The Western Ghats of India spans 180,000km2 and represent 30 percent of the plant and animal biodiversity of India while only occupying six percent of the total land area of India. The Bandipur and Nagarahole National Parks are subsets of this ecosystem occupying about 1200km2 in Karnataka. It is home to many tribal groups and nearly 30 percent of the global population of tigers and Asian Elephants. Although humans and wildlife occupy separate habitats, growing human population pressure and varying ecological conditions have led to conflicts.
During a visit to schools for tribal children located in these forests, YSR realized that these villages are often inaccessible because they lack roads and trails. There is a need for mapping trails and land-use so that the local inhabitants could obtain basic goods needed for survival. Here we propose to apply deep-learning and computer vision algorithms to map out different types of land usage, like trails and roads, on sentinel-2 satellite imagery taken at 10m resolution on a yearly basis. Algorithms already developed have been shown to have a superior resolution of paths, edges and boundaries. Using satellite imagery directly can be updated more frequently than Google and OpenStreetMaps.
Results/Conclusions
YSR has collected training data by manually marking different types of land usage in a small subsection of the Western Ghats. YSR uses these training data to build convolutional neural networks to delineate different types of land use in this region including the existence of roads. Sentinel-2 data include not only red, green, blue wavelength but also near-infrared. These types of data are valuable when automating detection of different types of landcover. Computer vision techniques such as NDVI and NWDI use near-infrared to classify different types of vegetation and the existence of water. Similar tools have been used to measure deforestation. Using these techniques YSR will also be able to determine precise boundaries of unique objects and areas as well as map wildlife habitats and different types of agricultural encroachments. Further research will be conducted on other types of satellite imagery that are available on a daily basis at a 3m resolution. We plan on developing monitoring systems to alert wildlife conservation agencies of previously unknown encroachment areas. The ultimate goal of the project is to apply dynamic “algorithmic zoning” approaches to include the complex interaction of human and wildlife in developing and monitoring the changing ecological zones.