Integrating multi-source remote sensing data and machine learning for predicting tree density and cover in Argania spinosa

Abstract

This examination explores the application of remote sensing technologies, including Sеntinеl-2, Mohammed VI satellite imagery and Unmanned Aerial Vehicles (UAVs), to predict the cover and density of Argane forest stands in Morocco. The primary objective was to determine the most dependable dataset for estimating these parameters by assessing the performance of various machine learning models. We integrated multiple vegetation indices and compared algorithms such as XGBoost, LightGBM, GBDT, RF and ANN. XGBoost and LightGBM outperformed the other models in estimating tree density using UAV and Mohammed VI data, with XGBoost achieving an impressive R² of 0.99 and RMSE values of 0.05 and 2.85, respectively, demonstrating strong alignment between predicted and measured parameters. Sеntinеl-2 data was particularly effective in predicting vegetation cover for both algorithms, exhibiting an impressive R² of 0.99 and RMSE of 0.34, highlighting a strong correlation. XGBoost and LightGBM consistently delivered the best results for estimating Argane stands density and cover, followed by GBDT, RF, and ANN. Correlation analysis revealed strong positive relationships between vegetation indices (NDVI and SeLI) and Argane stands density and cover across all data sources. The research revealed substantial variability in tree density and cover across different studied regions, with XGBoost model results indicating that the highest density (76.01 trees/ha) was recorded in Essaouira, while the lowest density (43.03 trees/ha) was found in Tiznit/Aït Baha. These findings underscore the importance of selecting appropriate data sources and algorithms for precise ecological assessments and provide valuable insights into the dynamics and ecological status of Argane forest stands, supporting effective forest management and conservation strategies in the context of climate change and environmental degradation.