Soil suitability classification for crop selection in precision agriculture using GBRT-based hybrid DNN surrogate models

Abstract

The main reason for agricultural productivity decline is farmers' failure to choose the appropriate crop for their soil. It is important for farmers to understand which crops are suitable for different soil types based on their characteristics. Due to the vast variety of soil types worldwide, farmers often struggle to choose the most profitable crop for their land. To improve crop yields, a crop selection system has been developed using GBRT-based deep learning surrogate models. Gradient Boosted Regression Tree (GBRT) has been combined with a Bayesian optimization (BO) algorithm to determine the most optimal hyperparameters for the deep neural network. The optimized hyperparameters are then applied during the testing phase. Further, the impact of each input parameter on the individual outputs is evaluated using explainable artificial intelligence (XAI). The crop recommendation system comprises data preparation, classification, and performance evaluation modules. A classification method based on confusion matrices and performance matrices, as well as feature analysis using density plots and correlation plots, follows. The crop selection system categorizes the experimental dataset into 12 classes, with three for each of the four crops. The dataset includes soil-specific physical and chemical features such as sand, silt, clay, pH, electric conductivity (EC), soil organic carbon (SOC), nitrogen (N), phosphorus (P), and potassium (K). The developed surrogate model is highly accurate, precise, and reliable, with an F1-Score of 1.0 for all classes in the dataset, indicating exact accuracy and recall. The DNN-based classification model achieves an average classification accuracy of 1.00.