Domain Adaptation Techniques for Plant Disease Classification: Bridging the Gap between Laboratory and Field Conditions
Year: 2026
Type: conference-paper
Abstract: When deep learning models based on carefully selected laboratory datasets are applied to real-world field images, their accuracy tends to decrease. In our case, the model, which reached 98.9% for PlantVillage, dropped to just 41.9% for field-collected PlantDoc images, a gap that severely limits any practical use of such systems in agriculture. To solve this problem, we studied and compared three strategies for adapting to the subject area: self-study with pseudo-labels, a progressive multi-stage curriculum, and a collaborative learning scheme with a balanced sample by class. Progressive learning, despite its intuitive appeal, suffered greatly from catastrophic forgetting and peaked at 66.5%. On the other hand, due to joint training, the accuracy of measurements in the field increased to 79.1%, which is 37.2 percentage points higher than with the unadapted baseline. Looking at individual diseases, we found that visually different diseases such as powdery mildew of squash (95.9%) and late blight of tomatoes (83.2%) brought the greatest benefits, while it turned out to be more difficult to adapt to more inconspicuous diseases. Adding an increase in testing time gave an additional 0.6% increase. Our findings indicate that training on all classes at once, rather than in stages, sidesteps the forgetting problem and makes field deployment of plant disease classifiers considerably more feasible.
Institution: Astana IT University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: hybrid
OpenAlex ID: W7168252403