Deep learning-based open set domain adaptation for plant disease recognition across multiple greenhouse scenarios

Plant diseases and pests cause significant losses to farmers and threaten food security worldwide. For sustainable agriculture, monitoring the conditions of plants and detecting plant diseases is essential. Recent advances in plant disease recognition based on deep learning techniques have demonstrated the potential to provide appropriate environmentally friendly monitoring. These systems can reliably distinguish plant anomalies by using images as the primary source of information. However, existing approaches make a closed-set assumption and perform supervised domain adaptation. This assumption is challenging in real applications where the target domain contains multiple data, but only a small part belongs to the classes of interest. In practice, much of this information is novel to the system and is often associated with one of the source categories, leading to wrong predictions during inference. Therefore, in this paper, we propose an approach based on open-set domain adaptation to the task of plant disease recognition to allow existing systems to operate in novel environments with changing conditions and variations. Our system explicitly addresses diagnostics as an open-set learning problem. It works primarily in the source and target domains, exploiting an accurate estimate of unknown data while maintaining the performance of known classes. The proposed architecture consists of two modules that perform bounding box detection and open-set domain adaptation. First, the detector is responsible for detecting regions with probable diseases through bounding boxes. Then, open-set domain adaptation allows transferring features from the source to the target domain by leading data to be classified as one of the target classes or labelled as unknown otherwise. To validate the performance of the proposed method, we use our database of images of tomato diseases and insects obtained from three-domain farms showing different infrastructure, control techniques, and lighting environments. Experimental results indicate that our method can effectively address changes in new farm environments during field testing and observe consistent gains from explicit modeling of unseen data.