Modelling butterfly distribution based on remote sensing data

Publisher Website
Google Scholar
Abstract
Aim: We tested the usefulness of satellite based remote sensing data and geographical information system (GIS) techniques (1) in explaining the observed distribution of the threatened clouded apollo butterfly (Parnassius mnemosyne) and (2) in predicting the occurrence of the butterfly in two independent test areas with different landscape structure.Location: The three study areas are located along the rivers Rekijoki and Halikonjoki in a boreal agricultural landscape in south‐western Finland (60°40′ N; 23°20′ E).Methods: Landsat satellite images were used to generate habitat maps of the three study areas. Topographical variables were calculated from a digital elevation model. These data were used to construct a multiple logistic regression model fitted with the observed occurrence of clouded apollo in 126 grid squares of 0.25 km2within an area of 31.5 km2(model building area). The parameterized model was used to predict the occurrence of the butterfly in an adjacent test area with a similar landscape structure as in the model building area, as well as in a more distant test area with a more fragmented pattern of semi‐natural grassland.Results: In the model building area probability of clouded apollo occurrence increased with connectivity of semi‐natural grassland, cover of deciduous forest, cover of semi‐natural slope grassland and topographical heterogeneity. The model accuracy was high, the correct classification rate being 98.4% overall and 95.0% for the butterfly presence squares. In both test areas the overall correct classification of the model prediction was high (c. 92%). However, in predicting the actual butterfly presence squares the model succeeded substantially better in the adjacent than in the distant test area (correct classification rates 91.3% and 66.7%, respectively).Main conclusions: The results showed that the distribution of a habitat specialized butterfly may be quite successfully explained and predicted based on pure satellite imagery and topographical data. Nevertheless, the decrease of accuracy of the model prediction when applied to a different landscape structure than the one in which the model was parameterized suggests that the useful application of such models is limited by the environmental variability of the original model building data.