Detecting Crop Disease and Classification Via Convolutional Neural Network Architecture on Cassava, Plantain, and Cocoa Leaves and Pod Dataset

Abstract

Detection of crop diseases is crucial to avert the losses in productivity and in yield. It is not obvious to monitor the crop diseases manually as the act of disease detection is very critical. It needs a huge effort, along with knowledge of crop diseases and extensive processing times, as early and accurate detection is critical for ensuring food security and improving agricultural productivity. This study evaluated the performance of five CNN architecture, AlexNet, DarkNet-53, DenseNet-201, GoogLeNet, and ResNet-101 in detecting and classifying diseases in cassava, plantain, and cocoa from leaf and pod images. The models were trained and tested under two T-T split ratios (80:20 and 90:10) to investigate the impact of dataset partitioning on performance. Key evaluation metrics used to assess model performance included accuracy, precision, recall, and F1-score. Across both split ratios, DenseNet-201 consistently outperformed other models, while DarkNet-53 and AlexNet recorded the weakest results. Under the 80:20 split, the multi-crop approach achieved its best performance with DenseNet 201 (Accuracy: 89.58%, Precision: 90.00%, Recall: 89.00%, F1-Score: 89.00%) compared to DarkNet-53 with the lowest scores across all metrics (Accuracy: 73.33%, Precision: 73.00%, Recall: 72.00%, F1-Score: 72.00%). For the individual crop datasets, DenseNet-201 again yielded superior results, attaining up to 97.61% accuracy and an F1-Score of 96.50% on plantain and cocoa, whereas AlexNet on cassava performed the least (Accuracy: 93.45%, F1 Score: 92.00%). Similar trends were observed under the 90:10 split, where DenseNet-201 reached peak performance in both approaches, achieving 92.90% accuracy in the multi-crop setting and 99.00% accuracy with an F1-Score of 98.20% in the plantain dataset. DarkNet-53 recorded the lowest multi-class performance (Accuracy: 78.80%, F1-Score: 77.50%), while AlexNet and DarkNet-53 also underperformed on the cassava classes, with accuracies of 94.80% and 95.80% and F1-scores of 94.70% and 95.90%, respectively. Results showed that DenseNet-201 consistently outperformed other models across all crops and metrics, achieving its highest performance at the 90:10 split. Findings indicate that, while both splits produced strong results, the 90:10 T-T ratio yields superior results, whereas the 80:20 split offered a more balanced evaluation, reflecting both training accuracy and generalizability of the model under diverse field conditions. The study highlights limitations such as dataset diversity, image quality variability, and lack of real-time testing. These results show the potential of CNN-based models in advancing precision agriculture and hence, recommendations for future integration into mobile and IoT-based disease monitoring systems made. In conclusion, DenseNet-201 is recommended for deployment under farm conditions, given its consistently superior performance across all crops and evaluation metrics observed in this study.