SCARF: A new algorithm for continuous prediction of biomass dynamics using machine learning and Landsat time series

Abstract

We developed the SCARF (Spatial Mismatch and Systematic Prediction Error Corrected cAscade Random Forests) algorithm for continuous prediction of biomass dynamics using machine learning and Landsat Time Series (LTS). Our approach addresses the challenges posed by the cloudy subtropical forests in southern China, where monitoring biomass dynamics is notoriously difficult. To derive spectral-temporal features from the LTS, we applied the Continuous Change Detection and Classification (CCDC) algorithm (Zhu and Woodcock, 2014). Subsequently, we employed the cascade random forests machine learning algorithm for biomass prediction. This new approach corrects the spatial mismatch effects between plots and Landsat pixels as well as the systematic prediction errors in the machine learning model. As a result, it substantially enhances biomass prediction accuracy, with a coefficient of determination (R²) of 0.83 and a root mean square error (RMSE) of 6.27 Mg ha^-1. In comparison, the commonly used random forests approach yields an R² of 0.47 and RMSE of 8.52 Mg ha^-1. Additionally, it provides reliable spatial prediction beyond the model-training area, achieving an R² of 0.79 and an RMSE of 6.62 Mg ha^-1. Furthermore, we demonstrate that modeling five different forest age groups separately further improves prediction accuracies, resulting in an increased R² of 0.87 and a reduced RMSE of 3.65 Mg ha^-1. A comparison of the allometric model prediction from the field plots and those from the SCARF model revealed a strong agreement, indicating that this approach can provide a temporally continuous prediction of biomass dynamics. Our study presents a robust method for continuous, reliable, and explicit spatiotemporal prediction of biomass dynamics in cloudy subtropical forests using LTS.