Farmers often spray too early — and kill the insects that were protecting the crop.
Insect pests are one of the leading causes of rice crop damage. The early-warning method is sampling the insect population, and in practice that means sweep netting — slow, manual, and easy to skip. Pest Advisor does the sampling automatically, every day.
I trained YOLO v5, v7 and v8 on Colab. Most insects on a trap are tiny relative to a 5 MP frame, so plain YOLO missed them. SAHI slices the image into tiles, runs detection on each, and merges the results — it pushed v5’s recall from 46% to 63%, and gave v8 the best precision of the lot.
Farmers and agricultural extension workers see what landed on each trap, by species. When destructive insects spike, the app recommends insecticide — and when the trap is mostly spiders and lady beetles, it doesn’t. Devices are registered with their ngrok URL and pinned on a Google map.
I led the group, built the web app and trained the models. My teammates set up and retrieved the traps, labelled insects alongside me, and wrote the paper. It won Best in Thesis in the College of Computer Studies, LSPU – Santa Cruz, 2024.