Scaling functions evaluation for estimation of landscape metrics at higher resolutions

Understanding the relationship between landscape pattern and environmental processes requires quantification of landscape pattern at multiple scales. This will make it possible to relate broad-scale patterns to fine-scale processes and vice versa. In this study, we used class level landscape metrics calculated at multiple scales to fit scaling functions that were used to downscale metrics at higher resolutions. The main objectives were to assess the performance of different type of functions (i.e. power, logarithmic, etc.) to downscale metrics at the subpixel level and to analyze the variability of the accuracy of subpixel estimates among patch classes for each landscape metric. We used thirteen frequently used landscape metrics, computed on a land use/land cover map derived from Landsat imagery through visual interpretation and supervised classification using Support Vector Machines. The performance of scaling functions was assessed with the Accuracy Improvement percentage (AI). In general, the power function fitted better for most landscape metrics and classes; however, in several cases, more than one type of function showed similar R2 values. Accuracy of subpixel estimates was very variable among landscape metrics and also among patch classes within a metric. The amount of variation was such that no generalization about the predictability of a landscape metric calculated at the class level was possible. Indeed, predictability seemed to be more of a characteristic of the class than a characteristic of the landscape metric. Additionally, the goodness of fit of the scaling functions was not a good indicator of the functions' ability to downscale landscape metrics accurately, indicating that different scaling functions should be analyzed when downscaling landscape metrics at higher resolutions is required.