Main content area

Combining multi-scale textural features from the panchromatic bands of high spatial resolution images with ANN and MLC classification algorithms to extract urban land uses

Saboori, Mojtaba, Asghar Torahi, Ali, Reza Riyahi Bakhtyari, Hamid
International journal of remote sensing 2019 v.40 no.22 pp. 8608-8634
algorithms, altitude, bitumen, land use, neural networks, panchromatic imagery, regression analysis, satellites, spectral analysis, texture, vegetation
High spatial resolution images have been increasingly used for urban land-use classification, but high spectral variations within same land use, the spectral confusion among different land uses, and the shadow problem often lead to poor classification performance of the traditional per-pixel classification methods. Main objectives of this paper were to extract phenomena with different altitudes with the absence of elevation features and shadowed areas without defining a shadow class, identifying the most effective textural features in classification by Regression analysis and also class differentiation with similar spectral properties. To achieve these aims, the panchromatic image of WorldView2, GeoEye1, and QuickBird satellites were applied in order to extract the statistical features of the first and the second order of multi-scale texture analysis, due to high potential for providing more detailed and high spatial resolution in five different window sizes, four different cell shifts, and three different angles or directions. Overall, 137 features were used as input in two classification algorithms including Maximum Likelihood Classifier (MLC) and Artificial Neural Network (ANN). The results showed that the multi-scale textural features and ANN made possible to differentiate three major classes of asphalt, vegetation and building surfaces even with the presence of shadowed area and the absence of elevation features. The experiments also presented that the more the elevation of vertical objects, the more the effect of textural parameters on extraction of these classes. Furthermore, the investigations denoted the validity of the Regression analysis in the detection of most effective textural features in classification.