Automatic detection and mapping of European ground squirrel burrows on UAV-based multi- and hyperspectral imagery with classification methods

European ground squirrels (EGS) are members of the soil megafauna and part of the ecosystem engineers that shape physical, chemical, and biological characteristics of soil ecosystems in European grasslands. Thanks to their strict protection their abundance and distribution have been surveyed systematically and annually in Hungary. The results of their 20 year monitoring indicate that their population is declining, there are sudden extinctions of local populations, and a desynchronized variation of the abundance of local populations occur either spatially or temporally.

The monitoring protocol involves the estimation of their abundance in each colony by a strip-transect method and the habitat-colony area by visual observations or digital maps. Both approaches use animal burrows as proxies for either their presence (colony area) or density (colony size). These estimation methods, however, consist of systematical errors: first, they consider the animals’ density to be even over the entire habitat-area, second, they conjecture that animals occupy the habitable area completely, and third, evenly. If we were able to survey distribution and abundance of EGS more accurately, frequently, and efficiently, we could better intervene in time when local populations begin to decline or before they disappear. In addition, we could better estimate the effects (+ or -) of management strategies in real time.

The primary aims of our study were to develop a non-invasive, semi-automated method for (1) estimating abundance of EGS in the area of occupation of a colony, and (2) delineating their occupancy within the habitable area. We have defined burrow openings and mounds as quantitative proxies for the presence of animals. We have started to develop a monitoring technique to identify, locate, and count objects of interest in images automatically and to delineate the area of occupancy by identifying those objects of interest from the surroundings. To survey EGS colonies and habitats we have used a multirotor platform UAV equipped with either an RGB visible-range or a hyperspectral sensor.

To test our method several pilot areas with different vegetation and relief were surveyed. Acquired aerial images have been processed by photogrammetric software and resulting high spatial resolution orthomosaics are classified by machine-learning algorithms (randomforest, CART, C5.0) implemented in a custom R script. As detection of mounds and openings are visually restricted by vegetation height (e.g. grass, shrubs, weeds, herbs), we have studied the effect of grass height on detection success. Preliminary results suggest that successful classification can be performed either on RGB visible-range and hyperspectral images. However, the appropriate spatial resolution (below cm range) and the presence of high grass are more important key factors than number of spectral bands.

Detecting EGS burrow openings and mounds is based on surface characteristics of EGS burrow openings and mounds consequently the method is being developed for EGS specifically but can be modified to the characteristics of other burrowing mammals of this size (e.g. mole-rats, moles).