Robust Crop Yield Prediction via Convolutional Neural Networks: Benchmarking Against Classical Machine Learning Models

Abstract

In the face of climate variability and population growth, prediction of the yield is a crucial requirement for food security and sustainable agriculture. A novel deep learning model based on one-dimensional Convolutional Neural Network (1D-CNN) architecture was proposed in this paper to predict crop yields by all agroclimatic variables including soil properties, weather, nutrient level and crop type. Because of the high sensitivity of the data used in the model, a well selected and balanced dataset consisting of 996 samples of six major crops were used in the training (Rice, Wheat, Maize, Sugarcane, Cotton and Groundnut) and validation. CNN model was also strictly benchmarked on five known classical machine learning algorithms, namely, Linear Regression, Decision Tree, Random Forest, Gradient Boosting, and XGBoost. The results obtained from experiments have been found that the proposed CNN architecture has high predictive performance as Mean Squared Error (MSE) is found to be 102,385 kg2/ha2, Mean Absolute Error (MAE) is 242 kg/ha2 and the coefficient of determination (R2) is 0.958, which shows that the CNN model has a strong predictive performance and is very good. Under the analysis of the most predictive factors of yield outcomes, the nutrient content in soil moisture, nitrogen availability, seasonal rainfall and crop type were identified. The recommended approach provides a scalable solution, based on data, for precision-agriculture applications, resource-scheduling plans and policy development programs aiming to improve the productivity and sustainability of agriculture.