Spatial Prediction of Soil Attributes from PRISMA Hyperspectral Imagery Using Wrapper Feature Selection and Ensemble Modeling

Abstract

Soil organic matter (SOM), available phosphorus (P2O5), and exchangeable potassium (K2O) are vital indicators of soil fertility. Standard approaches for determining SOM, P2O5, and K2O in soil levels are usually through solution extraction of the nutrient followed by chemical analysis, which are costly and time consuming, particularly at large scale. Hyperspectral imagery, exemplified by PRISMA with its rich spectral data (231 bands), offers a promising approach for mapping soil properties like SOM, P2O5, and K2O. However, the potentially redundant or correlated bands, can hinder the accuracy and interpretability of predictive models. To address this, the present study investigates the potential of ensemble learning models based on Recursive Feature Elimination (RFE) for rapid and accurate quantification of SOM, P2O5, and K2O using hyperspectral imagery (PRISMA). We employ an ensemble of linear and non-linear learners (Partial Least Squares Regression, Support Vector Machine Regression, and Gaussian Processes Regression) within the RFE framework to identify the most informative spectral bands for predicting soil properties from a dataset of 1217 Moroccan agricultural soil samples splitted over Haouz and Settat, two disctinct agricultural regions in Morocco, where Haouz served for the training of model, and Settat for testing. Our results demonstrate significant improvements in prediction accuracy using RFE, particularly for Support Vector Machine Regression (SVR) and Gaussian Processes Regression (GPR) models. Comparing the ensemble learning models with and without RFE demonstrates a clear enhancement in performance, as the coefficient of determination (R2) for SOM improved significantly from 0.27 to 0.65, while the Root Mean Squared Error (RMSE) decreased from 0.93 to 0.66%, highlight-ing the positive impact of RFE on model accuracy. Similar improvements are observed for P2O5, and K2O. The Ratio of Performance to Inter-quartile range (RPIQ) demonstrates substantial improvement for all nutrients, highlighting the efficiency of RFE in boosting prediction accuracy. The model’s has been tested on a different geographic area and was able to perform with similar accuracy. These findings underscore the importance of feature selection for developing robust models for soil property estimation using hyperspectral data. This approach has the potential to enhance precision agriculture practices by advancing the development of remote sensing tools that support variable-rate fertilizer application.