Parameter-Efficient Fine-Tuning with Adapters for Sorghum Disease Classification
Year: 2026
Type: conference-paper
Abstract: Sorghum is a globally important cereal crop whose productivity is severely affected by plant diseases, necessitating accurate and timely disease identification. Although deep learning has shown strong performance in automated sorghum disease analysis, most existing approaches rely on full fine-tuning of large pretrained models, resulting in high computational cost, large data requirements, and limited suitability for resource-constrained agricultural environments. This paper investigates adapter-based parameter-efficient fine-tuning for multi-class sorghum disease classification using a pretrained EfficientNet-B3 backbone. Lightweight adapter modules and a classification head are trained while keeping the backbone frozen, reducing the number of trainable parameters to 15.58% of the full model. Experiments on a real-world, class-imbalanced sorghum disease dataset, handled using appropriate imbalance-aware training strategies, demonstrate that PEFT-B3 achieves up to 99.58% accuracy, matching full fine-tuning performance while significantly outperforming frozen-backbone baselines. Despite the high accuracy achieved, the dataset contains visually distinguishable disease symptoms, which contribute to the strong performance observed. The results highlight the effectiveness of parameter-efficient fine-tuning for accurate, efficient, and deployable sorghum disease classification in realistic agricultural settings.
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
Open Access status: hybrid
OpenAlex ID: W7204436997