0,1} the corresponding ground truth. The Tversky index (TI) is defined as:
      
      where α and β control the penalty for false positives and false negatives, respectively, and ε is a smoothing term.
   2. Then express the Focal Tversky Loss as:
      
      with γ modulating the focus on hard-to-classify pixels.
2. Hyperparameter configuration
   1. Apply a 5-fold cross-validation-based systematic grid search on the train set to select optimal values.
   2. Set the loss weighting coefficients for the change detection task to α = 0.3, β = 0.7, γ = 1.0, and ε = 1.0. These values emphasize minimizing missed detections over false positives, which is important for imbalanced datasets where changes are relatively rare.
      ​NOTE: Asymmetric weight with β > α focuses on recall, which means that false negatives are weighted more than false positives, and is consistent with design, especially in disaster monitoring applications.
3. Optimizer and learning strategy
   1. Employ the AdamW optimizer with a weight decay of 1×10⁻⁴ to improve regularization and prevent overfitting.
   2. Initialize the learning rate at 1×10⁻³ and apply a cosine annealing schedule to ensure smooth and stable convergence during training.
   3. Use a batch size of 8 to maintain compatibility with a 4-GB GPU memory environment.
4. Training schedule
   1. Train the model separately on OSCD, consisting of 14 predefined pairs for training and 10 pairs for testing, then on SZTAKI Airchange dataset with Szada/1 and Tiszabob/3 as standard test set benchmarks.
   2. Divide the training pairs into five folds, use one fold as a validation set in each iteration, and continue training for a maximum of 100 epochs.
   3. Monitor validation loss at the end of each fold and apply early stopping criteria of 10 epochs once convergence stabilizes to prevent overfitting and reduce unnecessary computation.
   4. Select the hyperparameters based on best average validation performance across all the folds.
   5. Perform change detection on the test set and quantify performance on single run using annotated image pairs from each dataset.
5. Memory optimization
   1. Activate gradient checkpointing to lower GPU memory usage by approximately 40%, achieved by recomputing intermediate activations during the backward pass.
   2. Enable mixed-precision training to utilize FP16 operations where numerically stable, thereby improving speed and reducing memory load.
   3. Use adaptive batch size scaling to dynamically adjust memory usage for maximum GPU efficiency.

7.Evaluation

1. Ensure that no patches from test cities are included in training or validation to prevent data leakage.
2. Organize image patches and corresponding ground truth maps for batch-wise evaluation.
3. Load the trained model weights from the best-performing epoch determined by validation loss.
4. Select a threshold of 0.5 to convert probability maps into binary change maps.
5. Calculate pixel-wise evaluation metrics to measure change detection performance¹⁰
   Precision (P): Fraction of predicted change pixels that are true change pixels.
   Recall (R): Fraction of true change pixels correctly detected.
   F1-score: Harmonic mean of precision and recall.
6. Produce image patch in real-time while training and evaluating models with dimensions of 128 x 128 pixels and a stride of 64 pixels.
   NOTE: Such dynamic generation guarantees memory effectiveness as well as consistency of boundaries.
7. Check the correctness of preprocessing and patch placement by examining a representative sample of the patches created intelligible, or middle-level feature images.
   NOTE: Encoded feature representations that are generated by the CNN-based encoder are stored in memory in the form of tensors to be used later in the stages, which serve as checkpoints when verifying features. Produced change maps are stored as regular image files (e.g. PNG or TIFF) to be viewed visually in order to evaluate the identified alterations.
8. Record the convergence indicators (loss curves, accuracy metrics) in model training, and check if the points of the model are saved at specified intervals so that the entire training process can be repeated, refined, or recovered.
   NOTE: Explicit presentation of the output format, verification steps, and checkpoint details helps in ensuring transparency, permitting the verification to happen step-by-step and also permitting the protocol to be reproduced and validated by other researchers. This protocol gives an in-depth, detailed workflow of how to prepare a dataset and proceed with the preprocessing phase, network construction, training, and evaluation. By adhering to the steps, one will be able to implement the proposed method under the specified conditions in a repeatable way. The results of such a procedure are given in the following Results section.

Results

The proposed change detection method was evaluated against established baseline approaches using two publicly available benchmark datasets to assess its performance across different spectral resolutions and environmental conditions. The experimental results reveal significant improvements in balancing detection accuracy with false positive control, particularly demonstrating the method's ability to maintain high recall rates while substantially enhancing precision compared to traditional approaches.

Discussion

This study presents an interactive multi-scale attention fusion-based change detection network for detecting changes in bi-temporal satellite images. The integration of a bi-directional attention mechanism, unlike unidirectional methods, facilitates a complete cross-temporal feature-to-feature interaction with complementary forward and backward attention signals. Progressive temporal changes of the forward attention stream are captured by the attention difference i.e. older to newer temporal acquisitions, whereas the cha...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
CUDA Toolkit	NVIDIA Corporation	Version 11.3	GPU acceleration for deep learning computations
cuDNN	NVIDIA Corporation	Version 8.2	Optimized GPU deep learning library
Intel  processor	Intel Corporation	Core i7 10870H	Computational platform used for training and evaluating the proposed deep learning model
NumPy	Open Source	Version 1.21.0	Numerical array processing
NVIDIA GeForce	NVIDIA Corporation	GTX 1650 Ti (4 GB VRAM)	Graphics processing unit used for accelerated model training
Onera Satellite Change Detection (OSCD) Dataset	ONERA	Publicly available optical RGB and Multispectral image dataset used for training and evaluation.	https://rcdaudt.github.io/oscd/
Python	Python Software Foundation	Version 3.8	Programming language used for algorithm implementation
PyTorch	Meta AI (Open Source)	Version 1.12.1	Open-source deep learning framework used to develop and train the proposed model
Scikit-learn	Open Source	Version 1.0.2	Performance evaluation metrics
SZTAKI AirChange Benchmark Dataset	SZTAKI	Publicly available optical RGB aerial image dataset used for training and evaluation.	http://web.eee.sztaki.hu/remotesensing/airchange_benchmark.html
Torchvision	Meta AI (Open Source)	Version 0.13.1	Dataset loading and image transformations
Windows OS	Microsoft	11	Operating System