Crop yield prediction in terms of fresh weight is one of the many applications of machine learning in agriculture. Accurate yield prediction models are crucial as they guide growers in making appropriate decisions on what and when to cultivate under varying circumstances influenced by climatic and crop growth parameters, and market conditions. Production of sugarcane involves activities that heavily rely on accurate and timely cropping and harvest forecasting. This study aims to introduce a machine learning model, that estimates sugarcane yield based on various agronomic and field data collected in southwest Iran. Radial basis function neural network (RBF-NN), a shallow feedforward neural network distinguished for its simple structure, universal approximation, and fast learning speed, was employed to develop a yield prediction model. By utilizing datasets containing nine types of input variables and determining optimal values for network hyperparameters, different algorithms were examined for network training. The RBF-NN trained by Levenberg–Marquardt algorithm, containing 75 neurons in the hidden layer, bandwidth value of 0.9, while using 80% of the total data for training, achieved the highest accuracy with an efficiency of 92%, and least amount of estimation error of 4.77%. Further analysis revealed that crop harvest schedule, electrical conductivity of the soil, and crop variety have the most significant impacts on the estimation accuracy of the model. In terms of performance, RBF-NN may not always demonstrate a significant advantage over similar machine learning models. Yet, faster learning and quick convergence are its remarkable points, particularly when dealing with large datasets.

Keywords