Automatic identification of erosion gullies with ASTER imagery in the Brazilian Cerrados

Gully erosion is a serious problem at many locations worldwide, but little is known about its importance at large spatial scales. The remote sensing contribution for the spatial assessment of gullies has thus far been confined to visual image interpretation. The current study was conducted to determine whether automatic classification of optical ASTER imagery could accurately discern permanent erosion gullies in the Brazilian Cerrados. A maximum likelihood classifier (MLC) was trained with two classes, gullies and non-gullies, and applied to images of March (wet season) and August (dry season). Moreover, a bi-temporal classification was performed by labelling a pixel as gully when both for the March and August image it was classified as such. Validation was done with a gully map obtained from a panchromatic QuickBird image and field data. For mono-temporal classification, the March image performed much better than the August image, because spectral differences are more pronounced during the wet season. Based on spatial analysis of output maps, the bi-temporal classification performed best in identifying gullies, as user's accuracy was above 90%, while two of 17 actual gullies were not detected and two small locations were erroneously identified as gully. The combination of ASTER bands 1, 2, 3, and 4 gave highest accuracies. It is concluded that accurate automatic identification of permanent gullies is possible with optical satellite data. Because the Cerrados occupy a vast area, it is expected that the approach presented could be applied to larger areas with similar characteristics