Machine Learning-Based Soil Productivity Estimation from Multitemporal Spectral Features
Main Article Content
Abstract
Reliable spatial information on soil condition is important for precision agriculture in Haryana, where intensive cultivation and spatial variation in soil fertility create a need for more detailed assessment than can be obtained from sparse point-based testing alone. This study develops a multitemporal Sentinel-2 machine-learning framework for mapping soil productivity across agricultural areas of Haryana, India. The framework integrates cloudand shadow-screened surface reflectance, bare-soil identification, temporal compositing, spectral-feature engineering, feature optimization, multi-property regression, and an agronomically interpreted Soil Productivity Index (SPI). Soil organic carbon (SOC), nitrogen (N), phosphorus (P), potassium (K), and pH are treated as complementary indicators of soil fertility and chemical suitability. Four regression approaches partial least squares regression (PLSR), support vector regression (SVR), random forest (RF), and extreme gradient boosting (XGBoost)—are compared under spatially independent validation. Feature selection is performed within the model-development process to avoid information leakage, and the final predictions are evaluated using R², RMSE, MAE, and RPD. The predicted indicators are transformed into standardized suitability scores and integrated into an SPI using agronomically justified response functions, with sensitivity analysis of alternative weighting schemes. SHAP analysis is used to interpret the contribution of image-derived predictors, while resampling-based uncertainty estimation is used to identify areas in which the productivity estimates are less reliable. The study is positioned as an integrated framework rather than as an algorithmic novelty claim: Sentinel-2, machine learning, feature selection, SHAP, and composite soil indices are established individually, whereas their coordinated use for spatially validated, multi-indicator productivity assessment in Haryana provides the principal methodological contribution. The final manuscript reports empirical performance only from the Haryana ground-reference dataset and independent spatial validation actually used in the experiment.