A Comprehensive Review of Modified Recursive Feature Elimination and Related Feature Selection Methods in Precision Agriculture and Soil Health Assessment

Abstract

High-dimensional agricultural datasets integrate soil measurements, nutrient profiles, climatic records and remote-sensing observations. Selecting informative variables from these datasets is important for developing parsimonious predictive models and supporting interpretation. Recursive Feature Elimination (RFE) is a wrapper-based feature-selection procedure that repeatedly fits a learner, ranks predictors and removes lower-ranked variables. Modified Recursive Feature Elimination (MRFE) and related approaches adapt aspects of this procedure, although MRFE does not denote a single universally standardized algorithm. This review examines applications of direct MRFE, conventional RFE and related feature-selection methods in land-suitability classification, soil-property mapping, soil-fertility modelling, crop-yield prediction and nutrient recommendation. The reviewed studies include an MRFE–bagging application reporting approximately 95% classification accuracy, an SVM-RFE study that reduced 30 environmental variables to five for soil-pH mapping, and a modified greedy feature-selection study that retained nine predictors from an initial set of 392 covariates. Other applications include a hybrid correlation-based filter and RF-RFE approach, cross-validated RFE for maize nutrient recommendation, and comparative evaluations of feature-selection methods for soil-fertility modelling. Reported results vary across prediction targets, datasets, algorithms and validation designs and should not be interpreted as a pooled estimate or direct ranking of methods. Nitrogen, phosphorus, temperature-related variables, topographic indices, soil moisture and vegetation indices recur among the predictors discussed in the reviewed examples. The review also identifies methodological considerations concerning feature-selection leakage, subset-size determination, correlated predictors, spatial and temporal validation, interpretability and computational cost. Future research should emphasize transparent algorithm definitions, leakage-safe validation, independent multi-season testing and reproducible benchmarks.