Cross-modal fusion approach with multispectral, LiDAR, and SAR data for forest canopy height mapping in mountainous region

Abstract

The study introduces a technique for integrating multispectral, LiDAR, and Synthetic Aperture Radar (SAR) data within a machine-learning (ML) framework. By leveraging ML models, including Random Forest (RF), Gaussian Process Regression (GPR), and k-Nearest Neighbors (k-NN), successfully provides a comprehensive methodology for mapping forest canopy height (CH) and analyzes seasonal changes from 2019 to 2023 in the mountainous region of Vodno Mountain, North Macedonia. The RF model achieved the highest accuracy (R2 = 0.91, RMSE = 1.2 m), outperforming the other models when trained with Aerial LiDAR data. The forest CH models were validated against field measurements, Aerial LiDAR, and Global Ecosystem Dynamics Investigation (GEDI) data, confirming the accuracy of the approach and showing solid correlations between predicted and observed CH values. This research is significant due to its innovative approach to forest CH modeling in a region with minimal prior studies. Integrating multi-source data enables more accurate and detailed CH mapping, essential for monitoring forest biomass and carbon stocks, detecting forest disturbances, and assessing future forest management activities.