A UNIFIED MULTI-MODAL FRAMEWORK FOR CROP STRESS DETECTION: COMBINING TABULAR ENVIRONMENTAL DATA AND LEAF-IMAGE CLASSIFICATION

Abstract

Accurate crop stress detection is essential for precision agriculture; however, most existing approaches rely on binary labels that collapse distinct stress processes water deficit, nutrient deficiency, disease, and pest damage into a single "stressed" category.We demonstrate empirically that this binary formulation is the primary barrier to classification performance: five model architectures achieve ROC-AUC values within ±0.01 of the random baseline (0.50) on binary stress classification, regardless of feature engineering strategy. Decomposing the binary label into stress-type-specific categories enables anXGBoost classifier to achieve 91.4% accuracy and a macro-averaged F1-score of 0.93 using the same underlying features.To extend coverage to visual disease symptoms, we train a MobileNetV2-based CNN on paddy leaf images, achieving 93.7% binary accuracy (healthy vs. disease_stress) with 100% healthy recall.We combine both modalities in a fusion ensemble that merges tabular and image predictions through rule-based priority logic, achieving 94.6% accuracy on the evaluated image subset.

Plant phenotyping relevance

葉画像から健全・病害ストレス状態を推定するCNNと、画像・表形式データの融合分類法が研究の中心であり、植物状態の取得・推定手法を評価している。

Code and data availability

The paper uses third-party public Kaggle datasets (CHES, Crop Recommendation, Paddy Leaves) as inputs, but provides no authors' code, trained models, derived phenotype datasets, or any availability/deposit statements with URLs. The referenced datasets are cited prior-work resources, not paper-specific assets.

No evidence-backed public reproduction asset is currently recorded.