Mapping badland areas using LANDSAT TM/ETM satellite imagery and morphological data

Abstract Soil erosion is the most influential component of land degradation for its strong impacts on both natural and agricultural environments. In order to support effective intervention and recovery policies for eroded areas, monitoring techniques should take into account the space–time variability of the processes involved, and make use of assessed mapping methodologies as baseline criteria for studying the dynamics of landform development. When using multispectral data for mapping eroded areas, low spectral separability is a significant limit in areas with complex features, where soil materials are frequently remixed by surface runoff. Since multispectral satellite images are a valuable data source for multi-temporal analyses of erosion processes at the medium scale, we assessed how accurately a badland area can be identified from LANDSAT TM and ETM data. A protocol for an optimal mapping was built up by testing the performance of different supervised algorithms and input layers (spectral and morphological). Tests were carried out in a well-known badland area of Basilicata, Italy, with an extension of ∼ 8000 ha. Results obtained from the use of spectral bands (with and without thermal channel) and principal components returned an overall accuracy ranging from 53% (for classification on first three components) to 72% (for classification on all bands from TM), with low values for the kappa coefficient (0.30–0.50), showing that the spectral information alone are insufficient to accurately identify badland areas. In order to improve mapping, we found that the integration of slopes and aspects derived from a Digital Elevation Model (DEM) can overcome problems inherent to the low separability of spectral signatures. The use of morphological data was tested for different classification algorithms and integration approaches. In discriminating badlands, the better performing algorithm was MLC (Maximum Likelihood Classifier) and the best results were obtained by integrating all seven bands (including TIR) with slope and aspect maps as input within the classifier ( A  > 0.85 and K  ∼ 0.75 for both the sensors). We selected such parameters because they play an important role in characterising badlands of study area from a morphological perspective but the proposed approach is also conceptually simple and can be easily exported to other areas. The obtained results support the hypothesis that the combined use of remote sensing imagery and auxiliary morphological data significantly improves the mapping of badlands over large areas with heterogeneous features, thus providing a useful methodology for long-term studies on soil erosion processes.