Pest management in agriculture, particularly for mitigating threats posed by pests such as aphids, presents a significant challenge due to their detrimental impact on crops, notably potatoes. Traditional pest control methods often resort to widespread pesticide application, underscoring the necessity for more precise and efficient strategies. In response, this study investigates the deployment of deep learning for the accurate detection of insect infestations in agricultural landscapes. At the core of this inquiry lies the utilization of Class Incremental Learning (CIL) frameworks, offering adaptability to detection models without necessitating extensive retraining. The Memo (Memory Expanded Model) approach, a method of class incremental learning, emphasizing memory optimization through layer expansion to preserve information on previous classes. Our investigation delves into optimizing memory efficiency by exploring smaller backbone architectures and varying the number of expanded layers on a constrained dataset.
Additionally, we evaluate the effectiveness of incorporating EfficientNet and SqueezeNet backbones, introducing the Efficient Memo and Squeeze-Memo models. Results indicate that the Efficient Memo model achieves a solid 5% increase compared to the original Memo, albeit with a slightly higher GPU memory usage. Conversely Squeeze-Memo although it is significantly smaller than the ResNet34, demonstrates a 6% decrease in performance . These models hold potential for deployment on edge devices in agricultural settings, contingent upon specific requirements and resource availability, thereby enabling real-time pest detection and management. The outcomes of this research contribute to advancing pest detection capabilities in agriculture, providing pragmatic solutions for optimizing resource utilization and enhancing crop management practices. Beyond pest detection, the demonstrated efficacy of CIL techniques suggests broader applications in domains necessitating adaptive learning systems, thereby laying the groundwork for future progress in applied class incremental learning.
