AN ARTIFICIAL INTELLIGENCE FRAMEWORK FOR REAL-TIME DETECTION OF CASSAVA MOSAIC DISEASE USING LOW-COST ANDROID SMARTPHONES

In Katsina State, Nigeria, this study created and assessed CassavaGuard AI, an offline Android software for image-based cassava mosaic disease (CMD) screening. The program makes use of a pretrained MobileNetV3-Small network that has been compressed using static INT8 post-training quantization and adjusted for four operational leaf-image classes. Training (5,915), validation (1,268), and holdout test (1,267) sets were created from an 8,450 image dataset. The model obtained a macro F1-score of 92.45% and an accuracy of 92.58% on the holdout test set. The model file size was decreased from 21.4 MB to 5.9 MB using quantization. Total image-processing time was 42.6 ms on the Tecno Spark 8P and 58.4 ms on the 2 GB Infinix Hot 10i; peak application memory use was 145 MB. The usability assessment involved 45 farmers and 15 agricultural extension officers. The group mean System Usability Scale (SUS) scores were 87.75 and 94.00, respectively, giving a weighted overall score of 89.31/100; this calculation should be checked against the original questionnaire records. The study integrates on-device classification, local record storage, optional synchronization, and Hausa–English user support. The findings indicate the potential of lightweight deep learning for offline cassava disease screening on low-cost smartphones, although independent validation is needed before wider deployment