Predicting species diversity in agricultural environments using Landsat TM imagery

Maps based on classified Earth observation (EO) imagery have been used to model biodiversity, but errors associated with the classification process itself and the resulting discretization of land cover may ultimately limit such efforts. Among other issues, discrete land cover maps can often be costly to produce and validate. Alternatively, the original continuous spectral information in EO imagery can be used. The primary objective of this study was to compare predictors based on continuous and discrete information derived from Landsat TM imagery for modeling biodiversity in agricultural landscapes. In 46 landscapes throughout Eastern Ontario, Canada, landscape metrics (mean field size, the percentage of landscape in agriculture, and crop diversity) derived from a discrete image classification, along with several measures of crop productivity based on the continuous Normalized Difference Vegetation Index (NDVI), were used as predictors of field-based measures of species diversity for birds, butterflies, and plants. Using an Information-Theoretic approach for model-averaging and inference, we compared and interpreted the magnitude and direction of model-averaged coefficients, model evidence ratios, and overall fit of model-averaged predictions. Our findings indicate that when using Landsat TM imagery in agricultural environments, models using predictors derived from continuous information consistently outranked models based on discrete information derived from classified imagery.