Estimation of forest canopy height by integrating multi‑source remote sensing images from GEDI, ZY-3 and OHS in Jiyuan City, China

Forest canopy height (FCH) is a key parameter for estimating forest carbon stocks and sequestration. The Global Ecosystem Dynamics Investigation (GEDI) satellite emits laser pulses, which obtain vertical forest structure information. To solve the difficulties in estimating FCH in complex mountainous areas and the low accuracy of regional scale estimation, the GEDI LiDAR, ZY-3 stereo imagery and OHS hyperspectral data were integrated to estimate FCH of Jiyuan City in China. The spectral, textural and topographic variables derived from OHS data were selected using the method of variance inflation factor (VIF) and the Generalized Additive Model – Least Absolute Shrinkage and Selection Operator (GAM-LASSO), which aimed to improve the estimation accuracy and precision, and the variables were used to construct three machine learning models including random forest (RF), bidirectional recurrent neural network (BRNN) and convolutional neural network (CNN) to estimate FCH. Results indicate that multi-source data (GEDI and ZY-3) with RF method has the highest estimating accuracy (R2 = 0.529, RMSE = 3.701 m), followed by BRNN (R2 = 0.508, RMSE = 3.782 m) and CNN (R2 = 0.494, RMSE = 3.836 m). These results demonstrate that the integration of LiDAR, OHS and spectral remote sensing data is an effective approach to improving the accuracy of FCH estimation in mountainous areas.