Scholarlink

<p>Accurate monitoring of tropical deforestation is critical for climate change mitigation, biodiversity conservation, and sustainable land management. However, many remote sensing approaches overlook seasonal variability, which can significantly affect classification accuracy in humid tropical environments due to phenological changes and atmospheric conditions. To address this limitation, this study proposes the Deep Learning Spatio-Temporal Multi-Temporal Degradation (DL-SMTD) framework, which explicitly incorporates seasonal dynamics while integrating multi-sensor satellite data. The framework combines high-resolution Planet NICFI optical imagery and Sentinel-1 SAR data to improve the detection of deforestation and forest degradation. Two models were developed and compared: a deep learning U-Net architecture and a traditional Random Forest (RF) classifier. Both models were trained and evaluated using labeled image tiles (256 × 256 pixels) derived from Nigeria’s Cross River tropical rainforest, covering six bi-annual seasonal composites between 2020 and 2023. Each composite contains seven spectral features, including four optical bands from Planet NICFI (blue, green, red, and near-infrared) and three SAR features from Sentinel-1 (VV, VH, and VV/VH). Model performance was evaluated using different input configurations, including single-sensor datasets and multi-sensor data fusion. Results indicate that integrating optical and SAR data improves classification performance compared with single-sensor approaches. The U-Net model using fused Planet and SAR data achieved an overall accuracy of 0.93 and an Intersection over Union (IoU) of 0.90, while the RF model achieved 0.92 accuracy and 0.89 IoU with the same input configuration. Building on these results, the proposed DL-SMTD framework incorporates temporal consistency constraints and degradation detection rules to account for seasonal variability, further improving performance to 0.98 overall accuracy and 0.91 IoU. Beyond classification, DL-SMTD generates multi-temporal forest change products, including degradation frequency, first degradation occurrence, and transition maps, enabling detailed spatio-temporal analysis of forest disturbance dynamics. By explicitly modeling seasonal variation and leveraging the complementary strengths of optical and radar observations, the proposed framework provides more robust and consistent deforestation detection than conventional single-sensor or static classification approaches. The DL-SMTD framework offers a scalable and reliable tool for tropical forest monitoring and supports REDD+ implementation, conservation planning, and near real-time forest disturbance detection in data-limited tropical regions.</p> Graphical Abstract <p></p> <p>The graphical abstract illustrates the workflow and key components of the DL-SMTD framework for accurate monitoring of tropical deforestation. On the left, multi-sensor satellite inputs are shown, including high-resolution Planet NICFI optical imagery and Sentinel-1 SAR bands (VV, VH, VV/VH), which are combined to capture complementary spectral and structural information. These inputs are organized into six bi-annual seasonal composites spanning 2020–2023, highlighting the integration of temporal dynamics to account for seasonal variation in tropical rainforests. In the center, two classification approaches are depicted: a deep learning U-Net architecture and a traditional Random Forest (RF) model, both trained on labeled image tiles of 256 × 256 pixels. The fusion of optical and radar data is represented as a converging pipeline, emphasizing how combining multiple data sources enhances feature representation for improved detection of deforestation and forest degradation. The outputs of the DL-SMTD framework are visualized, including maps of deforestation, degradation frequency, first degradation occurrence, and transitions over time, enabling detailed spatio-temporal analysis. Key performance metrics, such as overall accuracy and Intersection over Union (IoU), are highlighted to demonstrate the superior results achieved with the fused U-Net model and the further improvement provided by the full DL-SMTD framework. The graphical abstract conveys the significance of integrating multi-sensor and multi-temporal data with deep learning to capture seasonal variability, improve classification accuracy, and provide actionable insights for conservation, REDD+ implementation, and sustainable land management in tropical regions. It underscores the framework’s potential as a scalable, reliable tool for real-time monitoring of forest disturbances in data-limited environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Mapping Deforestation Dynamics Using a Multi-Sensor Spatio-Temporal Deep Learning Framework

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Isiaka Lukman Alage,
  • Ahmed Wasiu Akande,
  • Seyi Festus Olatoyinbo

摘要

Accurate monitoring of tropical deforestation is critical for climate change mitigation, biodiversity conservation, and sustainable land management. However, many remote sensing approaches overlook seasonal variability, which can significantly affect classification accuracy in humid tropical environments due to phenological changes and atmospheric conditions. To address this limitation, this study proposes the Deep Learning Spatio-Temporal Multi-Temporal Degradation (DL-SMTD) framework, which explicitly incorporates seasonal dynamics while integrating multi-sensor satellite data. The framework combines high-resolution Planet NICFI optical imagery and Sentinel-1 SAR data to improve the detection of deforestation and forest degradation. Two models were developed and compared: a deep learning U-Net architecture and a traditional Random Forest (RF) classifier. Both models were trained and evaluated using labeled image tiles (256 × 256 pixels) derived from Nigeria’s Cross River tropical rainforest, covering six bi-annual seasonal composites between 2020 and 2023. Each composite contains seven spectral features, including four optical bands from Planet NICFI (blue, green, red, and near-infrared) and three SAR features from Sentinel-1 (VV, VH, and VV/VH). Model performance was evaluated using different input configurations, including single-sensor datasets and multi-sensor data fusion. Results indicate that integrating optical and SAR data improves classification performance compared with single-sensor approaches. The U-Net model using fused Planet and SAR data achieved an overall accuracy of 0.93 and an Intersection over Union (IoU) of 0.90, while the RF model achieved 0.92 accuracy and 0.89 IoU with the same input configuration. Building on these results, the proposed DL-SMTD framework incorporates temporal consistency constraints and degradation detection rules to account for seasonal variability, further improving performance to 0.98 overall accuracy and 0.91 IoU. Beyond classification, DL-SMTD generates multi-temporal forest change products, including degradation frequency, first degradation occurrence, and transition maps, enabling detailed spatio-temporal analysis of forest disturbance dynamics. By explicitly modeling seasonal variation and leveraging the complementary strengths of optical and radar observations, the proposed framework provides more robust and consistent deforestation detection than conventional single-sensor or static classification approaches. The DL-SMTD framework offers a scalable and reliable tool for tropical forest monitoring and supports REDD+ implementation, conservation planning, and near real-time forest disturbance detection in data-limited tropical regions.

Graphical Abstract

The graphical abstract illustrates the workflow and key components of the DL-SMTD framework for accurate monitoring of tropical deforestation. On the left, multi-sensor satellite inputs are shown, including high-resolution Planet NICFI optical imagery and Sentinel-1 SAR bands (VV, VH, VV/VH), which are combined to capture complementary spectral and structural information. These inputs are organized into six bi-annual seasonal composites spanning 2020–2023, highlighting the integration of temporal dynamics to account for seasonal variation in tropical rainforests. In the center, two classification approaches are depicted: a deep learning U-Net architecture and a traditional Random Forest (RF) model, both trained on labeled image tiles of 256 × 256 pixels. The fusion of optical and radar data is represented as a converging pipeline, emphasizing how combining multiple data sources enhances feature representation for improved detection of deforestation and forest degradation. The outputs of the DL-SMTD framework are visualized, including maps of deforestation, degradation frequency, first degradation occurrence, and transitions over time, enabling detailed spatio-temporal analysis. Key performance metrics, such as overall accuracy and Intersection over Union (IoU), are highlighted to demonstrate the superior results achieved with the fused U-Net model and the further improvement provided by the full DL-SMTD framework. The graphical abstract conveys the significance of integrating multi-sensor and multi-temporal data with deep learning to capture seasonal variability, improve classification accuracy, and provide actionable insights for conservation, REDD+ implementation, and sustainable land management in tropical regions. It underscores the framework’s potential as a scalable, reliable tool for real-time monitoring of forest disturbances in data-limited environments.