Aerial photogrammetry and machine learning techniques to enable improved monitoring of human-wildlife interaction

Wildlife population monitoring is crucial to human wildlife interactions, general conservation and sustainable wildlife use. It may, among other applications, determine wildlife population health, species composition and... [ view full abstract ]

Wildlife population monitoring is crucial to human wildlife interactions, general conservation and sustainable wildlife use. It may, among other applications, determine wildlife population health, species composition and distribution. It may also help wildlife managers predict the location and severity of human-wildlife conflict events, rangeland degradation, human settlement encroachment and poaching activity. Although technological improvements such as telemetry and camera traps have improved wildlife managers’ ability to monitor wildlife abundance and distribution, the advent of inexpensive aerial photogrammetry and affordable aerial vehicles have made it possible to produce high resolution imagery of extensive areas at low cost. Despite this, wildlife managers have largely persisted with inaccurate and outdated aerial census and ground monitoring techniques. Our study explores the use of aerial imagery at a resolution of 4-9 cm, and its effectiveness in identifying and counting wildlife across 2.74 million hectares of Namibian and Angolan protected areas, as well as 75 000 ha of southern African savannah. We demonstrate a multi-disciplinary approach, which includes photogrammetry, geospatial science, remote sensing and machine learning techniques to monitor wildlife populations and human settlement characteristics in protected areas. We found it possible to accurately identify southern African wildlife of springbok (Antidorcas marsupialis) size and larger in imagery of resolution higher than 8 cm, in the arid desert and semi-arid savanna of Namibia and Angola. In savannah it was possible to accurately detect elephant (Loxodonta Africana) at a resolution of 8-10 cm. We described diagnostic features for manual identification of wildlife and differentiating species. We further used machine learned computer algorithms to detect wildlife at a 75% true positive rate for a single false detection per image processed, when trained on approximately 10 000 images of a particular species. The combination of machine learning and specialized software (to verify results and improve detections) greatly reduced the time needed for manual identification.  We further found the imagery able to detect rangeland degradation, human settlement encroachment, potential poaching activity, as well as botanical and archaeological features. We explored the necessary future steps, including overcoming obstacles and limitations, to the large-scale uptake of the technology in the human-wildlife interactions discipline.