Self-Supervised Learning for Agricultural Image Classification: Detecting Plant Diseases and Pests in Cotton

Abstract

The early detection of plant diseases and pest infestations is essential for improving agricultural productivity and ensuring crop health. Cotton, as a major commercial crop, is highly susceptible to a wide range of diseases and pests that can significantly affect yield and quality if not identified at an early stage. Conventional deep learning approaches for plant disease classification rely heavily on large labeled datasets, which are often difficult and costly to obtain in real-world agricultural environments. In this paper, a self-supervised learning-based framework is proposed for agricultural image classification, focusing on the detection of plant diseases and pests in cotton. The proposed approach begins with dataset curation and cleaning using an automated image inspection technique to remove noisy, blurry, and duplicate samples, thereby improving data quality. Data augmentation techniques are then applied to enhance model generalization. A contrastive learning strategy inspired by SimCLR is employed to learn robust feature representations from unlabeled images using a ResNet18 encoder. The extracted features are further refined using Principal Component Analysis (PCA) for dimensionality reduction and K-Means clustering to improve feature separability. Finally, a pretrained VGG16 model is fine-tuned using labeled data for classification. The integration of self-supervised learning with feature refinement and clustering enhances representation quality while reducing dependency on labeled datasets. The proposed framework is scalable, efficient, and suitable for real-world agricultural applications, including automated crop monitoring and intelligent farming systems.

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.