A Multi-Feature Fusion Deep Neural Network with Attention for Multi-Task Crop Monitoring

Engineering Research Express. 2026. Article 7.

Obsie E. Y., Zhan Y., Liu Y.

Maintaining plant health is crucial for crop quality, food security, and agricultural productivity. Although deep learning methods have shown strong potential for species, disease, and severity classification, most existing approaches tackle these tasks separately, failing to exploit shared representations and interdependencies among them. Furthermore, the variability of disease symptoms, such as subtle differences in size, shape, and texture, presents additional challenges for accurate recognition and diagnosis. In this study, we present a deep multi-task learning framework, PlantSDS-Net, designed to perform species, disease, and severity classification in a single model. The proposed architecture comprises three main components: a multi-scale block, a deep feature-extraction backbone, and task-based attention methods. First, an improved Res2Net is employed to extract multi-scale features, enabling the model to recognize patterns at various spatial resolutions. These features are combined with high-level representations obtained from a pretrained MobileNetV2 model. Finally, task-based self-attention heads are used to direct the model’s attention to the most relevant features for each diagnostic task. The results on the AI ​​Challenger 2018 dataset demonstrate the remarkable performance of PlantSDS-Net. The proposed framework achieved accuracies of 99.89%, 99.58%, and 92.18% in species, disease, and severity classification, respectively.

Keywords: transfer learning, multi-task learning, self-Attention, deep learning, multi-scale, plant disease