The method of the aerial census is widely adopted and the standard in conservation track. This practice raises concerns, however, due to the difficulty of finding well-trained RSO’s and their ability to get fatigued on long missions. RSO’s performance is also less precise and reliable compared to human annotators who are trained to track and count wildlife from taken images.

Using new technology like cheap digital imaging, photographic aerial survey (PAS), for data capture and unmanned aerial vehicles (UAVs) are promising solutions for reducing the financial and logistical cost of flights. However, both PAS and UAV surveys are designed to take continuous photographs along flight lines, which results in tens of thousands of images and requires intensive labor efforts to sort through images, driving up the cost. Results suggest that manual photo counts after the UAV, PAS and SRF surveys exceed the accuracy of RSOs from SRF. However, the analysis time increases from days to months due to the volume and difficulty of counting complex images - typically less than 2% of images have any desired targets in them.

Shifting to PAS will require an order-of-magnitude improvement in photographic review times - an AI-assisted approach can provide this improvement. The potential improvement in the reliability of results (improved consistency together with human RSOs) and the reduced costs will make a photographic, AI-assisted aerial census very attractive in the immediate future. Machines will review the massive amounts of photographs and direct our mappers and analysts to areas where they provide the most value to sort images, track and count wildlife, and identify other objects of interest as they appear in images. Our end-to-end workflow from training dataset creation, deep learning models trained with the cloud computers, model iteration, model output validation, model inference, PostgreSQL database design, and data exportation to create an early warning of potential human-wildlife conflicts and proximity maps.