Multifractal Parameters in Prediction of Land-use Components on Satellite Images

The aim of the study was to assess the usefulness of multifractal parameters as global features describing the content of very high-resolution satellite images. It was done in a research experiment aiming at prediction of basic land-use classes shares within the image tiles cut from satellite EROS-A images. The reference land use data were obtained by on-screen digitizing of aerial ortophotomaps. The shares of built-up areas, agriculture areas and forests in the area of every image tile (ca. 1 sq. km) were calculated. Image tiles were also characterized using wide set of selected global textural features. Apart from multifractal parameters and fractal dimension we used histogram-based features as well as parameters based on co-occurrence matrix, run length matrix, absolute gradient, autoregressive model and wavelet analysis. The Cubist algorithm was applied to predict the percentages of each land use class within the image tiles. The results were evaluated using RMSE, MAE and R squared. When particular groups of textural parameters are considered, the best results in our experiment were obtained for absolute gradient-based features and multifractal features. The outcomes of presented study confirmed our previous findings that multifractal parameters should be considered as useful descriptors of high-resolution satellite image content.