Abstract : Accurate estimation of above-ground biomass (AGB) in mountainous pastures is important for monitoring of grazing livestock management. From previous studies it can be observed that integration of field data with the satellite images (sentinel-1 and/or sentinel-2) is limited. Hence, in this study our goal is to fuse both optic and SAR satellite images at the feature level as inputs in order to evaluate machine learning models to predict biomass in the mountainous of Azerbaijan. Results will be compared with the field-destructive biomass measurements collected in September 2020. Remotely sensed features were extracted from temporally aligned Sentinel-1 SAR (γ⁰_VV_dB, γ⁰_VH_dB, VH– VV_dB, RVI) and Sentinel-2 optical imagery (NDVI, NDMI, MSAVI), alongside topographic covariates (altitude, slope). Considering number of features used as an input for the ordinary least squares (OLS) regression and Random Forest (RF) models, we run the models with original number of features as well as dimensionality reduction analysis. After the initial analysis, we observed that, vegetation indices, SAR-derived features and grass, legume, forb as separate features are highly correlated with each other, therefore applying principal component analysis (PCA) became imminent. OLS and RF models achieved R² values up to 0.63 and 0.30 (cross-validated), respectively, with RMSEs corresponding to approximately 271 - 400 kg/ha. Based on feature importance analysis optical vegetation indices and altitude turns out to be most important predictors, while SAR-derived features add stability to the model rather than being significant predictor. Our results demonstrated the integrating multi sourced satellite data with on field dataset for stable summer pasture biomass estimation.

Keyword : Biomass estimation, Data Fusion, Machine Learning, NDVI, Sentinel-2, Sentinel-1, PCA.