Evaluation of EnMAP imagery for predictive modelling of soil salinity in highly saline soils

Soil salinization is a paramount issue affecting crop yields and soil productivity, specially threatening the soil in arid and semi-arid regions. In Bajo Guadalquivir (southern Spain), one of the main rice production areas of Spain, soil salinity has been reported by local stakeholders as the main ecological stressor affecting rice crops. EnMAP hyperspectral mission might rise promising opportunities to improve the monitoring of soil salinity and other environmental stressors. This mission provides continuous vis-NIR spectral data with a moderate temporal resolution. The aim of this study is to evaluate EnMAP imagery in two different predictive modelling workflows based on Random Forest and Support Vector Machines.

100 samples of electrical conductivity (EC) measures were collected in May-June 2023 in the study area. A EnMAP image was acquired over the study area on 22 March 2023. Vegetated and water surfaces were masked out, resulting in 80 samples of bare soils for the date of Enmap acquisition. Raw bands and soil salinity indices (SSI) were used as predictive features. SSI were based on an iterative procedure calculating normalized indices between all pair of bands, selecting the 1% of indices with higher correlation with EC. Two ML algorithms, Random Forest and Support Vector machine, were used together with a Sequential Feature Selection method built with each modelling algorithm.

The sampling results showed high soil salinity contents, with a median value of 10.96 dSm^-1. EnMAP image reached the higher accuracy using RF with R² = 0.14, RMSE = 3.14, RPIQ = 1.53. SVR performed worse, with a model achieving R² of -0.37, RMSE of 3.38 and RPIQ = 1.42. Both models selected the same two features, two SSI built with the 756/871 nm and the 972/1234 nm pairs. Given that the features were similar, differences might be derived from modelling algorithms. The results suggested that hyperspectral images are promising data sources, but their processing to get meaningful features is perhaps the most important task to obtain accurate soil salinity products.