A Localization-Aware Heterogeneous CNN Ensemble with Neural Meta- Fusion for Plant Disease Classification, with a Component Analysis on Laboratory and Field Images

Abstract

Abstract Deep convolutional networks now classify leaf images on curated benchmarks such as PlantVillage with accuracies close to the measurement ceiling of those datasets, which has shifted the open questions away from raw accuracy toward two under-reported issues: which components of a composite pipeline actually cause the result, and whether the components that matter on laboratory images are the same ones that matter on field photographs. We address both with a classification framework evaluated under a protocol that fixes every development decision before the independent test data are read. A Mask R-CNN stage localizes the dominant leaf, the accepted box is expanded by a validation-selected 5% margin and resized to a shared 224 by 224 input, and the same localized image is passed to ResNet50, InceptionV3, and MobileNetV2 together with an extractor that produces eighteen colour and shape descriptors. The forty-five class probabilities and eighteen descriptors form a sixty-three-dimensional input to a neural meta-classifier developed by three repetitions of stratified five-fold cross-validation. On a locked 3,101-image PlantVillage test partition of fifteen classes the framework reached 99.77% accuracy and 99.75% macro F1 with seven misclassifications, and a one-component-at-a-time ablation confirmed that every stage contributed. The same design was then trained and evaluated entirely within a separate thirteen-class PlantDoc field-image dataset, where it reached 90.03% accuracy and 89.40% macro F1. This is a within-PlantDoc experiment and not a controlled-to-field transfer test, so the figure measures how the pipeline behaves on field imagery rather than how a PlantVillage-trained model survives a domain shift. The central finding comes from running the identical ablation on both datasets: on clean images the ensemble breadth and descriptors provide the incremental gains, but under field conditions the ordering changes, and leaf localization and learned fusion become the decisive components. Removing localization cost 2.83 accuracy points and replacing the neural fusion with soft voting cost a further 2.08 points, the two largest effects on PlantDoc. The contribution is a controlled and transparent account of where each component of a localization-aware plant disease classifier earns its place, and of how that ordering shifts between laboratory and field acquisition.

Plant phenotyping relevance

葉画像から植物病害状態を推定する分類パイプラインの開発・アブレーション検証が中心であり、単なる病害実験や routine measurement ではない。

Code and data availability

保存済みの本文根拠を更新済みルールで再検証し、公開資産1件を確認しました。