<![CDATA[Purpose – Corn leaf disease diagnosis in resource-constrained agricultural settings requires mobile-deployable models that maintain a practical balance between classification accuracy, model size, and on-device latency. This study aims to provide empirical guidance for selecting lightweight Convolutional Neural Network (CNN) architectures by systematically analyzing the accuracy–efficiency trade-offs of five models for corn leaf disease classification. Design/methods/approach – MobileNetV2, MobileNetV3-Small, MobileNetV3-Large, EfficientNetB0, and NASNetMobile were evaluated on the PlantVillage Corn dataset comprising 4,188 images across four classes under identical experimental settings. The models were trained using a two-phase strategy and converted into standard and dynamic-range quantized TensorFlow Lite formats. Evaluation covered classification accuracy, macro F1-score, model size, Android on-device inference latency, Pareto frontier and radar analyses, and pairwise McNemar's tests with Yates continuity correction. Findings – EfficientNetB0 achieved the highest accuracy (95.25%) and macro F1-score (93.77%). MobileNetV3-Small offered the strongest efficiency under the tested Android CPU setting, reaching 94.54% accuracy with a 1.18 MB dynamic-range quantized TensorFlow Lite model and 3.89 ± 0.04 ms standard inference. The top three models were statistically comparable (p = 0.6625-1.0000). Research implications/limitations – Standard TensorFlow Lite is preferable for low-latency Android CPU deployment, whereas dynamic-range quantized TensorFlow Lite supports storage-constrained offline distribution. However, the findings are limited to the PlantVillage benchmark dataset and testing on a single mid-range Android device. Originality/value – This study integrates lightweight CNN benchmarking, TensorFlow Lite deployment, real-device Android testing, accuracy–efficiency trade-off analysis, and statistical validation to support evidence-based mobile agricultural AI model selection.]]>
