Evaluation of convolutional neural network models' performance for estimating mango crop yield

Abstract

In agriculture, crop yield estimation is essential; producers, industrialists, and consumers all benefit from knowing the early yield. Mango manual counting typically involves the utilization of human labor. Experts visually examine each sample to complete the process, which is time-consuming, very difficult and has little precision. For commercial mango production to produce high-quality fruits from the orchard to the consumer, a quick, non-destructive, and accurate variety classification is required. Because of its effectiveness in computer vision, a convolutional neural network—one of the deep learning techniques—was chosen for this investigation. For yield prediction, a total of eight popular mango cultivars were utilized. A comparison with previously trained models was used to assess the suggested model.The performance of the classifiers was evaluated using evaluation metrics such as accuracy, loss, Roc-AUC score, precision, recall, F1-score, sensitivity, specificity, Positive Predictive Value (PPV), Negative Predictive Value (NPV), and Cohen-Kappa performance measures. In terms of performance evaluation criteria, it was discovered that the proposed approach performed better than the pre-trained models. It was discovered that the suggested model produced 98.85% accuracy in the test set, which had 800 images. This outcome has demonstrated the tangible applicability of the proposed methodology for mango crop estimation.