TrIncNet: a lightweight vision transformer network for identification of plant diseases

Abstract

In the agricultural sector, identifying plant diseases at their earliest possible stage of infestation still remains a huge challenge with respect to the maximization of crop production and farmers' income. In recent years, advanced computer vision techniques like Vision Transformers (ViTs) are being successfully applied to identify plant diseases automatically. However, the MLP module in existing ViTs is computationally expensive as well as inefficient in extracting promising features from diseased images. Therefore, this study proposes a comparatively lightweight and improved vision transformer network, also known as "TrIncNet" for plant disease identification. In the proposed network, we introduced a modified encoder architecture a.k.a. Trans-Inception block in which the MLP block of existing ViT was replaced by a custom inception block. Additionally, each Trans-Inception block is surrounded by a skip connection, making it much more resistant to the vanishing gradient problem. The applicability of the proposed network for identifying plant diseases was assessed using two plant disease image datasets viz: PlantVillage dataset and Maize disease dataset (contains in-field images of Maize diseases). The comparative performance analysis on both datasets reported that the proposed TrIncNet network outperformed the state-of-the-art CNN architectures viz: VGG-19, GoogLeNet, ResNet-50, Xception, InceptionV3, and MobileNet. Moreover, the experimental results also showed that the proposed network had achieved 5.38% and 2.87% higher testing accuracy than the existing ViT network on both datasets, respectively. Therefore, the lightweight nature and improved prediction performance make the proposed network suitable for being integrated with IoT devices to assist the stakeholders in identifying plant diseases at the field level.

Plant phenotyping relevance

植物の病徴画像から疾病状態を推定する新規画像解析モデルを開発し、複数データセットと既存モデルで性能比較しており、植物フェノタイピング手法が中心である。

Code and data availability

The paper used two plant disease image datasets. The PlantVillage dataset is openly available at the authors' stated public URL (matching an allowed URL), qualifying as a public paper-specific asset. The in-field Maize dataset is only available from the corresponding author on request, so it is listed as request_only.