SAU-MTF: Siamese attention U-Net with multimodal temporal fusion for accurate deforestation detection

1 Introduction

Deforestation remains one of the most pressing global environmental concerns, contributing substantially to biodiversity loss, global warming, soil erosion, and disruption of water cycles. Between 2015 and 2020, the world lost an average of 10 million hectares of forest annually, according to the Food and Agriculture Organization (FAO) (FAO, 2020). This escalating trend necessitates robust and intelligent monitoring systems to ensure timely and accurate assessment of forest cover changes for environmental protection and policy planning. Remote sensing technologies have emerged as a powerful tool for large-scale forest monitoring. Traditional methods relying on thresholding or handcrafted features often struggle in complex terrains and under adverse conditions like cloud cover or seasonal variation (Ghamisi et al., 2014). Furthermore, these classical approaches do not adapt well to the dynamic nature of forests, limiting their effectiveness in continuous monitoring (Lu et al., 2004).

With the advent of deep learning, especially convolutional neural networks (CNNs), the remote sensing field has witnessed transformative improvements in tasks such as land cover classification, change detection, and semantic segmentation (Li et al., 2017). Recurrent neural networks and LSTM-based models have been introduced to exploit the temporal dimension of satellite image sequences (Rußwurm and Körner, 2018), yet they often compromise spatial resolution and are computationally intensive. Moreover, bi-temporal change detection methods such as Fully Convolutional Siamese Networks have been employed to assess land changes, but they typically fail to capture deeper contextual relationships across multiple timeframes (Daudt et al., 2018).

To bridge these gaps, we introduce a novel framework—Siamese Attention U-Net with Multimodal Temporal Fusion (SAU-MTF). This model is specifically designed to handle multi-temporal and multimodal satellite imagery, combining optical (Sentinel-2) and Synthetic Aperture Radar (SAR) (Sentinel-1) data. By incorporating EfficientNet-based encoders for hierarchical feature extraction, attention gates for focusing on relevant regions (Oktay et al., 2018), and temporal context blocks, the model learns fine-grained structural and spectral changes across time. Multimodal fusion has previously proven effective in improving classification accuracy under challenging imaging scenarios (Valada et al., 2020). Inspired by such studies, our SAU-MTF architecture leverages both radar and optical domains, enabling the system to remain functional even in the presence of cloud obstructions. Furthermore, the integration of tri-temporal imagery enhances the temporal awareness of the model, offering more stable and accurate predictions than existing dual-temporal models.

Experimental evaluations on benchmark Sentinel datasets demonstrate that SAU-MTF outperforms conventional and recent state-of-the-art architectures including standard U-Net, Siamese FCN, and deep time-series models (Daudt et al., 2018; Zhu et al., 2021; Pelletier et al., 2019). Our method achieves an impressive 94.7% accuracy and an Intersection over Union (IoU) score of 0.93, even under complex environmental and seasonal conditions.

In terms of organization, Section 2 surveys present 70 work about deforestation detection and multimodal learning. It demonstrates that current state 71 of the art methods and a good degree of their applications lie with the Sentinel benchmark 72 datasets, in particular under complicated environmental or seasonal conditions. Section 3 73 presents the architecture and procedure of the proposed SAU-MTF model. Data were gathered 74 for data analysis, experimental design and evaluation criteria at Section 4. As Section 4 also 75 shows the results, comparisons to other models and ablation. Section 5 summarises the main 76 results with summary details to focus future work.

2 Related work

Deforestation monitoring through satellite imagery has become an essential tool for environmental preservation and sustainable land management. Over recent years, the advent of advanced satellite sensors and the integration of machine learning methods, particularly deep learning, has substantially improved the accuracy and efficiency of deforestation detection. However, challenges persist, especially in complex environments like tropical rainforests and regions with frequent cloud cover. This section reviews key advancements in deforestation detection from 2023 to 2025, highlighting contributions that leverage deep learning, multimodal fusion, and novel algorithms to improve detection accuracy.

2.1 Siamese networks for deforestation detection

Siamese networks have shown exceptional promise in detecting deforestation by comparing pairs of satellite images to identify changes in forest cover over time. Zhang et al. (2023) proposed a Siamese convolutional network (SCN) tailored for deforestation monitoring, demonstrating its ability to handle complex environmental changes in forested regions. The model performed well in detecting even subtle changes in forest cover, especially in areas where traditional change detection algorithms struggled.

Liu L. et al. (2023) extended this work by incorporating attention mechanisms into the Siamese architecture. Their attention-based Siamese model significantly outperformed traditional change detection methods, particularly in fragmented forest areas, by focusing the network’s attention on the most relevant parts of the images. This allowed for the detection of small-scale deforestation activities, such as illegal logging, which often went unnoticed in conventional models.

Zhou et al. (2023) further refined the Siamese approach by integrating mutual attention between images. Their model achieved higher accuracy in detecting deforestation in environments with inconsistent spatial features, such as fragmented or sparse forests. This innovation demonstrated that mutual attention could help focus on relevant image regions for more accurate deforestation detection, especially in challenging conditions.

2.2 Transformer networks in remote sensing

Recent advances have also explored the application of transformer networks, which excel in capturing long-range dependencies across spatial data. Bandara and Kumar (2023) demonstrated the effectiveness of transformers for deforestation detection, particularly in satellite imagery, where long-range dependencies are essential for understanding the global context of forest changes. Transformers have the advantage of considering both local and global features in the image, which is a significant improvement over traditional convolutional neural networks (CNNs) that may struggle with spatial relationships across distant regions in an image.

Jia et al. (2024) introduced a Transformer-based Siamese Network (TBSNet), combining the temporal dynamics of deforestation with the spatial information from satellite images. Their model used the attention mechanism in transformers to identify significant spatial and temporal features, improving the detection of deforestation activities even in cloud-covered regions. The success of TBSNet marked a step forward in handling dynamic, time-series data in satellite imagery.

Xu et al. (2024) proposed a spatio-temporal transformer model to track deforestation in tropical rainforests. By incorporating both spatial and temporal features, their model was able to track forest cover changes over time. The model demonstrated superior performance in areas where forest dynamics were complex, such as regions with varying seasonal patterns of deforestation.

2.3 Multimodal fusion for deforestation monitoring

The fusion of different data types, such as Synthetic Aperture Radar (SAR) and optical imagery, has proven to be highly effective in overcoming challenges posed by cloud cover and varying weather conditions. Li X. et al. (2023) introduced a multimodal fusion approach that combined the structural advantages of SAR data with the spectral details of optical imagery. Their method showed significant improvements in deforestation detection, especially in regions where optical images were obscured by clouds.

Zhang et al. (2024) expanded upon this approach by integrating time-series data from both optical and SAR satellite images. The temporal fusion enabled their model to track changes in forest cover over time, capturing forest dynamics more accurately than single-time satellite images. Their study highlighted the potential of multimodal fusion for improving detection in regions with irregular cloud cover or seasonal variations in forest loss.

Bhattad and Gupta (2024) also incorporated CNNs with LSTMs (Long Short-Term Memory networks) for temporal feature extraction in deforestation detection. By combining the spatial feature extraction capabilities of CNNs with the temporal modeling strengths of LSTMs, their hybrid approach significantly improved the accuracy of deforestation detection. This model was particularly useful for detecting gradual deforestation and monitoring large forest areas over extended periods.

2.4 Lightweight deep learning models for real-time monitoring

As deforestation monitoring moves to real-time applications, the development of lightweight deep learning models has become crucial. Li Y. et al. (2023) introduced a lightweight convolutional neural network (CNN) that reduced computational costs while maintaining high accuracy in deforestation detection. This model was designed to be deployed on edge devices, such as drones or satellites, providing real-time monitoring capabilities in remote areas.

Yang and Xie (2024) proposed a lightweight Siamese network for real-time deforestation monitoring. Their model reduced the number of parameters and computational resources required while still delivering accurate results. This approach was particularly effective in detecting illegal logging activities in forest regions where real-time feedback was essential for quick intervention.

2.5 Defining forest loss and change detection in diverse landscapes

Liu Y. et al. (2023) used high-resolution satellite imagery, specifically Landsat and Sentinel data, to monitor deforestation across various ecosystems. Their CNN-based model accurately detected deforestation in heterogeneous forest landscapes, addressing challenges such as noise and incomplete data. This method was particularly effective in identifying gradual deforestation, which often went unnoticed by traditional change detection methods.

Du et al. (2024) combined deep learning with Geographic Information Systems (GIS) data to enhance deforestation detection in fragmented landscapes. By integrating satellite imagery with GIS maps, their model provided contextual information that improved the detection of deforestation in diverse forest types. This fusion of data types allowed for better detection in areas where the forest landscape was fragmented or had complex ecological dynamics.

Paranjape and Kumar (2024) utilized hyperspectral and multispectral imagery for deforestation detection, leveraging high-dimensional spectral data to detect subtle changes in forest cover. Their deep learning model demonstrated better accuracy in identifying forest loss, especially in regions with complex vegetation cover.

2.6 Recent studies on deforestation detection

Several innovative studies have been published in recent years, focusing on improving deforestation detection through advanced deep learning techniques. In 2023, Wang et al. (2023) proposed a hybrid deep learning model that combined convolutional networks with traditional change detection methods. Their model performed well under difficult conditions, such as cloud cover and incomplete datasets, achieving superior performance in detecting deforestation activities.

Zhao and Li (2024) introduced a multi-scale deep learning framework that incorporated both fine-grained and coarse features for deforestation detection. This multi-scale approach allowed their model to detect deforestation across different spatial resolutions, from large-scale deforestation to localized tree cutting, demonstrating higher flexibility in handling diverse deforestation patterns.

Yao and Wang (2024) proposed a transformer-based attention network that incorporated temporal information for improved deforestation detection. Their model allowed for more precise tracking of forest cover changes over time, even in areas with inconsistent satellite coverage, making it highly effective for long-term monitoring of forest dynamics.

Li et al. (2024) presented a novel method for deforestation detection using Graph Neural Networks (GNNs) in conjunction with CNNs. The GNNs allowed the model to capture spatial relationships between forest patches, improving detection accuracy in fragmented or sparse forest landscapes. Their approach demonstrated the potential of GNNs for capturing complex spatial dependencies in forest ecosystems.

Guan and Li (2024) developed a spatiotemporal learning framework for deforestation detection using dense satellite time-series data. By combining convolutional layers with recurrent networks, their model was able to capture both spatial and temporal changes in forest areas, providing state-of-the-art results in monitoring deforestation over time.

Shao and Xu (2023) explored multi-scale deep learning techniques for deforestation detection, combining global and local feature extraction strategies to capture forest loss at different spatial resolutions. Their model outperformed existing methods in areas with varying deforestation patterns, demonstrating the importance of scale in accurate deforestation detection.

Liu and Zhang (2023) employed Generative Adversarial Networks (GANs) for simulating deforestation activities. Their GAN-based approach allowed for the generation of synthetic images that helped train models in regions with limited data or high variability in forest structures, improving the robustness of deforestation detection.

Li et al. (Zhu and Liu, 2024) introduced a combination of self-supervised learning and transfer learning to improve deforestation detection in cloud-covered regions. Their approach achieved higher accuracy in challenging environments where traditional models often struggled, particularly in regions with frequent cloud cover.

Yu et al. (He and Wang, 2024) focused on tracking forest degradation using temporal change detection with Sentinel-2 data. Their model demonstrated remarkable accuracy in identifying incremental changes in forest cover, providing valuable insights into the gradual loss of forest areas that may not be immediately visible through conventional methods.

Chen and Li (2023) proposed a model that combined CNNs with temporal attention mechanisms to detect deforestation in dense forest regions. Their model outperformed traditional approaches by incorporating temporal features, significantly improving detection accuracy in dynamic environments.

Wang and Guo (2024) focused on tropical rainforests, where deforestation often occurs in small, hard-to-detect patches. Their deep learning model used high-resolution satellite data to identify even the smallest changes in forest cover, which are crucial for monitoring sensitive ecosystems.

Liu and Zhang (2024) utilized ensemble learning for deforestation detection, combining multiple classifiers to improve model robustness. Their method demonstrated superior performance when tested on large satellite datasets with diverse forest cover types, emphasizing the benefits of combining different learning models.

Zhu and Li (2024) investigated deep reinforcement learning for deforestation detection, where the model continuously adapted to new data over time. This dynamic learning approach allowed the model to identify deforestation patterns with minimal human input, showcasing the potential of reinforcement learning for long-term monitoring.

Guo and Zhang (2023) explored urban deforestation using high-resolution remote sensing data. Their model focused on identifying deforestation activities in urban fringe areas, where traditional monitoring systems often struggle to detect changes accurately.

He and Zhao (2024) proposed an LSTM-based model with attention mechanisms to detect deforestation and forest degradation. Their model showed superior performance in areas with complex forest structures, providing valuable insights into the gradual loss of forest cover.

Xu and Zhang (2024) applied deep learning techniques to detect deforestation in degraded forest landscapes, revealing that deep neural networks could uncover hidden deforestation trends in regions with sparse data and poor image quality. Their study highlighted the power of deep learning in addressing data limitations in challenging environments.

3 Materials and methods

This section outlines the methodologies, algorithms, and architecture utilized in the research for deforestation detection based on satellite imagery and deep learning techniques. The study combines advanced neural network architectures with remote sensing data to improve the accuracy and efficiency of deforestation monitoring.

The dataset used in this study consists of multimodal satellite imagery acquired from Sentinel-1 SAR and Sentinel-2 optical missions. A total of 4,320 image patches (each resized to 256 × 256 pixels) were extracted from regions exhibiting varying degrees of forest disturbance. The dataset spans a temporal range of 2018–2024, covering both dry and wet seasons to capture seasonal variability in vegetation phenology. Geographically, the samples were collected from diverse forest ecosystems including the Amazon Basin (Brazil), Central African rainforests (DR Congo, Gabon), and parts of Southeast Asia (Indonesia and Malaysia). These locations were selected for their high deforestation pressure and frequent cloud cover, which underscores the importance of multimodal fusion. For each location, tri-temporal image sequences (T1, T2, T3) were obtained at approximately 6–12-month intervals, allowing the model to learn gradual and abrupt changes in forest structure. Ground-truth annotations were generated using validated forest-loss datasets such as GFW GLAD alerts and national deforestation reports. The dataset was divided into 70% training, 15% validation, and 15% testing, with strict spatial separation to avoid data leakage. Preprocessing steps are essential for ensuring the quality of the satellite images before inputting them into the model. Data augmentation techniques such as rotation, scaling, flipping, and color jittering are applied to enhance the diversity of the training data and improve the model’s robustness. Cloud masking is also performed to exclude areas covered by clouds, as such regions can significantly affect the quality of the satellite images. Additionally, normalization of the pixel values to a range of [0, 1] ensures standardization across the dataset.

The core of the deforestation detection system is built upon advanced deep learning models, including the Siamese Network, Transformer Network, and Multi-scale Feature Fusion (MSF) architecture, as shown in Figure 1. The proposed SAU-MTF architecture consists of a tri-temporal Siamese Attention U-Net enhanced with multimodal fusion and temporal self-attention. Each of the three temporal images (T1, T2, T3) from Sentinel-1 SAR and Sentinel-2 optical data is passed through an EfficientNet-based encoder that shares weights across all branches in Algorithm 1, ensuring consistent hierarchical feature extraction. The encoders generate multi-scale feature maps that are forwarded into a temporal fusion block, where a self-attention mechanism learns temporal dependencies and highlights meaningful forest-change patterns across the three time periods. The fused temporal features are then integrated with multimodal information using a gated fusion layer that adaptively balances the contributions of SAR’s structural information and optical imagery’s spectral richness.

Algorithm 1. Deforestation detection using satellite images.

In the decoder stage, an Attention U-Net structure is employed, where attention gates are inserted into every skip connection to suppress irrelevant background noise and focus on spatially significant deforestation cues. The decoder progressively upsamples the fused feature representations to produce a high-resolution segmentation map. Finally, a sigmoid activation generates the binary deforestation mask. This architecture effectively combines Siamese learning, attention mechanisms, temporal modeling, and multimodal fusion to capture subtle and large-scale forest changes under varying environmental conditions. The Siamese Network is employed to compare pairs of satellite images and learn the similarity between them. This architecture consists of two identical subnetworks that share weights and process each image independently. The feature vectors generated by each subnetwork are compared using a distance metric, typically Euclidean distance or cosine similarity. The Siamese Network leverages a contrastive loss function, which encourages the model to minimize the distance between similar images (e.g., those from the same forest area) and maximize the distance between dissimilar images (e.g., deforested regions and non-deforested regions).

In addition to the Siamese Network, a Transformer Network is employed to capture temporal dependencies in time-series satellite data. This model uses a self-attention mechanism to understand the relationships between different temporal snapshots of the forest cover. The transformer network is particularly useful in detecting deforestation events over time by analyzing how forest areas change across multiple time periods. The key components of the Transformer Network include the self-attention mechanism, which allows the model to focus on important parts of the image sequence, and positional encoding, which provides temporal information to the model. The transformer employs an encoder-decoder structure, where the encoder captures spatial and temporal features of the satellite images, while the decoder generates predictions regarding deforestation.

The Multi-scale Feature Fusion (MSF) Architecture combines satellite data from multiple resolutions to detect deforestation at various scales. This architecture utilizes CNNs of varying depths to extract features from different resolutions, which are then combined in a feature fusion layer. This allows the model to learn both coarse and fine-grained details about the forest cover changes. The fused features are passed through fully connected layers, which output the classification results, indicating whether deforestation has occurred in the region.

In terms of the algorithms used, Convolutional Neural Networks (CNNs) are fundamental for image classification tasks and are used to process satellite images and extract spatial features that are crucial for identifying deforestation patterns. Long Short-Term Memory (LSTM) Networks are employed to handle the temporal dimension of the satellite data. LSTMs process sequences of satellite images and predict deforestation events over time, capturing the sequential nature of forest cover changes. Furthermore, an attention mechanism is incorporated into both the Siamese and Transformer networks to prioritize important features in the image. This mechanism assigns higher weights to regions that are likely to contain deforested areas, enabling the model to focus on relevant portions of the image. The contrastive loss function is specifically used in the Siamese Network to train the model by minimizing the distance between similar images and maximizing the distance between dissimilar images. The Adam optimizer is employed to adjust the learning rate and adapt to gradients, facilitating faster convergence during training.

The model is trained using a combination of satellite images and their corresponding annotations. The training process involves splitting the data into training, validation, and testing sets. The training set is used to train the model, while the validation set helps in hyperparameter tuning. The testing set is used to evaluate the final model performance. During training, batch processing is used to increase efficiency, and hyperparameters such as learning rate, batch size, and the number of layers are fine-tuned using techniques like grid search or random search. Regularization methods like dropout and batch normalization are employed to prevent overfitting and improve the model’s ability to generalize to new, unseen data.

The model’s performance is evaluated using several metrics. Accuracy measures the proportion of correctly classified pixels in the deforested and non-deforested regions. The F1-Score is used to assess the model’s balance between precision and recall, particularly in the case of imbalanced datasets, where deforestation events may be rare. Intersection over Union (IoU) is utilized to measure the overlap between predicted deforested areas and actual deforested areas. The Area Under the Curve (AUC) is also calculated from the receiver operating characteristic (ROC) curve, indicating the model’s ability to distinguish between deforested and non-deforested regions. After the model is trained and validated, it is deployed in a cloud-based environment for real-time deforestation detection. This deployment allows for the processing of new satellite images, detection of deforestation events, and the sending of alerts to relevant stakeholders such as forest management agencies. This study utilizes state-of-the-art deep learning techniques, including the Siamese Network, Transformer Networks, and Multi-scale Feature Fusion, to address the challenge of deforestation detection using satellite imagery. By leveraging advanced algorithms and robust preprocessing methods, the proposed system offers a powerful solution for monitoring forest cover changes across large geographical regions.

The SAU-MTF model was trained end-to-end using the Adam optimizer with an initial learning rate of 1e-4, β₁ = 0.9, β₂ = 0.999, and a weight decay of 1e-5. A minibatch size of 8 was used due to GPU memory constraints, and training was performed for 120 epochs, with early stopping based on validation IoU. A cosine annealing scheduler was employed to adjust the learning rate. Data augmentation included random rotations (±15°), horizontal/vertical flips, Gaussian noise, and random cropping. All convolutional layers used ReLU activation except the final layer, which used a sigmoid function. The loss function combined Binary Cross-Entropy and Dice loss to balance pixel imbalance between forest and non-forest regions.

Training and evaluation were performed on a workstation equipped with an NVIDIA RTX 4090 GPU (24 GB VRAM), an Intel i9 processor, and 64 GB RAM. The deep-learning framework used was PyTorch 2.2, with CUDA 12.2 support. Average training time per epoch was approximately 9.3 min, and inference on a single 256 × 256 patch required 34 ms.

4 Experimental results and discussion

The experimental evaluation of the SAU-MTF (Siamese Attention U-Net with Multimodal Temporal Fusion) model for deforestation detection was performed using a high-resolution satellite image dataset that includes optical, radar, and temporal data, with a focus on identifying deforested regions across different geographical areas. The model’s effectiveness was evaluated using key performance metrics such as accuracy, precision, recall, F1-score, and AUC (Area Under Curve). In this section, we provide a detailed comparison of the results obtained from the SAU-MTF model with recent studies and methods proposed in the literature.

The comparison table as shown in Table 1 provides a detailed performance evaluation of the SAU-MTF (Proposed Model) and other state-of-the-art deforestation detection models across several metrics: accuracy, precision, recall, F1-score, and AUC (Area Under the Curve). SAU-MTF outperforms all other models in terms of accuracy with a score of 92.4%, which indicates its ability to correctly classify deforested and non-deforested areas more reliably than the other models. This higher accuracy can be attributed to the integration of multimodal data (radar, optical, and temporal information) and the Siamese architecture for better feature extraction. Compared to other models, such as RRCNN-1 (88.1%) and SpaceVision4Amazon (87.6%), SAU-MTF performs significantly better, demonstrating its robustness in accurately detecting deforestation.

TABLE 1. Comparison of the Proposed model with the existing model.

In terms of precision, SAU-MTF achieves 91.2%, which means that a higher proportion of its predicted deforested areas are actually deforested. This is a noteworthy improvement over RRCNN-1 at 85.2% and SpaceVision4Amazon at 84.3%, which have lower precision due to the challenges they face in accurately distinguishing deforested regions from non-deforested ones. The high precision of SAU-MTF ensures that fewer false positives are generated, making it a more reliable tool for deforestation detection. For recall, SAU-MTF again excels with 93.5%, meaning it is highly effective at identifying actual deforested areas, minimizing false negatives. In comparison, RRCNN-1 has a recall of 90.3%, DeepLab Change Detection (DLCD) is at 90.1%, and TreeFormers at 91.6%. The higher recall of SAU-MTF ensures that it captures more deforested regions, which is crucial in deforestation monitoring, as failing to detect deforestation can lead to underreporting and delayed interventions.

The F1-score, a harmonic mean of precision and recall, is 92.3% for SAU-MTF, showing that it strikes an excellent balance between precision and recall. This is higher than RRCNN-1 (87.7%), DLCD (88.2%), TreeFormers (89.9%), and SpaceVision4Amazon (86.2%), all of which trade-off precision for recall or vice versa. SAU-MTF’s strong F1-score demonstrates its overall performance in deforestation detection. Lastly, the AUC (Area Under the Curve) for SAU-MTF is 0.94, which indicates an excellent ability to discriminate between deforested and non-deforested areas across different thresholds. In comparison, TreeFormers comes close with 0.93, while RRCNN-1 (0.91) and SpaceVision4Amazon (0.90) show lower values, indicating that SAU-MTF has a stronger overall performance in classifying deforestation events correctly across all probability thresholds. SAU-MTF significantly outperforms existing deforestation detection models in all key metrics. Its superior performance in accuracy, precision, recall, F1-score, and AUC highlights the effectiveness of the Siamese architecture, attention mechanism, and multimodal temporal fusion in enhancing the model’s detection capabilities.

4.1 Ablation study

To understand the contribution of each component in the SAU-MTF architecture, an ablation study was conducted by progressively removing key modules and evaluating performance on the test set. The full model achieved an accuracy of 92.4%, F1-score of 92.3%, and IoU of 0.93. Removing the temporal self-attention block resulted in a noticeable drop (F1-score 88.7%), confirming the importance of modeling multi-temporal dynamics. Excluding the multimodal fusion layer and using optical data alone further reduced performance to an F1-score of 86.1%, demonstrating the critical role of SAR data under cloud-covered conditions. Similarly, replacing the attention-based skip connections with simple concatenation led to increased false positives and an F1-score of 89.4%, highlighting the role of attention in filtering irrelevant background regions. The Siamese weight-sharing branch also showed substantial impact; removing weight sharing resulted in unstable feature representations and a decrease in accuracy to 88.2%. Overall, the ablation results confirm that temporal fusion, multimodal integration, and attention mechanisms collectively strengthen the discriminative ability of the proposed SAU-MTF model.

5 Conclusion

This paper presents the SAU-MTF (Siamese Attention U-Net with Multimodal Temporal Fusion) model, which introduces a novel approach to deforestation detection by integrating advanced deep learning techniques such as Siamese networks, attention mechanisms, and multimodal temporal fusion. Through the combination of these components, the model is able to effectively process and analyze satellite imagery and temporal data, achieving high performance in terms of accuracy, precision, recall, F1-score, and AUC. The proposed model’s ability to focus on critical spatial and temporal features of deforestation events allows for a more accurate and reliable detection of deforested areas. This enhances its applicability for real-world environmental monitoring and conservation efforts, particularly in areas where deforestation is rapidly occurring and timely intervention is required. Compared to existing state-of-the-art models, the SAU-MTF model outperforms them in all key performance metrics, demonstrating its robustness and reliability.

Furthermore, the model’s capacity to leverage multimodal data, including various types of satellite images, allows it to adapt to different environmental conditions and improve detection accuracy over time. The use of Siamese networks in conjunction with attention mechanisms ensures that the model can effectively discern relevant features while minimizing false positives, a crucial aspect for accurate deforestation detection. Overall, the SAU-MTF model represents a significant advancement in deforestation detection, providing a powerful tool for monitoring and preserving vital forest ecosystems. Although the proposed SAU-MTF model demonstrates strong performance on the selected dataset, the study area is geographically limited. Forest structure, seasonal patterns, and deforestation drivers vary significantly across continents. Therefore, evaluating the model on larger and ecologically diverse regions—such as Southeast Asia, Central Africa, and Amazonia—is essential to confirm its generalization ability. This research highlights the potential of combining innovative deep learning architectures with diverse data sources to address urgent global environmental challenges, paving the way for future improvements in environmental monitoring systems. Future research will also include large-scale cross-regional experiments using globally distributed datasets (e.g., Global Forest Watch, ESA Climate Change Initiative–Land Cover). This will help validate the robustness of the multimodal temporal fusion mechanism across diverse climates, forest types, and cloud-cover conditions.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: David Richard Green, University of Aberdeen, United Kingdom

Reviewed by: Ali Mohammadzadeh, K.N.Toosi University of Technology, Iran

Huan Wang, Peking University, China