Machine Learning-Based Crop Suitability Prediction: An Emerging Technique for Sustainable Agricultural Production in the Desert Region of India

Machine learning (ML) algorithms can be applied to predict the suitability of soil for crop cultivation based on digital soil mapping. We used three distinct models viz. Multinomial Logistic Regression (MnLR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) to predict the suitability of wheat and pearl millet in the Barmer district of the Thar Desert. After the computation of crop suitability classes, ML techniques were used to develop suitability maps of wheat and pearl millet in the study area. The study found that the RF and XGBoost models worked well to classify crop suitability. The RF model showed that 11.9% of the total area was highly suitable, 1.6% was moderately suitable, 14.9% was marginally suitable, and 71.6% was not suitable for wheat crop. RF model for pearl millet showed that 15.5% of the area is highly suitable. Soil suitability mapping for wheat showed a Kappa index ranging from 0.23 to 0.57 and an overall accuracy ranging from 0.79 to 0.86, whereas the prediction of suitability for pearl millet showed a moderate range of Kappa index from 0.31 to 0.58 and accuracy from 0.63 to 0.77. The area under curve (AUC) for wheat crop was 0.72, 0.88, and 0.91 for MnLR, RF, and XGBoost models, respectively. Overall, the RF model performed better than the MnLR model, showing a 16% increase in accuracy. Therefore, the developed suitability maps using ML provide valuable details on agricultural potential in the Indian desert region while harmonizing its impact on the environment and the economy.