STN-MobileNetV2: A Hybrid Lightweight Deep Learning Model for Maize Disease Classification

Abstract

Abstract Agriculture occupies an essential role because of the demand for food, and it is a crucial source of human income in many countries, especially in developing countries. In the ‎world, China is a large agricultural country and many people ‎rely on agricultural production for a ‎living. Maize is widely cultivated as a kind of the major dominant crop, while the diseases of ‎maize ‎not only influence the maize plantation but also the economic development. Severe maize diseases ‎may even result in no ‎harvest of grains. Thereupon, looking for an accurate, fast, automatic, and ‎low-cost approach to conduct maize disease recognition is ‎of great realistic importance. In this study, we put forward a novel image-based network architecture for maize disease identification. The proposed network integrates a Spatial Transformer Network (STN) with MobileNetV2, forming a hybrid architecture termed STN-MobileNetV2. This model leverages MobileNetV2’s pre-trained efficiency while enhancing spatial invariance through STN, enabling robust recognition of maize disease types.We also improved the Focal-Loss (FL) function to enable it to handle multi-class problems and keep more attention on minor lesion characteristics. When benchmarked against other state-of-the-art (SOTA) techniques, the presented approach exhibits superior efficacy. Specifically, it attains an average recognition accuracy of 88.00% and a specificity of 92.00% using the publicly available dataset.Even when eight disease types are considered, the proposed approach realizes a mean accuracy and specificity of 92.84% and95.85% on the ‎locally captured maize disease images. Results derived from the experiments demonstrate that the proposed model is highly effective in the identification of maize-related diseases.

Plant phenotyping relevance

トウモロコシ病害の病斑画像から植物の病害状態を推定する画像ベース手法を提案・ベンチマークしており、分類モデルの開発が研究の中心である。

Code and data availability

The paper's locally captured maize disease image dataset (~466 expert-labeled field images) is only available 'upon reasonable request' from the corresponding author, so it is not a public asset. The PlantVillage maize subset used is a public dataset but is cited prior work, not a paper-specific deposit. No author code

No evidence-backed public reproduction asset is currently recorded.