Explainable Multimodal Deep Learning for Forest Canopy Height Estimation Using LiDAR, SAR, Optical, and Thermal Remote Sensing Data

Accurate estimation of forest canopy height (FCH) is essential for biomass assessment, carbon accounting, biodiversity conservation, and sustainable forest management. This study presents an Explainable Multimodal Deep Learning (EMDL) framework and an experimental protocol for FCH estimation across four geographically distributed forest domains representing tropical evergreen, temperate deciduous, boreal coniferous, and mixed forest ecosystems. The final empirical implementation will report the exact countries or regions, geographic coordinates, spatial coverage, GEDI footprint counts, and acquisition periods for the four study areas, as required for reproducible geographic evaluation. The framework integrates GEDI LiDAR-derived RH98 canopy-height observations with Sentinel-1 synthetic aperture radar (SAR), Sentinel-2 multispectral imagery, Landsat thermal observations, and Digital Elevation Model (DEM)-derived terrain information. The proposed architecture combines modality-specific feature extraction, convolutional learning, transformer-based multimodal fusion, reliability-aware attention, temporal aggregation, and probabilistic prediction. Explainability is incorporated through SHAP-based feature attribution and gradient-based spatial visualization to identify influential variables, sensor modalities, and spatial regions contributing to canopy-height prediction.

The experimental and reporting framework evaluates the proposed model against Random Forest, XGBoost, multilayer perceptron, convolutional neural network, ResNet, and Vision Transformer baselines using geographically independent partitions. Evaluation includes overall comparison, modality and architectural ablation, biome-wise performance, canopy-height and terrain stratification, missing-modality robustness, cross-biome generalization, uncertainty calibration, spatial consistency, explainability, and computational efficiency. In the illustrative reporting configuration comprising 12,480 test observations, the proposed EMDL framework achieves an RMSE of 4.82 m, compared with illustrative overall RMSE values of 6.84 m for Random Forest, 5.72 m for CNN, and 5.31 m for Vision Transformer. Across the four forest domains, illustrative EMDL RMSE values range from 4.31 m in temperate deciduous forests to 5.14 m in tropical evergreen forests. Controlled missing-modality analysis further indicates the intended importance of complementary Earth observation sources: removal of optical information increases illustrative RMSE from 4.82 m to 6.12 m (+26.97%), followed by removal of SAR (5.37 m; +11.41%), terrain information (5.21 m; +8.09%), and thermal information (5.05 m; +4.77%). Thus, in the proposed reporting scenario, optical information represents the largest performance contribution, followed by SAR, terrain, and thermal information. Cross-biome evaluation is designed to assess geographic transferability, with illustrative held-out-domain RMSE values ranging from 5.21 m to 6.08 m and values from 0.72 to 0.78. The uncertainty evaluation framework further assesses prediction-interval coverage and sharpness, while computational analysis jointly considers predictive performance, memory demand, training cost, and inference latency.