Machine Learning Models for Predicting Forest Canopy Height in the “Souariz, Dar Chioukh” Djelfa Region of Algeria Using Multisensor and GEDI Data

Modelling Forest Canopy (FCH) height is an important and complex challenge in developing remote sensing and managementforests, and helps to understand ecosystem processes, assessing forest health, monitoring growth it, etc. The study focuses on developing model to predict canopy heights in the Djelfa « Souariz, Dar Chioukh » Region of Algeria. by Leveraging from Sentinel-1 (S-1), Sentinel-2 (S-2), and Global Ecosystem Dynamics Investigation (GEDI) data, topographical data like information on terrain elevation obtained from the Shuttle Radar Topography Mission (SRTM) data, slope and aspect data by following a complex process of data pre-processing, extraction of explanatory variables and canopy height modelling. the research employs advanced machine learning (ML) techniques to predict canopy height accurately. Two regression models were selected topredict canopy height: Random Forest (RF) and Gradient Tree Boosting (GTB). where model performance was assessed using measures such as Root Mean Squared Error (RMSE), coefficient of determination (R-Squared) and bias. In our research, the RF model revealed “Elevation” as the most influential variable, with $\left(\right.$ RMSE $=0.641 \mathrm{~m}, \mathbf{R}^{2}=0.899$, and a bias of 0.005 m) during training, and slightly lower performance in validation (RMSE = $0.804 \mathrm{~m}, \mathrm{R}^{2}=0.831$, and bias $=\mathbf{- 0. 0 4 3} \mathbf{~ m}$). In contrast, the GBT model highlighted “B3: green band “as the top influential variable, with (RMSE $=0.871 \mathrm{~m}, \mathrm{R}^{2}=0.869$, and a bias of $\mathbf{- 0. 0 8 0} \mathrm{m}$) during training, and ($\mathrm{RMSE}=0.966 \mathrm{~m}, \mathrm{R}^{2}=0.805$, and bias $=0.041 \mathrm{~m}$) during validation. Thus, the RF model slightly outperformed the GBT model in both training and validation phases for predicting forest canopy height (FCH), as the RF model has a slightly lower values for RMSE, bias and higher value for R-squared.