Estimating canopy height in tropical forests: Integrating airborne LiDAR and multi-spectral optical data with machine learning

Accurate assessment and mapping of biomass in tropical forests is essential for understanding the contributions of forests to the global carbon budget and informing environmental policies. Canopy height models, an important predictor for estimating above-ground biomass, can be enhanced using fine-resolution remote sensing including airborne Light Detection and Ranging (LiDAR). However, these methods are expensive to deploy, and airborne LiDAR lacks continuous global coverage. To address this, various studies have incorporated freely available Synthetic Aperture Radar (SAR) and optical data to extrapolate these fine-grained observations. This study aims to compare the performance of three machine learning algorithms (Multiple Linear Regression (MLR), Random Forest (RF), and Convolutional Neural Networks (CNN)) when using PlanetScope and Sentinel-2 imagery to improve the accuracy of height predictions. We suggest using RF due to its ease in model building, computational time, and efficient feature selection for predicting canopy height. The S2 data at 10 m spatial resolution combined with RF were most appropriate, yielding an R2 of 0.68, RMSE of 3.52 m, and MAE of 2.63 m. While a combined dataset of PlanetScope and Sentinel-2 had a slightly higher R2 of 0.69, PlanetScope data are not always freely available in certain geographical regions and in non-commercial applications, making S2 a more readily available and consistent data source. Overall, this study reveals that the inclusion of PlanetScope and the application of more complex CNN models is not necessary for canopy height predictions in tropical forests, as open-access optical data combined with simpler tree-based models, such as RF, demonstrate comparable performance.