A framework for montane forest canopy height estimation via integrating deep learning and multi-source remote sensing data

Quantitative remote sensing-based forest parameter estimation is challenging in tropical mountainous conditions with complex topography and vegetation. To address this issue, we conducted a study utilizing Landsat 8, ALOS2 PALSAR, and GEDI data. We applied an effective deep learning framework-Deep Markov Regression (DMR)- along with Random Forest Regression (RF) and 3D Regression Kriging (3DRK) methods to estimate canopy height in subtropical mountain forests. Our goal was to explore effective modeling techniques for this task. Additionally, we treated "slope" as a dummy variable and incorporated factors such as slope and geographic coordinates into the model. The results showed that optical remote sensing provided the highest estimation accuracy in mountainous terrain, significantly outperforming both GEDI and SAR data. The combination of multiple remote sensing datasets further enhanced the estimation accuracy. Incorporating slope and geographic location data also improved model performance. Among all methods, the RF model was most sensitive to topographic variations, whereas the DMR model consistently delivered excellent performance across different slope conditions. The R2 of the DMR model was 0.772, the RMSE was 2.968 m, and the prediction accuracy approached 80 %.