Soil Salinity Estimation from Sentinel‐2 Satellite Data and Machine Learning Algorithms

Soil salinization is a critical environmental problem in arid and semiarid regions since it seriously affects the ecological sustainability of limited land resources. As a form of land degradation, soil salinization greatly impacts on ecosystem services [1]. Accurate monitoring of soil salinization thus plays a key role in the ecological security and sustainable agricultural development of semiarid regions. The objective of this study was to estimate soil salinity from salt-affected soils in a south Mediterranean region using Sentinel-2 multispectral imagery and machine learning algorithms. Regression models between the salinity indices and ECSoil were all significant (p≤0.001) and showed stronger relationships compared to vegetation and chlorophyll spectral indices; the highest coefficient of determination was observed for the Salinity Index 1 [2]. The use of the spectral indices is however restricted to formulas that use only a few bands, with at most three or four bands, which implies a decreasing of the complete-spectrum dataset information. Among machine learning algorithms, the Gaussian Process Regression (GPR) models presented the best results and attained a higher performance and accuracy while estimating the soil EC. Indeed, GPR models are efficient machine learning algorithms for retrieving bio-physical parameters and at the same time provide a ranking of relevant bands (features) from input spectral data. Random forest showed also good accuracy. Generally, decision tree-based algorithms require no assumptions about the data distribution, adapt to outliers by isolating them in small regions of the feature space, and have no hidden layers in their structure. The Gradient Boosting presented good statistic results when estimating ECSoil as it reduces model complexity, avoids over-fitting and increases the model’s generalization ability, which improves the efficiency of model construction and ensures prediction accuracy. This research provides a further contribution to the ability of Sentinel-2 spectral data to predict soil electrical conductivity in an agricultural area in central Tunisia. The performance of twenty-five regression algorithms was compared and quantitively assessed. Our results indicate that the combined use of machine learning algorithms and multi-spectral Sentinel-2 bandset data can represent a suitable tool for soil EC estimation. From future perspectives, the higher spectral resolution of the newly launched hyperspectral satellites (e.g., PRISMA and EnMAP) would allow the exploitation of narrow absorption spectral features in the Vis‐NIR and SWIR spectral ranges associated with salinity estimation.