Deep Learning Techniques for Insect Pest Life-stage Classification and Detection from Images

Fatin Faiaz Ahsan

doi:10.60867/00000125

Insect pests, whether as adults, larvae, or eggs, pose a significant risk to agriculture by reducing crop yields. Advanced deep learning (DL) techniques have the potential to automate various tasks involved in insect pest management. However, detecting pests in the real world from digital imagery is challenging because pests often have small, camouflaged appearances that change across different life stages (e.g., adult, larva, pupa, egg). Recognising insect species and their life stages is important for implementing targeted control measures, as these can vary depending on the prevailing life stage. This thesis primarily focuses on enhancing the performance of deep learning models for recognising insect pest life stages from digital images. This thesis has the following contributions: First, a literature review has been presented, which sheds light on the existing challenges in developing automatic methods using DL based techniques for classifying insect pest species and their life stages. Second, one of the most extensive publicly available datasets of insect pests, IP102, was repurposed by assigning appropriate life-stage labels to the insect instances. Several neural network models were evaluated for classifying insect pest life stages and establishing a benchmark for this task. The dataset was assessed using multiple models, and both insect species and life-stage classifications were investigated, achieving high-throughput identification. Third, we proposed a semi-supervised learning pipeline for identifying insect species and life stages, with a unified class weighting method that prioritises minority classes, as class imbalance highly impacts the classification performance. We also observed that Focal Loss, combined with a simple class-weighting strategy, improves training performance in a class-imbalance scenario. Fourth, a weakly supervised detection framework has been proposed that leverages pseudo-labels generated by a pre-trained segmentation model. These labels are used to train an object detector on the IP102 dataset, which was repurposed in Chapter 3, to classify insect life stages (Adult, Larva, Pupa, and Egg). By converting segmentation masks into bounding boxes and refining class labels iteratively, the weakly supervised method eliminates the need for manual bounding-box annotation during training. Evaluation on both pseudo-labelled and manually annotated test sets demonstrates strong performance for visually distinctive stages, such as Adult and Pupa, while highlighting reduced robustness for Larva and, particularly, Egg due to small-object sensitivity and class imbalance. These findings confirm that segmentation-driven pseudo-labelling can serve as an effective and scalable alternative to fully supervised detection, while highlighting the need for further improvements to enhance reliability across all life stages. Finally, 10 widely used activation functions (AF) in the context of insect pest detection are investigated. The Borda count voting method is employed to rank AFs across multiple evaluation metrics, providing a fair comparison. Results show that SiLU almost consistently outperforms other AFs. A set of recommendations based on this analysis is provided, highlighting the importance of AF selection and offering practical suggestions for enhancing insect pest detection models in future research. Together, this work demonstrates the potential of DL-based techniques to accurately detect and classify insect species and their life stages, providing a tool to support more effective monitoring and management of agricultural pests.

Deep Learning Techniques for Insect Pest Life-stage Classification and Detection from Images

Files and links (1)

Abstract

Details

Metrics