Dynamic recognition on plant species and diseases using mean teacher network

Recent advances in deep learning have greatly improved plant disease and species recognition. However, discovering new or previously unseen diseases remains a major challenge. Most existing approaches assume that all unknown cases belong to a single “unknown” category and that models are trained once on a fixed dataset. In reality, agricultural data arrive over time, often without labels, and unknown diseases may belong to different categories that need to be distinguished. In this paper, we reformulate the task as a dynamic recognition problem: models must learn from incoming data, classify known diseases, and at the same time discover and group new ones. We propose a plant-specific framework called Plant-relevant Dynamic Recognition (PDR), which is tailored to the unique visual patterns of crops and their growing environments. To support this, we also introduce the Large-Scale Plant-relevant Dataset (LSPD), a benchmark that captures diverse crops, plant organs, and real-world farm conditions. Experiments demonstrate that our approach outperforms existing methods, achieving average accuracies of (93.44 ± 0.02)% for known diseases and (77.77 ± 0.03)% for new ones across five datasets. These results represent improvements of +(11.77 ± 0.01)% (paired t-test p = 0.002) and +(15.92 ± 0.01)% (paired t-test p = 0.005), respectively, over the state-of-the-art method. By supporting more accurate and timely disease discovery, this research can help farmers detect emerging threats earlier, minimize crop losses, and enhance agricultural productivity.