Examining the Research Gap in Embedded Machine Learning for Real-Time Crop Disease Detection on Low-Cost Nigerian Farming Devices

On-device machine learning inference for crop disease detection on affordable embedded hardware offers transformative potential for Nigerian smallholder farmers, but a systematic gap exists in understanding the accuracy-hardware trade-offs achievable on the commodity microcontrollers and development boards accessible within Nigerian agricultural extension budgets. This study examined this gap through systematic review and original experimental evaluation. A scoping review of 54 embedded ML crop disease detection publications from 2019 to 2024 identified that 91 percent used development boards priced above 40 USD (approximately 65,000 naira), far exceeding the 15,000 to 25,000 naira hardware budget realistic for Nigerian agricultural extension deployment. An empirical study trained and evaluated MobileNetV2, EfficientNet-B0, and SqueezeNet classifiers for cassava, maize, and tomato disease detection on an augmented dataset combining PlantVillage and an original 2,400-image field dataset collected in Benue and Oyo States. Models were quantised to INT8 using TensorFlow Lite and deployed on Raspberry Pi Zero 2W (12,500 naira), ESP32-S3 (9,800 naira), and Coral USB Accelerator-augmented Raspberry Pi Zero (28,000 naira) platforms. MobileNetV2 on the Coral-accelerated platform achieved 87.4 percent top-1 accuracy with 0.9-second inference latency. On the ESP32-S3 alone, EfficientNet-B0 INT8 achieved 79.2 percent accuracy at 3.4-second inference, representing the best accuracy-to-cost ratio within the 15,000 naira budget. The study recommends the ESP32-S3 platform for scalable deployment and outlines a data collection roadmap for North West and North East crop disease varieties underrepresented in current training datasets.

Keywords: embedded machine learning, crop disease detection, TensorFlow Lite, Nigeria agriculture, edge AI