Multi-sensor Fusion for Tropical Forest Carbon Stock Prediction

Abstract

Tropical forests store approximately 25% of terrestrial carbon, making accurate quantification of Aboveground Biomass Density (AGBD) critical for global climate mitigation and national reporting under REDD+ mechanisms. However, operational biomass mapping in dense tropical environments faces significant challenges due to signal saturation in optical and C-band Synthetic Aperture Radar (SAR) sensors. While recent studies have explored neural networks for biomass estimation, many rely on incomplete sensor stacks that fail to capture the complex structure of mature forests. This study proposes a robust Multi-Sensor Data Fusion framework to predict AGBD in the Special Region of Yogyakarta, integrating Sentinel-1 (C-band SAR), Sentinel-2 (Optical), and ALOS PALSAR-2 (L-band SAR) imagery. Calibrated estimates from the Global Ecosystem Dynamics Investigation (GEDI) Level 4A product were utilized as high-fidelity reference data. To identify the optimal modeling strategy, sixteen experimental configurations were rigorously evaluated, testing four feature selection algorithms (Recursive Feature Elimination, Mutual Information, SelectKBest, and PCA) paired with four regression models (Linear Regression, Random Forest, SVR, and Multi-Layer Perceptron). The experimental results demonstrate that the proposed fusion architecture, utilizing Recursive Feature Elimination (RFE) with a Multi-Layer Perceptron (MLP), achieved the most robust performance, yielding an R2 of 0.3432, an RMSE of 74.02 Mg/ha, and a Mean Absolute Percentage Error (MAPE) of 69.98%. While the Support Vector Regression (SVR) model yielded a lower percentage error (MAPE 54.41%), it was rejected due to a systematic negative bias (-14.25 Mg/ha), which resulted in the severe underestimation of high-carbon stock areas. Conversely, univariate filter methods (Mutual Information) proved ineffective (R2 < 0.26) as they prioritized optical indices that saturate early, discarding essential L-band volume scattering information. This thesis confirms that incorporating L-band SAR is scientifically necessary to overcome the saturation limits of C-band-only approaches. The resulting model provides a scalable, scientifically justified method for high-resolution carbon stock mapping, supporting Indonesia’s commitments to the Paris Agreement.