Abstract

The analysis of soils and proper prediction of crops are very important to help grow more productive agriculture and sustainable food production. A Hybrid CNN–Machine Learning (CNN-ML) Framework for Soil Classification and Crop Prediction is proposed in this paper to combine deep learning and machine learning approaches for intelligent farming decision-making. The proposed framework uses Convolutional Neural Networks (CNNs) for automatic soil classification and machine learning algorithms for crop recommendation. There are five types of soil which are described by their image, such as: Black Soil, Cinder Soil, Laterite Soil, Peat Soil and Yellow Soil. CNN, MobileNetV2 and ResNet50 were implemented and compared to assess the effectiveness of different deep learning architectures. The CNN model was found to be the most effective soil classification model with 98.71% accuracy, which is better than MobileNetV2 and ResNet50, as it can automatically extract discriminative texture, color features from soil images for classification. After soil classification, soil-specific nutrient characteristics (Nitrogen (N), Phosphorus (P), Potassium (K), pH, temperature, humidity and ranges of rainfall) were gathered from literature. A dataset of crops was then downloaded from Kaggle and categorised into five soil types according to these nutrient profiles found in literature. For soil-specific crop prediction, several machine learning algorithms such as Logistic Regression, K-Nearest Neighbors (KNN), Random Forest and XGBoost were evaluated. Each soil category was analyzed independently with Logistic Regression, K-Nearest Neighbors (KNN), Random Forest and XGBoost algorithms for crop prediction. The final performance was calculated by taking mean of the results across the five soil datasets. Experimental results demonstrated that among the different machine learning models, XGBoost model outperforms others with the highest average accuracy of 88.0% in crop prediction. Experimental results showed that XGBoost model has the highest accuracy of 88.0% in crop prediction with the best predictive capability and generalization performance among the other machine learning models.