Comparison of effects of spatial autocorrelation on distribution predictions of four rare plant species in the Watarase wetland

We tested the effectiveness of distribution-prediction models for four rare herbaceous wetland species in the Watarase wetland, Japan, based on data obtained from aerial images. We used visible and near-infrared aerial images from three seasons, and elevations and vegetation heights derived from the images. Because spatial autocorrelation in species distribution data often biases the estimated effects of certain variables and reduces the prediction accuracy of distribution models, we compared the predictions of an intrinsic conditional autoregressive (CAR) model, which accounts for spatial autocorrelation, with those of a standard logistic regression model. The four study species had different distribution patterns: Ophioglossum namegatae and Impatiens ohwadae had aggregated distributions, whereas Galium tokyoense and Thalictrum simplex var. brevipes had scattered distributions. Predictions based on remote sensing images performed well for O. namegatae with the intrinsic CAR model and for I. ohwadae with both the logistic and CAR models; performance was poor for G. tokyoense and T. simplex var. brevipes with both models. Prediction accuracy improved by the CAR model in comparison to the logistic model most in O. namegatae and least in I. ohwadae. Impatiens ohwadae’s distribution was explained well by ground height. In contrast, the apparent improvement in the prediction for O. namegatae resulted from a substantial spatial random effect, suggesting the presence of determinants that could not be detected by remote sensing. The number of explanatory variables with large effects decreased in the intrinsic CAR model in three species possibly by avoiding spatial pseudoreplication, but not for T. simplex var. brevipes.