Abstract:
Production of dragon fruit is of economic importance but is severely affected by fungus and insect outbreak, especially stem canker and scale insects that reduce production and quality. The conventional disease identification approaches that rely on manual observation are labour intensive, subjective and inapplicable to extensive agriculture. The paper is based on the deep learning-powered computer vision model that can simultaneously classify fruit and leaf health as good, bad, good, stem canker, and scale insect. The data set was a collection of 2,547 in-field pictures; this data was pre-processed using the technique of augmentations to enhance generalization. Four models were trained and analyzed through transfer learning and optimization with the help of InceptionV3, DenseNet201, MobileNetV2, and a custom CNN. DenseNet201 had the highest accuracy (99.63) with good precision and recall whereas MobileNetV2 had good performance (96.75) with computational efficiency that can be applied to mobile development. InceptionV3 provided high quality classification of 95.05 and custom CNN was used as a baseline of 86.14. Findings point to a case of trade-off between accuracy and deployability, showing that DenseNet201 is better suited to controlled settings and MobileNetV2 to real-time and resource-constrained settings. It is proposed that the system would decrease the use of manual inspection, better disease suppression and would offer a solution that is scalable and sustainable to increase productivity and farmer profitability in growing dragon fruit.
