Answers are grounded in this paper and cite their sources.
Municipal mosquito control programs depend on timely counts and species identification, yet these tasks take staff time, require scarce expertise, and are often handled in separate tools. We developed MOSAIC (Mosquito Observations and Surveillance using AI for Communities), a web platform bringing trap collections, photographs, counts, laboratory results, and expert review into a single record and turning them into weekly trap priorities. We evaluated it with the vector control program of Corpus Christi, Texas, USA. Counting uses classical computer vision (OpenCV) without a trained model; for identification, we compared multimodal large language models (LLMs) with a laboratory-trained convolutional neural network (CNN) and tested LLMs given a local species list, taxonomic knowledge, and labeled examples. OpenCV counts were within two mosquitoes of the operators’ blind counts for 69% of 175 collections. On citizen-science photographs, the zero-shot LLM was correct on 89% of the genera it named but named a genus for 83% (overall accuracy 74%, versus 77% for the CNN). Across three runs, added knowledge and examples raised mean forced-choice species accuracy from 17 to 20% to 63–70%, but when abstention was allowed, the models named a species for only 23–49% of specimens. Automated tools can assist surveillance while experts confirm the outputs.
No takes yet. Share an insight, caveat, or question.
Kuchavaram et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: