Satellite-based estimation of soil organic carbon in Portuguese grasslands

• Autor(a) principal: Morais, Tiago G.
• Data de Publicação: 2023
• Outros Autores: Jongen, Marjan, Tufik, Camila, Rodrigues, Nuno R., Gama, Ivo, Serrano, João, Gonçalves, Maria C., Mano, Raquel, Domingos, Tiago, Teixeira, Ricardo F. M.
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.5/29346
• topic: remote sensing; satellite; cross-validation; features selection; sown biodiverse pasture

Soil organic carbon (SOC) sequestration is one of the main ecosystem services provided by well-managed grasslands. In the Mediterranean region, sown biodiverse pastures (SBP) rich in legumes are a nature-based, innovative, and economically competitive livestock production system. As a co-benefit of increased yield, they also contribute to carbon sequestration through SOC accumulation. However, SOC monitoring in SBP require time-consuming and costly field work. Methods: In this study, we propose an expedited and cost-effective indirect method to estimate SOC content. In this study, we developed models for estimating SOC concentration by combining remote sensing (RS) and machine learning (ML) approaches. We used field-measured data collected from nine different farms during four production years (between 2017 and 2021). We utilized RS data from both Sentinel-1 and Sentinel-2, including reflectance bands and vegetation indices. We also used other covariates such as climatic, soil, and terrain variables, for a total of 49 inputs. To reduce multicollinearity problems between the different variables, we performed feature selection using the sequential feature selection approach. We then estimated SOC content using both the complete dataset and the selected features. Multiple ML methods were tested and compared, including multiple linear regression (MLR), random forests (RF), extreme gradient boosting (XGB), and artificial neural networks (ANN). We used a random cross-validation approach (with 10 folds). To find the hyperparameters that led to the best performance, we used a Bayesian optimization approach. Results: Results showed that the XGB method led to higher estimation accuracy than the other methods, and the estimation performance was not significantly influenced by the feature selection approach. For XGB, the average root mean square error (RMSE), measured on the test set among all folds, was 2.78 g kg−1 (r2 equal to 0.68) without feature selection, and 2.77 g kg−1 (r2 equal to 0.68) with feature selection (average SOC content is 13 g kg−1). The models were applied to obtain SOC content maps for all farms. Discussion: This work demonstrated that combining RS and ML can help obtain quick estimations of SOC content to assist with SBP management