Classification model METHODOLOGY AND CLASSIFICATION MODEL

Figure 1. Test areas: No. 1 – suburban area of Warsaw, No. 2 – rural landscape of Western Pomerania. The research on the formulas and values of certain indices consisted in analyzing spectral for NDVI and textural granulometric maps data regarding different examples of 5 land cover types being the subject of the classification. Different types of granulometric analysis based on different types of opening and closing operations such as simple operations and operations with a multiple structuring element were tested regarding separability of different classes on specific granulometric maps. Figure 2 presents samples of land cover classes visualized on a spectral colour composition column A and a colour composition consisting of different granulometric maps based on operations with multiple structuring element – column B. A B 1. 2. 3. 4. 5. Figure 2. Exemplary samples of land cover classes in a spectral colour composition and in a colour composition consisting of granulometric maps: water 1, forest 2, low vegetation 3, bare soil 4, built-up area 5. The performed tests and analysis allowed to establish the final form of the classification model as well as to evaluate the optimal values of indices.

2.3 Classification model

Basing on the research performed on test images, we proposed the following model of classification, presented in figure 3. The model consists of 3 steps. The first step, an analysis of spectral data, relies on NDVI values. Its purpose is to distinguish basic classes, such as water, vegetation and non-vegetation, which all differ significantly spectrally, and can thus be easily extracted basing on spectral analysis. Previous analyses allowed to evaluate V1 and V2, values of NDVI: V1 distinguishing water and non-vegetation i.e. bare soil and built-up area, and V2 distinguishing non-vegetation from vegetation i.e. forest and low vegetation. This step allowed to extract 3 groups of classes according to the model - “GROUPS”. The second step relies on granulometric maps. The goal of this step is to distinguish classes previously grouped based on NDVI values: low vegetation and forest, and bare soil or built-up area. Due to the different texture form of high texture classes, i.e. forest and built-up area, significant textural information is provided by different granulometric maps for different classes. Thus, values G1 and G2 refer to two indices based on values of granulometric maps: from C1 to C7 closing and from O1 to O7 opening. This step allowed to distinguish 5 preliminary classes within the groups see figure 3. The third and final step again relies on spectral values. The NIRB index helped to distinguish certain parts of forest from built-up area. Because of low NDVI values caused by low NIR values, the shaded parts of forest were assigned to the non- vegetation group in the first step of the process, and due to high texture, they were classified as built-up area. The NIRB index and the B1 parameter allow to correct the assignment of these misclassified pixels. This step allowed to fix the final classes. Figure 3. Block diagram of the classification model. Number of classes according to the list presented in section 2.2. As it is shown in the model scheme figure 3, each class shall meet a set of conditions presented below. 1. ‘Water” – this is the only class in Group 1, and therefore all the pixels belonging to this group are assigned to This contribution has been peer-reviewed. doi:10.5194isprsarchives-XLI-B7-277-2016 279 Class 1. The condition for belonging to Class 1 can therefore be determined as follows: GROUPS = 1 2. “Forest” – it is a class of Group 2 it also contains class 3 “low vegetation” which is characterized by a stronger, more heavy texture than Class 3 see figure 2.2. In order to distinguish it from class 3, four granulometric maps are used after opening and closing with the lowest index C1, C2, C3, C4, O1, O2, O3, O4. The “Forest” Class is represented by pixels with high values on those granulometric maps, where the sum of all 8 maps exceeds the limit described as G1. The condition of pixels belonging to Class 2 can be determined as follows: GROUPS = 2 AND C1+C2+C3+C4+O1+O2+O3+O4 G1 3. „Low vegetation” – is a class in Group 2 characterized by a weaker texture and, therefore, lower than Class 2 values on the granulometric maps compare figure 2.3. In order to distinguish it from Class 2, the same eight granulometric maps are used: C1, C2, C3, C4, O1, O2, O3, O4. The “low vegetation” class is represented by pixels with high granulometric values on these maps, where the sum of all 8 maps does not exceed the limit value, defined as G1. The condition of pixels belonging to Class 3 can be determined as follows: GROUPS = 2 AND C1+C2+C3+C4+O1+O2+O3+O4 = G1 4. “Bare soil” – this is the class of Group 3 characterized by less heavy texture, and hence lower values on granulometric maps compared to the second class of this group – “Built-up area” Class 5. Since the difference in the texture of these two similar spectral classes is most evident on a slightly smaller scale, granulometric maps with higher indexes must be used: C4, C5, C6, C7, O4, O5, O6, O7 . The “Bare soil” class is represented by pixels with low values on those granulometric maps, where the sum of all 8 maps is less than the limit defined as G2. The condition of pixels belonging to Class 4 can be determined as follows: GROUPS = 3 AND C4+C5+C6+C7+O4+O5+O6+O7 G2 5. “Built-up area” – this is another class of Group 3 which has a clear, strong texture, and consequently, high values on granulometric maps, due to the size of the texture grain – that is on those maps with higher indexes. According to the proposed solution, the “Built-up area” Class is represented by pixels, whose sum of the values in all 8 maps is not less than the limit defined as G2. The condition of pixels belonging to Class 5 can be determined as follows: GROUPS = 3 AND C4+C5+C6+C7+O4+O5+O6+O7 = G2 As observed on the test images in the case of deciduous forest with a specific form i.e. not fully spring, non-consistent canopy or sparse forest, there may be tiny shaded areas. These surfaces belong, according to the specification of classification, to Class 2 –“Forest”, but are characterized by low values of NDVI and high values on granulometric maps. Consequently, in the second step, it can be assigned to Class 5 “Built-up area”. For this type of surface, a condition relating to the value of the quotient indicator NIR B was introduced. This surface receives higher values than actual built-up areas. Pixels classified as built-up area Class 5, but having high values of this ratio and exceeding the limit value – defined as B1 – will be assigned to Class 2. The condition of changes from Class 5 to Class 2 can be determined as follows: CLASSES1 = 5 AND NIRB B1

3. RESULTS