Reconstruction of understory terrain based on machine learning combined with GEDI and AW3D30 data

Accurate reconstruction of understory terrain is essential for environmental monitoring and resource management. This study integrates 1:10, 000 Digital Elevation Model, Global Ecosystem Dynamics Investigation (GEDI), and AW3D30 Digital Surface Model data, combined with three machine learning algorithms—Random Forest (RF), Back Propagation Neural Network (BPNN), and Extreme Gradient Boosting (XGBoost)—to evaluate the performance of canopy height inversion and understory terrain reconstruction. The analysis emphasizes the impact of topographic and vegetation-related factors on model accuracy. Results reveal that slope is the most influential variable, contributing three to five times more to model performance than other features. In low-slope areas, understory terrain tends to be underestimated, whereas high-slope areas often result in overestimation. Moreover, the Normalized Difference Vegetation Index (NDVI) and land cover types, particularly forests and grasslands, significantly affect prediction accuracy, with model performance showing heightened sensitivity to vegetation characteristics in these regions. Among the models tested, XGBoost demonstrated superior performance, achieving a canopy height bias of –0.06 m, a root mean square error (RMSE) of 4.69 m for canopy height, and an RMSE of 9.82 m for understory terrain. Its ability to capture complex nonlinear relationships and handle high-dimensional data underlines its robustness. While the RF model exhibited strong stability and resistance to noise, its accuracy lagged slightly behind XGBoost. The BPNN model, by contrast, struggled in areas with complex terrain. This study offers valuable insights into feature selection and optimization in remote sensing applications, providing a reference framework for enhancing the accuracy and efficiency of environmental monitoring practices.