Znaczenie pola powierzchni i długości obiektów w półautomatycznej klasyfikacji obiektowej użytków zielonych na zdjęciach satelitów serii LANDSAT

Semi-automatic method for object classification of the grassland procedure involves two stages: 1) the creation of image segments as a representation of natural spatial complexes, 2) classification of the segments. So far, the classification algorithms were used refer to the three categories of characteristics: spectral, panchromatic or geometric. In the first stage of the work segmentation were performed of the composition of the two satellite images Landsat7 acquired at different seasons of the year: in September 1999 and the beginning of May 2001. Panchromatic data were used for distinguishing complexes due to the greater (in comparison with spectral data) spatial resolution. In the area of grasslands landscape-vegetation complexes (Matuszkiewicz, 1990, Kosiński, Hoffmann -Niedek, Zawiła, 2006) were distinguished of approximately a hundred to a few hundred meters in length and of about 20 ÷ 200 panchromatic image pixels. Semi-automated delimitation of complexes were carried out under the visual control, using as auxiliary material aerial photographs and topographic maps. In the second stage (classification of segments) an attempt were taken to assess the suitability of selected geometrical features to distinguish grasslands in use (currently or potentially) from grasslands unfit for production use due to excessive or insufficient moisture. The classification algorithm used GIS tools for measuring area and length of segments and artificial neural networks as a tool for classification. The previous studies of the Piotrkowska Plain show that the complexes of meadows used differ from those abandoned in terms of size and shape of objects (Kosiński, Hoffmann- Niedek, 2006, Fig. 1). Hypothesis that area and length of the landscape -vegetation complex are cues of identification in relation to the use and moisture of grasslands. 43 complexes of the grassland have been established as training samples on the Piotrkowska Plain in the Pilsia valley. In order to avoid overfitting classification algorithm to data from the Piotrkowska Plain, in order to allow the application of the algorithm for another mezoregionu 10 complexes have been selected as a validation set in the Szczercowska valley. To evaluate the classification results 32 complexes have been collected from Szczercowska Basin (test set). All treining set objects were described in terrein. Validation and test set objects were classified by a more accurate metod (based on biteporal image: Kosiński, Hoffmann -Niedek, 2008) and checked at random in the field. Objects of learning, validation and test set have been grouped into five categories according to use and habitat moisture (Kosiński, Hoffmann -Niedek, 2008; Table 1). For learning neural networks fife categories of objects of the learning and validation set were generalised into the three classes. In the Szczercowska Valley combination of characteristics (area and length) of the abandoned complexes is more close to the meadows in use than on the Piotrkowska Plain (Table 2). Therefore, the classification algorithm of the Piotrkowska Plain can not be directly applied to Szczercowska Basin. To obtain the correct result of classification, the classes of test set has been interpreted differently than in the learning and validation sets (Table 3, Figure 2). In the test sample 3/4 of the 23 complexes of meadows potentially used were classified correctly, while of nine abandoned ones due to unfavorable moisture habitats correctly classified 2/3. Thus confirmed the working hypothesis. Application of artificial neural networks can cancel the designation of non parametric empirical indicators of the size and shape of the complexes (Fig. 1). Neural networks auto-uwilgotmatically builds a morpfometric model based on simple indicators such as area and length of the object. Two model types of artificial neural network have been tested: 1) multilayer perceptrons (MLP) wich use hyperplanes to divide up feature space, 2) radial basis function network (RBF) wich use hyperspheres. MLP networks have proved to be more suitable to build the model than the RBF network.