Predictive Modeling of soil salinity integrating remote sensing and soil variables: An ensembled deep learning approach

Accurate predictions of soil salinity can significantly contribute to achieving the UN- Sustainable Development Goal (SDG-2) of ensuring 'zero hunger.' From this perspective, the current research aimed to predict soil electrical conductivity (EC) from remote sensing and soil data using advanced deep learning (DL) architectures. A total of 109 soil samples were analyzed for agricultural land use in the Middle Indus Basin of Pakistan. Seven salinity indices (SI-1 to SI-7) were derived from the 10m to 20m wavelength bands of Sentinel-2, along with vegetation and topographic covariates. Initially, Recursive Feature Elimination was implemented as a featureselection method to select the most effective predictors. Subsequently, deep learning architectures, including a Feedforward Neural Network (FFNN), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM), were employed to predict soil salinity. Research findings showed that EC ranged between 0.57dS/m to 11.5 dS/m in the study area. The evaluation metrics of the DL models revealed that a simple FFNN with three fully connected dense layers achieved the highest R2 = 0.88 for model training. However, the ensemble of improved FFNN and LSTM outperformed with the highest R2 and NSE = 0.84, and the lowest RMSE and MAE = 1.38 and 1.01, respectively, on the testing dataset. Optimized deep learning architectures with adjustments to the learning rate, dropout rate, and activation functions achieved the highest prediction accuracy with the lowest validation loss. Finally, SHapely Additive exPlanations (SHAP) revealed that elevation, pH, NDVI, SI-1, and SI-7 had highly significant impacts on EC predictions. This research provides insight into implementing advanced and interpretable DL architectures, supporting informed decision-making by agricultural stakeholders.