Machine Learning-Based Forest Canopy Height Modeling with Multi-Sensor and GEDI Data

In our study, we modelled forest canopy height (FCH) using multi sensor data, including Sentinel-1, Sentinel-2, and Global Ecosystem Dynamics Investigation (GEDI - Relative Height (RH98)) Lidar data. We employed advanced machine learning (ML) techniques, specifically Random Forest (RF) and Classification and Regression Trees (CART) for their accuracy, robustness, and effectiveness in handling complex remote sensing data, on the Google Earth Engine (GEE) platform in the Djelfa “Souariz, Dar Chioukh” Reforestation Region of Algeria by following several phases, beginning with loading the data and proceeding to map the predicted forest canopy height for each model. We have used both statistical metrics, coefficient of determination (R squared), and Root Mean Square Error (RMSE) to evaluate the accuracy of prediction for each model. The RF model indicates that “Elevation” is the most significant variable, followed by (NIR, Green, and SWIR1) bands,” while the interquartile range (IQR) of the VV polarization “VV_iqr” has the least impact. It achieves high predictive accuracy with an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.899 in training and 0.831 in validation, with RMSE values of 0.6407 m and 0.8036 m, respectively. In comparison, the CART model also highlights Elevation, (SWIR1 and Green) bands, but with slightly lower performance, showing an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.865 in training and 0.732 in validation, and RMSE values of 0.7356 m and 1.0442 m, respectively. Therefore, the RF model is preferred here for its superior predictive performance.