Unclassified landsat TM predicts bird distributions at fine resolutions in forested landscapes

Background/Question/Methods

Land-use and land-use change are important drivers of biodiversity and species distributions yet their use in species distribution modeling has been limited by the availability of fine resolution data appropriate for most species responses at broad scales.  Remote-sensing technology offers great potential for predicting species distributions at large scales, but the costs of processing and required expertise is prohibitive for many applications especially in developing countries.  Raw reflectance data is readily available for all parts of the globe, can be obtained at minimal cost and can be useful in predicting species distributions.  We compiled data from remote-sensing spectral reflectances to test whether they can be useful predictors of species distributions of 40 commonly-occurring bird species in western Oregon.  Information on bird observations was collected from 4,375 sample points using fixed-radius point counts.  Reflectance data was obtained using Landsat imagery acquired from the USGS; spectral values for 6 bands were summarized at scales of 150m, 500m, 1000m and 2000m.  To analyze relationships between species distributions of birds and reflectance values, we used boosted regression tree (BRT) models.  BRT models have several advantages for model fitting, being able to model non-linear relationships and interactions between predictor variables.  We evaluated predictive performance of the models using AUC values and compared the relative influence of predictor variable across six bands and four scales.

Results/Conclusions

Prediction success was high with an average AUC value of 0.87 (SD = 0.067) across all models.  Reflectance bands varied in their contribution to the prediction of species distributions.  The relative influence of predictors for band 4 was three times greater than the next highest band indicating the high predictive success of these variables which are often associated with vegetation biomass and photosynthetic activity.  Relative influence also varied by scale with the average influence of predictors at the 2000m scale being over four times greater than other scales suggesting that species may be responding to land-use patterns at these scales. In this paper, we demonstrate that variables obtained from raw unclassified remote-sensing imagery can be used to produce species distribution models with high predictive ability. Our study is the first to identify some general patterns in the usefulness of spectral reflectances for species distribution modeling of multiple species in a region. We conclude that raw remote-sensing data can be useful for several applications including forest management to address habitat loss and disturbance, change detection studies, and reserve selection and evaluation.