Pestoscope: AI-Based Pest Detection

Abstract

Before humans became hyper-specialized, most people primarily relied upon farming as a means of food acquisition. However, up to 40% of global crop production is lost to plant pests. With erratic weather linked to global warming, crop-destroying insects have begun thriving in these newfound conditions (Down To Earth 2021). Thus, addressing this challenge has become increasingly urgent to safeguard the livelihoods of millions of farmers and ensure food security for billions. This study aims to develop early pest detection tools that offer significant economic, environmental, and health benefits. We hypothesized that deep neural networks trained through transfer learning could be utilized to detect specific pest types, enabling timely treatment of infestations before they reach critical levels. To build the model, we collected data for six insect pests (bees, moths, slugs, snails, wasps, and weevils) from public datasets available in Kaggle (Marionette 2023) (Pestopia 2023). We applied transfer learning to three convolutional neural networks (CNNs): MobileNetV2 (Sandler et al. 2018), VGG16 (Simonyan and Zisserman 2014), and ResNet50 (He et al. 2016). During training, ResNet50 achieved the highest accuracy of 99.22%, with 20 epochs and a learning rate of 0.001. Subsequently, we added caterpillars to the dataset and retrained the model, resulting in a slightly lower accuracy of 99.00% with the same hyperparameters. The results from these experiments are promising, and with the integration of hardware, we can detect pests early and effectively. This solution offers farmers a proactive way to prevent crop destruction and enhance overall yields.