HieraChange: A Hierarchical Vision Transformer Model for Change Detection in Satellite Images

Song et al. [1] introduced an Object-Based Image Analysis (OBIA) approach that proves more effective than traditional pixel-level methods for identifying changes in very high-resolution (VHR) satellite imagery. Their work presents a DL-driven OBIA framework that incorporates object-level uncertainty estimation to perform CD without relying on ground truth labels. The method begins by generating change objects using unsupervised techniques, which are then refined through a deep network employing 3D convolutions and ConvLSTMs. Object labels are iteratively updated based on uncertainty measures until all regions are confidently classified. Experimental results demonstrate that this approach outperforms conventional methods, offering strong noise resistance and precise object-level change identification.

Siddique and Ahmed [2] introduced a DL-based coherent change detection model (CCD-Conv1D) using Sentinel-1 SAR images to monitor floods and predict affected areas. The model combines image segmentation, log-ratio analysis, and DL to generate accurate flood maps. Experiments in Ayodhya and Basti showed improved CD accuracy and significant inundation over vegetation areas. The DL-based forecasts using Conv1D and Naïve Forecast models proved more reliable than traditional methods, supporting future development of effective flood monitoring systems.

Zhu et al. [3] emphasized that high spatial resolution (HSR) RS images are crucial for examining land-use transitions, yet many traditional methods fail to integrate semantic class information. To overcome this limitation, a new Siamese Global Learning (Siam-GL) framework is introduced, integrating a shared-parameter Siamese network with a global hierarchical sampling strategy to address sample imbalance and improve semantic CD. Experimental results across several HSR datasets demonstrate that Siam-GL delivers higher accuracy and stronger generalization capabilities than existing techniques.

Xiao et al. [4] addressed multi-spatial-resolution CD, which predicts the differences between image pairs of varying resolutions by aligning them and using unsupervised deep feature learning. The proposed framework uses co-registration, denoising autoencoders, and a mapping network to bridge resolution gaps and generate accurate change maps, showing superior results across four real datasets.

Farooq and Manocha [5] discussed that satellite-based CD is crucial for monitoring urban growth, land use, disasters, and environmental changes using multi-temporal RS data. DL methods have shown superior performance over traditional techniques, though some challenges remain. The review also highlights AI-based approaches and outlines future directions for advancing urban CD research.

Patil et al. [6] presented EffCDNet, which integrates a Siamese EfficientNet encoder with an Attention-UNet decoder and UDWT-based post-processing to improve feature extraction and produce noise-resistant change maps. Results on benchmark datasets demonstrate that EffCDNet surpasses existing approaches in both accuracy (IoU) and inference speed.

Gite and Gupta [7] proposed a novel approach that integrates Fuzzy Neural Network (FNN) classification with segmentation performed using a Taylor Shuffled Shepherd Optimization (TSSO)-based GAN. The TSSO method combines the Taylor series with Shuffled Shepherd Optimization for more effective feature learning. Experimental results confirm that the method performs highly with 0.932 accuracy, 0.0704 error rate, and 0.911 kappa coefficient. Compared to KPCA-MNet, MSVM, PPCNET, and IFN, the approach improves accuracy by 2.28%, 4.78%, 0.33%, and 13.26%, respectively, proving its superiority.

Dong et al. [8] introduced EfficientCD, a DL framework that employs EfficientNet for feature extraction along with a novel ChangeFPN module to enhance bi-temporal feature interaction. The model also incorporates a progressive upsampling module that uses Euclidean distance to improve feature fusion and reconstruction. Experiments on the LEVIR-CD, SYSU-CD, CLCD, and WHU-CD datasets indicate that EfficientCD achieves state-of-the-art performance in CD.

Chughtai et al. [9] compared traditional and modern approaches, highlighting the effectiveness of post-classification techniques. Among them, the Maximum Likelihood Classifier (MLC) consistently delivers the highest accuracy across diverse regions and remains the most reliable method for LULC CD.

Lim et al. [10] proposed a CNN-based CD method for high-resolution satellite imagery that removes the need for preprocessing steps such as ortho-rectification or classification. One of the major challenges in multi-temporal image analysis is separating genuine changes from variations in viewpoint or color, particularly in urban areas where registration errors from tall buildings often undermine traditional techniques. To overcome this, the authors designed three encoder-decoder CNNs capable of generating change maps directly from RGB image pairs. They also built a large, fully annotated dataset from Google Earth images spanning different years and seasons to enable supervised training. Experimental results demonstrate that the CNN models can reliably identify true changes despite imperfect alignment, and their ensemble outperforms each individual model.

Afaq and Manocha [11] provided a comprehensive review of digital CD techniques, noting that change vector analysis delivers the highest accuracy among algebra-based methods, while discrete wavelet transformation outperforms other transformation-based approaches. The study also highlights that neural networks and fuzzy-based techniques generally surpass traditional methods, and although DL shows strong potential, continued advancements are still needed within the RS field.

Satellite imagery received at ground stations is frequently degraded by noise such as impulse, speckle, and Gaussian noise. Conventional denoising techniques can reduce noise but often blur edges, resulting in a loss of important details. To overcome this issue, Asokan and Anitha [12] developed an optimized bilateral filter enhanced with nature-inspired optimization algorithms. The filter’s parameters are fine-tuned using Particle Swarm Optimization (PSO), Cuckoo Search (CS), and Adaptive Cuckoo Search (ACS). Experimental results indicate that the ACS-optimized bilateral filter outperforms traditional methods, including Median and RGB spatial filters, across metrics such as PSNR, MSE, FSIM, entropy, and processing time.

Feizizadeh et al. [13] proposed an OBIA method for landslide mapping and CD using multi-temporal satellite imagery. The approach integrates spatial and spectral features, including shape, texture (via GLCM), and fuzzy logic after initial image segmentation. Multi-resolution segmentation was applied to IRS-1D, SPOT-5, and ALOS datasets, incorporating DEM-derived slope and flow direction information. Classification was performed using fuzzy membership functions and operators, with Fuzzy Synthetic Evaluation (FSE) used for validation. The AND operator produced the highest accuracy (93.87% in 2005 and 94.74% in 2011), demonstrating the effectiveness of object-based CD in monitoring landslide evolution in northern Iran.

Zhang et al. [14] proposed GAS-Net, a global-aware Siamese network developed to model comprehensive contextual relationships between scenes and foreground regions. The framework integrates a Global-Attention Module (GAM) and a Foreground-Awareness Module (FAM), enhancing contextual reasoning and joint feature learning. Experiments conducted on the LEVIR-CD and Lebedev-CD datasets demonstrate strong robustness, yielding F1-scores of 91.21% and 95.84%, respectively.

Liu et al. [15] proposed a hierarchical transformer architecture that leverages self-attention to efficiently capture long-range dependencies in multi-temporal imagery. Local windowed self-attention is used to lower computational cost while improving feature representation. The extracted features are fused and processed through a nested U-Net, with an additional model fusion strategy further enhancing accuracy. Evaluation on the LEVIR-CD and SYSU-CD datasets demonstrates superior performance compared to existing methods.

Priyadarshi and Singh [16] presented a fully convolutional Siamese auto-encoder framework that produces change maps from bi-temporal, co-registered satellite imagery. The method is applied to detect land-cover changes such as encroachment, building expansion, and unauthorized construction. Due to the absence of a public multi-label CD dataset, a new hand-annotated dataset was created. The proposed approach achieves a 3.9% improvement in F1-score over state-of-the-art methods and significantly reduces training requirements, needing only 50 hours for any number of labels, compared to 30 hours per label in previous CNN-based techniques.

Wang et al. [17] introduced a Spatiotemporal Enhanced CD (STECD) algorithm for monitoring flood changes using multi-source heterogeneous earth observation image time series (EOITS). The method combines spatial clustering, where a sparse Markov random field (SMRF) optimizes flood boundaries using contextual spatial features, with temporal enhancement, which leverages historical flood information to reduce errors caused by terrain, cloud, and topographic shadows. By jointly exploiting spatial and temporal dependencies, the STECD algorithm improves the accuracy of flood area detection. Experimental results on real EOITS datasets confirm that STECD outperforms widely used CD methods. Overall, it provides a more robust solution for flood hazard monitoring.

Cheng et al. [18] proposed a Multi-scale Cross-modal Fusion Network (MCFNet) that leverages style transfer to integrate content and style features across multiple scales for improved fusion. Experimental evaluations on three datasets demonstrate that MCFNet surpasses existing approaches, achieving at least a 4% improvement in Kappa accuracy.

Recent research shows that translating multimodal RS images into a common domain can reduce incompatibility issues in CD. However, most methods ignore global context, causing semantic confusion and reduced accuracy. To overcome this, a pseudo-siamese GAN (PS-GAN) is proposed by Zhu et al. [19], which extracts all global and local features collectively via a two-branch encoder and fuses them using an adaptive multi-scale module, ensuring structurally intact translations. A U2Net+-based CD model with gated feature modulation and enhanced squeeze-excitation further improves semantic feature representation. Experiments on four datasets show PS-GAN outperforms 11 state-of-the-art multimodal CD methods with higher accuracy.

Heterogeneous RS with SAR and optical data enables reliable anytime, all-season CD. To address feature loss and noise in existing two-stage methods, a domain adaptation-based multi-source CD network (DA-MSCDNet) is proposed by Zhang et al. [20], aligning deep feature spaces across heterogeneous data for improved detection. Experiments on multiple datasets, including Sentinel, Landsat, QuickBird, and COSMO-SkyMed, show DA-MSCDNet outperforms six existing methods with higher accuracy (F1-score 82.58%) and 10× faster training efficiency.

Zhang et al. [21] introduced a bipartite unsupervised residual network (ResNet) for CD using heterogeneous radar and optical images captured at different times. Unlike traditional shallow pixel-level methods, the proposed approach operates at the feature level to better capture differences. The bipartite deep ResNet fully leverages joint multi-temporal information, generating representative features of image changes. A novel loss function is designed to optimize the ResNet training process. Experiments on both heterogeneous and homogeneous datasets show that the method outperforms existing CD approaches.

CD in heterogeneous RS imagery is crucial for accurate disaster damage evaluation, yet current homogeneous transformation techniques often degrade semantic information because they rely solely on low-level feature representations. To address this, a Deep Homogeneous Feature Fusion (DHFF) model based on image style transfer (IST) is proposed by Jiang et al. [22], which separates semantic content and style features for accurate transformation. An iterative IST strategy further enhances feature homogeneity across new subspaces, improving detection performance. Experiments on SAR and optical satellite images show that DHFF significantly improves accuracy and Kappa index compared to existing methods.

CD using RS images is essential for monitoring surface changes, but existing Siamese networks often lack sufficient contextual information, leading to false or missed detections and poor edge refinement. To overcome this, SSANet was proposed by Jiang et al. [23]. It is a spatial-temporal attention network that is self-weighted with a combined learning framework of different subnetworks that include fusion, difference, and decoding. The fusion subnetwork contains the MCA module that aggregates multiscale semantic context, while the difference subnetwork, which contains the FDR module, extracts and reconstructs temporal change features. An AW module in the decoder balances these features, enabling SSANet to deliver superior change maps and achieve state-of-the-art performance.

Wang et al. [24] introduced an unsupervised CD method tailored for optical sensors, leveraging a fusion framework that is robust. By augmenting a pre-trained optical image fusion network with a parallel network of identical architecture and integrating both into an adversarial learning framework, the method facilitates effective CD. Experimental comparisons show that the proposed method is both versatile and superior to existing state-of-the-art techniques.

Li et al. [25] proposed a spatially self-paced convolutional network (SSPCN) for unsupervised CD. Pseudo labels are first generated, and self-paced learning (SPL) dynamically selects reliable samples, starting with easier ones and gradually incorporating complex ones. The model assigns adaptive weights to samples, updating them during training to improve robustness and convergence. By integrating spatial information, SSPCN enhances stability against noisy data, and experiments on four heterogeneous datasets demonstrate its superior effectiveness.

Tong et al. [26] presented a novel method for detecting multiple types of changes in bi-temporal hyperspectral images using only a small set of source-image samples. The workflow consists of four key stages: performing unsupervised CD through uncertain area analysis, classifying the source image via active learning, classifying the target image using an enhanced transfer learning strategy, and finally generating a multiclass change map. Experiments on both synthetic and real datasets demonstrate substantial improvements in binary and multiclass CD performance, surpassing state-of-the-art methods.

Sun et al. [27] proposed an unsupervised CD framework built on the patch similarity graph matrix (PSGM), which assumes that the graph structures of homogeneous or heterogeneous image pairs remain stable in unchanged regions. The PSGM for the first image is constructed using the self-expressive property, where each patch is represented as a graph vertex. Change intensities are then derived by assessing how closely the second image adheres to this learned structure. A sparse prior is applied to further refine the generated change map, and experiments on multiple datasets confirm the approach’s robustness and strong effectiveness.

Sun et al. [28] proposed a structure consistency-based CD method that compares image structures instead of pixel values, offering robustness to noise, illumination, and modality differences. The approach uses an advanced nonlocal patch-based graph (NLPG) for structural comparison while avoiding data leakage. It is validated across homogeneous and heterogeneous CD, including the rarely explored multichannel SAR data. Experiments on six scenarios with 12 datasets demonstrate its robustness and accuracy as a unified CD framework.

Sun et al. [29] proposed a robust graph-mapping-based framework for unsupervised heterogeneous CD in RS images. Because heterogeneous image pairs still share modality-invariant structural information of ground objects, the method builds K-nearest neighbor graphs for each image and compares them via graph mapping to produce forward and backward difference images. These different images are then integrated using a Markovian co-segmentation model with co-graph cut to accurately identify changed regions. An iterative feedback mechanism further refines the process by updating graph construction to minimize the influence of changed neighbors. Experimental results across multiple datasets confirm the effectiveness and stability of the proposed approach.

Unsupervised CD is often limited to optical images from sensors with the same spectral and spatial resolution in RS, but this is not always feasible in real-world scenarios like emergencies or defense. Existing methods typically handle only specific cases of resolution disparity, often with resampling-based preprocessing that can discard useful information. Ferraris et al. [30] proposed a robust fusion-based framework that models both observed images as degraded versions of latent high-resolution images, differing only in sparse regions where changes occur. The approach formulates CD as an inverse problem solved via alternating minimization, demonstrating improved accuracy and versatility over state-of-the-art methods on real multi-band optical datasets.

HieraChange is an advanced model designed for detecting infrastructure changes in satellite imagery. By combining a hierarchical transformer encoder with a lightweight MLP decoder, it effectively captures both fine-grained and large-scale patterns. The Siamese architecture in the model enables precise identification of meaningful differences in the images. The detailed architectural description is as follows:

At every stage of the encoder, the respective feature maps from both images are sent to a difference module. These modules compute differences at multiple resolutions, capturing both fine-grained and large-scale changes. This multi-level differencing helps the model identify changes in complex scenes, even when the spatial resolution varies or the changes are subtle.

All the different features are then processed by a decoder that uses MLP transformations and multi-scale upsampling to gradually restore full spatial resolution. Finally, a classifier produces an H × W change map, identifying which pixels have changed between the two time steps. This combination of multi-level transformer features, difference modules, and upsampling enables accurate and detailed CD.

The transformer block, as shown in Figure 2, begins with a Sequence Reduction step, which compresses the input token sequence to a shorter one. This reduces computational cost by lowering the number of tokens processed in attention. After that, the compact sequence is passed through a Multi-Headed Self-Reduction module, a variation of multi-head self-attention that operates on the reduced sequence. This step allows the model to capture global contextual relationships among tokens while maintaining efficiency.

Figure 2. Transformer block

The output of the reduction modules enters a block that integrates positional encoding, ensuring the model retains information about spatial order. Inside this block, the features pass through an MLP layer, followed by a depth-wise convolution. The depth-wise convolution enhances local spatial information-something traditional Transformers lack-making the representation more suitable for vision tasks. Another MLP further refines the representation, allowing the model to combine features that are local as well as global effectively.

The MLP upsampling module, as shown in Figure 3, takes a multi-scale difference feature $F_{\text {diff}}^i$ of size $\frac{H}{2^{i+1}} \times \frac{W}{2^{i+1}} \times C_i$ and first passes it through an MLP to unify or project the channel dimension to a fixed size CCC. This transformed feature is then fed into an upsampling layer, which increases its spatial resolution from $\frac{H}{2^{i+1}} \times \frac{W}{2^{i+1}}$ to the target resolution $\frac{H}{4} \times \frac{W}{4}$. The final output, $\widehat{F_{\text {diff }}^{\imath}}$, is therefore a spatially enlarged and channel-aligned version of the original difference feature, allowing multi-level features from different scales to be fused consistently in the decoder. 

Figure 3. Multilayer perceptron (MLP) upsampling process

4.1.1 Feature extraction

Feature extraction refers to the process of converting raw data into informative representations that emphasize key patterns and characteristics. In DL models, this is typically achieved through convolutional layers, transformers, or specialized encoders that capture spatial, structural, or contextual information. Effective feature extraction reduces data complexity while retaining essential details for accurate predictions. It plays a crucial role in tasks like image recognition, CD, and object tracking. Well-extracted features enhance model performance by improving generalization and reducing noise. At each stage ‘l’, the encoder produces feature maps as shown in Eqs. (1) and (2):

$F_{\text {pre}}^l=$Transformer Block ${ }^l\left(I_{\text {pre}}\right)$   (1)

$F_{\text {post}}^l=$Transformer Block ${ }^l\left(I_{\text {post}}\right)$ (2)

4.1.2 Self-attention mechanism

The self-attention mechanism is a fundamental component of transformer models that allows each input element to evaluate and weight its relevance to every other element in the sequence. It calculates attention scores using query, key, and value matrices, allowing the transformer to focus on the most relevant regions of the input. This mechanism captures long-range relationships and complex patterns, improving the model’s contextual understanding. By continuously adjusting attention weights, it enhances performance across tasks such as natural language processing, image interpretation, and CD. Its flexibility makes self-attention well-suited for handling variable-length inputs and diverse data types. In each transformer block, the self-attention mechanism computation is carried out in Eq. (3):

$\operatorname{Attention}(Q, K, V)=\operatorname{softmax}\left(\frac{Q K^T}{\sqrt{d_k}}\right) V$        (3)

where,

$Q=W t . X$ is the query;

$K=W t . X$ is the key;

$V=W t . X$ is the value;

$d$ is the dimension of the key vector.

4.1.3 Fusion of pre- and post-change features

The fusion of pre- and post-change features combines information from both time-stamped data to enhance CD accuracy. Techniques such as concatenation, element-wise addition, or attention mechanisms are used to merge features, highlighting meaningful differences. This fusion process strengthens the ability of the model to distinguish the changes from noise and irrelevant variations. This captures structural differences between the two images at each hierarchical level as stated in Eq. (4):

$F_{\text {diff }}^l=F_{\text {pre }}^l-F_{\text {post }}^l$ (4)

4.2 Feature aggregation

Feature aggregation is the process of combining multiple feature maps to create a richer and more informative representation. Methods like summation, concatenation, or weighted averaging are commonly used to combine the features from various layers or scales. This enhances the model’s ability to capture diverse patterns and improves overall performance. The extracted features are aggregated across multiple scales as shown in Eq. (5):

$F_{a g g}=\operatorname{Concat}\left(F_{d i f f}^1, F_{d i f f}^2, \ldots ., F_{d i f f}^L\right)$  (5)

where, $L$ is the total number of hierarchical stages and Concat(.) is the channel-wise concatenation.

4.3 Lightweight multilayer perceptron decoder

A lightweight MLP decoder is a simplified yet effective module designed to refine extracted features and generate precise output predictions. It typically consists of a few fully connected layers with minimal parameters, ensuring fast computation and reduced memory usage. Despite its compact design, it efficiently maps complex feature representations to output spaces like segmentation masks or change maps. The lightweight structure makes it ideal for real-time applications and deployment on resource-constrained devices. Its balance between efficiency and accuracy enhances model performance in tasks like CD and image analysis. The decoder takes the aggregated features and processes them through several fully connected layers to predict change maps as shown in Eq. (6).

$Y=\sigma\left(W_2 \cdot \operatorname{ReLU}\left(W_1 F_{a g g}+b_1\right)+b_2\right)$     (6)

where, $Y$ is the predicted output, $W_1$ and $W_2$ are the weights, $b_1$ and $b_2$ are the biases, $\sigma$ is the sigmoid activation function for binary CD. Finally, the aggregated features are reconstructed as:

$F_{\text {dec }}=M L P\left(F_{\text {agg}}\right)$

where, MLP(.) denotes the lightweight ‘multilayer perceptron’ decoder.

4.4 Loss function

A loss function is a mathematical tool used to measure how far a model’s predictions deviate from the true target values. It plays a crucial role in optimization by supplying feedback that helps adjust the model’s parameters. In this work, a combination of loss functions is employed during training to achieve precise and reliable CD.

4.4.1 Binary Cross-Entropy loss

BCE loss is widely used for binary classification problems. It measures how much the predicted probabilities differ from the actual labels. The loss is computed using the following formula in Eq. (7):

$L_{B C E}=-\frac{1}{N} \sum_{i=1}^N\left[y_i \log \left(y_i\right)+\left(1-y_i\right) \log \left(1-y_i\right)\right]$     (7)

This loss effectively penalizes incorrect predictions, encouraging the model to produce confident and accurate outputs.

4.4.2 Dice loss (for handling class imbalance)

Dice loss is a popular loss function in image segmentation, used to evaluate how well the predicted mask matches the ground-truth mask. It is based on the Dice Similarity Coefficient (DSC) and is expressed in Eq. (8) as:

$L_{\text {Dice }}=1-\frac{2 \sum y_i \widehat{y}_l+\epsilon}{\sum y_i \widehat{y}_l+\epsilon}$      (8)

4.4.3 Total loss

The total loss is often a weighted combination, which is given in Eq. (9):

$L_{\text {Total }}=\alpha L_{B C E}+\beta L_{\text {Dice }}$   (9)

where, $\alpha$ and $\beta$ are hyperparameters balancing the two losses.

4.4.4 Final prediction

The binary change mask is obtained as stated in Eq. (10):

$M(x)= \begin{cases}1, & P(x) \geq 0.5 \\ 0, & \text { otherwise }\end{cases}$      (10)

The encoder used in the proposed HieraChange framework is based on the standard Swin Transformer architecture with a shared-weight Siamese configuration for processing pre-change and post-change satellite images. The encoder comprises four hierarchical stages, each consisting of alternating Window-based Multi-Head Self-Attention (W-MSA) and Shifted Window Multi-Head Self-Attention (SW-MSA) blocks, followed by Patch Merging layers for progressive spatial downsampling and hierarchical feature learning. The model accepts an input image of size 256 × 256 × 3, which is partitioned into 4 × 4 non-overlapping patches before feature extraction. Multi-scale feature maps are generated at 1/4, 1/8, 1/16, and 1/32 of the original image resolution with progressively increasing channel dimensions, enabling the extraction of both local and global contextual information. The shared Siamese encoder processes the bi-temporal images using identical network weights to ensure consistent feature representation. The extracted multi-scale features are subsequently passed to the proposed hierarchical differencing module, where temporal differences are computed and aggregated across all encoder stages before being reconstructed using a lightweight MLP decoder.

The proposed HieraChange framework demonstrates effective performance for binary CD in new buildings using high-resolution bi-temporal satellite imagery. By integrating a shared-weight Siamese Swin Transformer encoder, hierarchical feature differencing, multi-scale feature aggregation, an MLP-based decoder, and a hybrid BCE–Dice loss function, the proposed HieraChange framework effectively detects the building changes in the LEVIR-CD dataset. Experimental results achieved a Precision of 97.56%, Recall of 96.12%, F1-score of 96.23%, and IoU of 92.70%, indicating competitive performance for building CD.

The present study is limited to evaluation on the LEVIR-CD benchmark dataset, and therefore the generalization of the proposed framework to other datasets and RS scenarios has not been investigated. Future work will focus on cross-dataset validation, comprehensive computational efficiency analysis, ablation studies, and the extension of the proposed framework to broader RS applications, including urban monitoring and disaster assessment.