ST-CFI: Swin Transformer with convolutional feature interactions for identifying plant diseases.

Abstract

The increasing global population, coupled with the diminishing availability of arable land, has rendered the challenge of ensuring food security more pronounced. The prompt and precise identification of plant diseases is essential for reducing crop losses and improving agricultural yield. This paper introduces the Swin Transformer with Convolutional Feature Interactions (ST-CFI), a state-of-the-art deep learning framework designed for detecting plant diseases through the analysis of leaf images. The ST-CFI model effectively integrates the strengths of the Convolutional Neural Networks (CNNs) and Swin Transformers, enabling the extraction of both local and global features from plant images. This is achieved through the implementation of an inception architecture and cross-channel feature learning, which collectively enhance the information necessary for detailed feature extraction. Comprehensive experiments were conducted using five distinct datasets: PlantVillage, Plant Pathology 2021 competition dataset, PlantDoc, AI2018, and iBean. The ST-CFI model exhibited exceptional performance, achieving an accuracy of 99.96% on the PlantVillage dataset, 99.22% on iBean, 86.89% on AI2018, and 77.54% on PlantDoc. These results underscore the model's robustness and its capacity to generalize across various datasets and real-world conditions. The high accuracy and F1 scores, in conjunction with low loss values, further validate the model's efficacy in learning discriminative features. The ST-CFI model signifies a substantial advancement in the early and accurate detection of plant diseases, serving as a valuable instrument for precision agriculture. Its capacity to integrate CNNs and Transformers within a unified framework enhances the model's feature extraction capabilities, resulting in improved accuracy in the identification of plant diseases. This study concludes that the ST-CFI model effectively addresses plant disease detection challenges, with significant implications for agricultural sustainability and productivity.

Plant phenotyping relevance

葉画像から植物病害を識別する深層学習手法を提案し、複数データセットで性能評価しているため、植物状態の画像ベース表現型推定が中心である。

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。