Study of The Transferability of Rfr-Based and Cnn-Based Algorithms for Canopy Height Prediction from Sentinel-2 Images

• Published in: IEEE International Geoscience and Remote Sensing Symposium proceedings pp. 3054 - 3057
• Main Authors: Liu, Xiaobo, Mishra, Rakesh, Zhang, Yun
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.07.2023
• Subjects: Canopy Height Map (CHM); Convolution Neural Network; Data models; Forestry; Laser radar; Lidar; Neural networks; Prediction algorithms; Predictive models; Random Forest Regression; Sentinel-2; Training
• ISSN: 2153-7003
• DOI: 10.1109/IGARSS52108.2023.10283068

Recently, studies have focused on integrating LiDAR data and satellite images to improve forest canopy height monitoring of large areas. Notably, algorithms based on Random Forest Regression (RFR) and Convolutional Neural Networks (CNN) have shown enhanced accuracy in predicting canopy heights. This study explores the transferability of RFR-based and CNN-based prediction algorithms using airborne LiDAR data and Sentinel-2 images from 2018 and 2021. The 2018 LiDAR and Sentinel-2 were used to train the RFR and CNN prediction algorithms. The trained RFR and CNN algorithms were then used to predict 2018 canopy height from 2018 Sentinel-2 and 2021 canopy height from 2021 Sentinal-2, respectively. Validation results reveal that the RFR-based algorithm achieved a mean absolute error (MAE) of 2.93m for 2018 canopy height and 3.35m for 2021 canopy height. The CNN-based algorithm yielded a MAE of 1.71m for 2018 and 3.78m for 2021. These findings demonstrate the feasibility of predicting forest canopy heights in the same and different years once the RFR and CNN prediction algorithms are properly trained.