,
Department of Computer Science and Engineering, GITAM (Deemed to be University), Hyderabad, India
Assistant Professor, Department of Computer Science and Engineering, GITAM (Deemed to be University), Hyderabad, India
It is essential in the context of precision agriculture to be able to give accurate advice about the types of crops to grow, as the suitability of a crop depends on the complicated, nonlinear interactions between soil fertility, nutrient levels, and climatic conditions. Although the existing body of literature currently evaluates each recommender system on its own, it fails to offer a single framework for comparing classical methods with more advanced learning methods, particularly when it comes to the individual components and the way in which they are combined within a consistent evaluation environment. In this study, Bagging, Random Forest, RBF-SVM, LightGBM, CatBoost, and a Keras multilayer perceptron
(MLP) are systematically compared using a standard preprocessing procedure and an 80:20 method of evaluation, with soil pH, nitrogen, phosphorus, potassium, temperature, humidity, and rainfall acting as the features. Missing values are dealt with, outliers are removed, encoding is performed, and z-score normalisation is applied in the same manner in all cases. The accuracies of Bagging, Random Forest and RBF-SVM are 47.97%, 50.41% and 60.16% respectively, while their Top-3 accuracies are 78.86%, 78.05% and 85.37% respectively. LightGBM achieves 92.61% accuracy and 97.13% Top-3 accuracy; CatBoost reaches 90.14% and 97.33%; and the Keras MLP attains 96.84% and 98.50%. Since there is a
class imbalance in the test subset, both macro and weighted metrics are included. A probabilistic fusion architecture is proposed for the three more advanced components. However, the final fusion weights and the results from the strict leave-one-component-out procedure are not given in the experimental records provided, so no performance figures are reported for a fused model that is not supported. The results indicate that boosting and neural network models are good choices for ranked crop decision support, on the condition that the full dataset provenance is available and independent geographic validation has been performed.
This is an open access article distributed under the Creative Commons Attribution Non-Commercial License (CC BY-NC) License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
0
The statements, opinions and data contained in the journal are solely those of the individual authors and contributors and not of the publisher and the editor(s). We stay neutral with regard to jurisdictional claims in published maps and institutional affiliations.