Crop-conditional semantic segmentation for efficient agricultural disease assessment

In this study, we introduced an innovative crop-conditional semantic segmentation architecture that seamlessly incorporates contextual metadata (crop information). This is achieved by merging the contextual information at a late layer stage, allowing the method to be integrated with any semantic segmentation architecture, including novel ones. To evaluate the effectiveness of this approach, we curated a challenging dataset of over 100,000 images captured in real-field conditions using mobile phones. This dataset includes various disease stages across 21 diseases and seven crops (wheat, barley, corn, rice, rape-seed, vinegrape, and cucumber), with the added complexity of multiple diseases coexisting in a single image. We demonstrate that incorporating contextual multi-crop information significantly enhances the performance of semantic segmentation models for plant disease detection. By leveraging crop-specific metadata, our approach achieves higher accuracy and better generalization across diverse crops (F1 = 0.68, r = 0.75) compared to traditional methods (F1 = 0.24, r = 0.68). Additionally, the adoption of a semi-supervised approach based on pseudo-labeling of single diseased plants, offers significant advantages for plant disease segmentation and quantification (F1 = 0.73, r = 0.95). This method enhances the model's performance by leveraging both labeled and unlabeled data, reducing the dependency on extensive manual annotations, which are often time-consuming and costly. The deployment of this algorithm holds the potential to revolutionize the digitization of crop protection product testing, ensuring heightened repeatability while minimizing human subjectivity. By addressing the challenges of semantic segmentation and disease quantification, we contribute to more effective and precise phenotyping, ultimately supporting better crop management and protection strategies.