PEDS-AI: A Novel Unmanned Aerial Vehicle Based Artificial Intelligence Powered Visual-Acoustic Pest Early Detection and Identification System for Field Deployment and Surveillance

Currently, there is an urgent need to develop a system for practical in-field pest early detection to combat grasshopper invasions of farmland. This research proposes the first novel visual-acoustic system created to expand the sensorial range for more field flexibility compared to the current unmanned aerial vehicle (UAV) detection. A visual-acoustic database is also designed to support this study. The proposed UAV system is designed and implemented to survey from an optimal 1m altitude and record 48k sampling rate audio to capture the unique visual-acoustic signatures of pests, allowing PEDS-AI to achieve UAV species detection, a problem difficult even for experts. PEDS-AI introduces a dual signal processing algorithm that combines deep learning source separation and spectral gating denoising to cope with the signal-to-noise ratios (SNR) caused by the UAV and quiet pests. PEDS-AI then uses segmented photos and scalograms of the cleaned recordings to perform species determination through convolutional neural networks enhanced with transfer learning. The current design can completely cover up to 12,500 $\frac{{{m^2}}}{{hr}}$ for less than 1000 dollars per unit. Improving the Yolov5 model with 46.5 million parameters with a UAV-collected dataset of 100+ infield 48MP photos, visual detection of grasshoppers achieved a precision of 0.92 and a recall of 0.84. In-field recordings from 16 species enhanced the Efficientnetv2 model with 52.9 million parameters, reaching 0.87 precision and 0.90 recall. This research has an immediate impact on in-field agricultural tracking of pest species and can also be applied to ecological management and ecosystem restoration.