A neural architecture search optimized lightweight attention ensemble model for nutrient deficiency and severity assessment in diverse crop leaves.

Abstract

Abstract The growth and productivity of banana crops are critically affected by micronutrient deficiencies, which are often difficult to detect at early stages. Lightweight deep learning models, optimized through neural architecture search (NAS) and attention mechanisms, are hypothesized to provide accurate and efficient classification of such deficiencies for real-time agricultural applications. In this study, multiple convolutional neural networks (CNNs) and mobile-friendly architectures, including ResNet50, VGG16, NASNetMobile, and MobileNet variants (V1, V2, V3), were evaluated using transfer learning on a curated banana leaf deficiency dataset. To improve robustness and prediction accuracy, modified classification layers and ensemble strategies–initially average ensembling and later a NAS-guided dynamic attention weighting mechanism were employed. This optimization resulted in a novel lightweight model, NASMobV2 (NASNetMobile + MobileNetV2), capable of both classifying nutrient deficiencies and assessing their severity levels. The proposed model achieved a validation accuracy of 98.57%, outperforming baseline and state-of-the-art counterparts in precision, recall, and F1 score. To improve generalization, banana crop diseases along with an additional Coffee crop dataset were included for evaluation. Finally, the practical utility of the model was demonstrated by deploying the trained system in both mobile and web applications, enabling farmers and agronomists to perform fast and accurate diagnostics directly in the field.

Plant phenotyping relevance

バナナ葉画像から栄養欠乏と重症度を推定する軽量深層学習モデルを開発・評価し、実運用アプリにも展開しており、植物状態の取得・推定手法が中心である。

Code and data availability

The paper's banana leaf nutrient-deficiency image dataset is publicly available on Mendeley Data and was directly used for the phenotyping/classification measurements. No author analysis code or trained model checkpoints are explicitly deposited; the GitHub/Streamlit links are deployment apps rather than deposited code