Estimation of vegetation parameter for modeling soil erosion using linear Spectral Mixture Analysis of Landsat ETM data

Soil conservation planning often requires estimates of soil erosion at a catchment or regional scale. Predictive models such as Universal Soil Loss Equation (USLE) and its subsequent Revised Universal Soil Loss Equation (RUSLE) are useful tools to generate the quantitative estimates necessary for designing sound conservation measures. However, large-scale soil erosion model-factor parameterization and quantification is difficult due to the costs, labor and time involved. Among the soil erosion parameters, the vegetative cover or C factor has been one of the most difficult to estimate over broad geographic areas. The C factor represents the effects of vegetation canopy and ground covers in reducing soil loss. Traditional methods for the extraction of vegetation information from remote sensing data such as classification techniques and vegetation indices were found to be inaccurate. Thus, this study presents a new approach based on Spectral Mixture Analysis (SMA) of Landsat ETM data to map the C factor for use in the modeling of soil erosion. A desirable feature of SMA is that it estimates the fractional abundance of ground cover and bare soils simultaneously, which is appropriate for soil erosion analysis. Hence, we estimated the C factor by utilizing the results of SMA on a pixel-by-pixel basis. We specifically used a linear SMA (LSMA) model and performed a minimum noise fraction (MNF) transformation and pixel purity index (PPI) on Landsat ETM image to derive the proportion of ground cover (vegetation and non-photosynthetic materials) and bare soil within a pixel. The end-members were selected based on the purest pixels found using PPI with reference to very high-resolution QuickBird image and actual field data. Results showed that the C factor value estimated using LSMA correlated strongly with the values measured in the field. The correlation coefficient (r) obtained was 0.94. A comparative analysis between NDVI- and LSMA-derived C factors also proved that the latter produced a more detailed spatial variability, as well as generated more accurate erosion estimates when used as input to RUSLE model. The QuickBird image coupled with field data was used in the validation of results. (C) 2007 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.