A robust machine learning-based engineering solution for crop recommendation in smart agriculture

Abstract

Selecting a crop that matches the soil and climate of a particular field is one of the most consequential decisions a farmer makes, yet it is still guided largely by convention and local experience. This paper treats crop selection as an information-engineering problem and reports an end-to-end decision-support system that couples a rigorously validated classifier to a deployed, explainable field application. Seven supervised algorithms—Gaussian naive bayes, random forest, gradient boosting, decision tree, support vector machine, logistic regression and k-nearest neighbours—were trained on 2 200 records covering 22 crop classes and seven agronomic inputs (N, P, K, temperature, humidity, pH and rainfall). All models were tuned by grid search with an inner five-fold cross-validation confined to the training partition and then compared under repeated stratified ten-fold cross-validation (three repeats, 30 folds), with 95% confidence intervals, paired t-tests with Holm correction and McNemar tests. Random forest achieved the highest mean accuracy, 99.55% (95% CI 99.42–99.67), with the lowest fold-to-fold dispersion (SD 0.34 percentage points); Gaussian naive bayes reached 99.48% (95% CI 99.34–99.63). We report explicitly that this 0.06 percentage-point gap is not statistically significant (paired t-test p = 0.423; McNemar p = 1.000) and therefore make no claim of superiority on accuracy alone; random forest is adopted because it is the most stable model across folds, makes no distributional assumption that the data violate, and supports exact tree-based attribution at deployment. Out-of-fold predictions over all 2200 samples give 99.59% accuracy with only nine errors, seven of which are rice recommended as jute—two crops that share almost identical nutrient profiles and differ mainly in rainfall. Permutation importance, SHapley Additive exPlanations (SHAP) attribution and leave-one-feature-out ablation identify rainfall and humidity as the dominant predictors, and per-recommendation SHAP explanations are surfaced in the deployed application so that a farmer can see why a crop was proposed. The model is served through a cloud-hosted Streamlit web application and a companion Flutter mobile client; the packaged model occupies 2.8 MB and returns a recommendation in about 2 ms on a single core. The contribution of the work lies not in the individual algorithms, which are standard, but in the combination of a transparent and statistically defensible evaluation protocol, model explainability and a documented deployment architecture within a single, reproducible engineering solution.