Assessing the potential of machine learning and integrated satellite data for canopy height mapping: a case study on the Iberian Peninsula

Abstract

Understanding the structure of forests is essential to gain insights into the diversity of ecosystems, which is key to determining their resilience to abiotic and biotic hazards. Forests play a critical role in the storage of biomass and are therefore affecting the global carbon cycle and future trajectories of atmospheric CO2 concentration. Earth observations and satellite remote sensing data have become valuable resources for observing forested areas on a global scale. From this viewpoint the Global Ecosystem Dynamics Investigation (GEDI) LIDAR (Light Detection and Ranging) mission, developed by NASA (National Aeronautics and Space Administration), measures an unprecedented range of parameters that describe forest structural complexity. In this paper, we investigate the ability of machine learning tools to discern variations in canopy height and evaluate their performance with a view to determine the most accurate one for the Iberian Peninsula. A novel approach is proposed for integrating Sentinel-1, Alos-Palsar, Sentinel-2 and spectral indices data to create an input dataset for machine learning models. Random Forest (RF), XGBoost. and Multi-Layer Perceptron (MLP) are trained to leverage canopy height data obtained from GEDI metrics. Our results show that the three algorithms perform similarly in terms of accuracy and errors, presenting MLP with the highest correlation with canopy height. Additionally, we evaluated the feature importance of each model, providing insights into the variables that influence the model's decision-making. Our analysis revealed that MLP gave greater importance to Sentinel-2 bands, whereas Random Forest relied more heavily on vegetation indices. In contrast, XGBoost exhibited a balanced approach and utilized information from both features evenly. It is noteworthy that the red-edge channels and their derived indices have demonstrated a significant level of importance across all models employed in this study.