Lightweight Deep Learning Models for Efficient Classification of Plant-Parasitic Nematodes on Resource-Constrained Devices

Abstract Agricultural productivity underpins global food security, yet it faces growing threats from plant diseases. Among these Plant Parasitic Nematodes (PPNs) contribute to an estimated annual crop loss exceeding \$157 billion. Existing methods for identifying PPNs are labor‑intensive and require specialized expertise, which limits their practical application in low resource settings. To address this challenge, we developed a lightweight deep learning models tailored for PPN classification using minimal computational resources. The models were trained on a curated dataset of 1,427 microscopic images representing eight widely distributed PPN genera in Ethiopia. Among the models evaluated, EfficientNetV2B0 has achieved superior performance compared with EfficientNetB0, NASNet Mobile, and MobileNetV2. Applying additional model compression techniques, including pruning and quantization, reduced the model size to 504 KB while incurring only a minimal accuracy loss of 0.63% (from 99.3% to 98.67%). The compressed model was deployed on a Raspberry Pi 3B and demonstrated inference times between 0.26 and 0.29 seconds per image. This compact and efficient model has strong potential for battery powered agricultural devices, enhancing the usability and accessibility of real time PPN classification in resource constrained environments. By enabling accurate identification of nematode threats, our research presents a promising technological solution to enhance crop management and drive gains in agricultural productivity.