Evaluating Airborne Hyperspectral Scanner (AHS) for the mapping of soil organic matter and clay in a Mediterranean forest ecosystem

• Soil organic matter and clay were mapped in a Mediterranean forest area. • Airborne Hyperspectral Scanner outperformed Sentinel 2. • Support Vector Regression with Feature Selection achieved the best results. • Hyperspectral narrow-band indices showed promising results for soil mapping. Soil mapping in Mediterranean ecosystems is experiencing a growing demand to face decision-making policies related to climate change mitigation. With several new and forthcoming hyperspectral missions, hyperspectral imagery and derived features could supply unprecedented data for soil mapping and monitoring, even in challenging areas with various land covers through indirect approaches. This study is aimed at evaluating airborne hyperspectral (Airborne Hyperspectral Scanner, AHS) and satellite Sentinel-2 multispectral data coupled with Machine Learning approaches for the modelling of soil organic matter (SOM) and clay content in a complex mountainous Mediterranean ecosystem dominated by forest areas, where the aboveground vegetation properties are exploited to predict underlying soil properties. One hundred field samples characterized for their SOM and clay content were taken from Sierra de las Nieves, southern Spain. Different features were derived from AHS and Sentinel-2 bands together with terrain features. Three modelling algorithms, Random Forest (RF), Support Vector Regression (SVR) and Partial Least Squares Regression were used together with four Feature Selection methods (Sequential Forward Selection, Sequential Flotant Forward Selection, Sequential Backward Selection and Sequential Flotant Backward Selection) built with RF and SVR. The resulting SOM maps with AHS data revealed spatial patterns controlled by land cover through remotely sensed features, while clay maps showed less defined patterns, with local differences. Hyperspectral narrow-band indices, visible region bands and terrain features were the most relevant for SOM and clay mapping. SVR showed the best modelling performances for SOM (RMSE 5.28 %, R 2 0.43) and clay (RMSE 10.25 %, R 2 0.33) modelling. Altogether, the results revealed an improved modelling performance when including feature selection methods and hyperspectral data as predictor features in soil mapping workflows, underpinning new possibilities to enhance soil mapping in complex canopy covered Mediterranean ecosystems.