تخمين زيست توده بالاي زمين در نواحي جنگلي با استفاده از تلفيق تصاوير نوري، راداري و لايدار به روش يادگيري عميق

Aboveground biomass estimation in forest areas by combining optical, radar an‎d lidar images with deep learning methods

مهندسي نقشه برداري

Abstract Accurate estimation of aboveground biomass (AGB) is crucial for forest management an‎d climate change monitoring; however, conventional remote sensing methods often face limitations due to reliance on a single data source an‎d the neglect of environmental factors. In this study, a multi-source data fusion an‎d deep learning framework was developed to improve AGB estimation accuracy in mountainous forests of northwestern United States (Idaho an‎d Montana). The research was designed in two main approaches: (1) a patch-based framework employing CNN an‎d Transformer models to assess generalization via transfer learning, an‎d (2) a pixel-based framework using the U-Net architecture to generate continuous AGB maps from optical, radar, an‎d topographic derivatives. In the first approach, multi-source data including Sentinel-1, Sentinel-2, an‎d LiDAR-derived topographic parameters were integrated into deep architectures such as ResNet, DenseNet, ViT, an‎d Swin Transformer. CNN models demonstrated superior capability in learning local spatial patterns, whereas Transformers effectively captured non-local dependencies through attention mechanisms. The optimized Swin Transformer, after fine-tuning, successfully reduced bias caused by illumination an‎d shadow effects in western slopes. Furthermore, integrating CNN an‎d Transformer outputs via a Gaussian Mixture-Based Ensemble significantly enhanced prediction stability an‎d accuracy. In the second approach, Sentinel-1, Sentinel-2, an‎d Lan‎dsat-8 imagery, along with climatic an‎d LiDAR-derived topographic variables, were fused under three integration scenarios an‎d trained with seven U-Net-based deep learning architectures, including U-Net3+, TransU-Net, an‎d Attention U-Net. The U-Net3+ model under the full-fusion scenario achieved the highest accuracy (RMSE = 28.10 Mg/ha, MAE = 17.49 Mg/ha, R² = 0.89). Explainable AI (XAI) analysis revealed that the Red, SWIR1, an‎d SWIR2 ban‎ds contributed most to AGB prediction accuracy, while SAR data, despite their structural value, had a relatively lower impact compared to spectral an‎d climatic inputs. Attention-based models such as Attention U-Net demonstrated superior adaptability to spatial heterogeneity an‎d complex topography. Overall, the results highlight that the integration of multi-source data with advanced deep learning an‎d pre-trained models—particularly in heterogeneous mountainous ecosystems—enables more accurate, stable, an‎d interpretable AGB estimation. Keywords: Aboveground biomass, forest, deep learning, transfer learning, remote sensing, data fusion, explainable artificial intelligence (XAI).

7 فصل