Field of Study :
Obtaining accurate buildings information is needed for many applications such as surveying, management of urban area, as well as for the tracking and prevention of uncontrolled urbanization, etc. In this paper, we proposed a multi-cue classification model to integrate spectral, height and textural features. This model is designed to improve the accuracy of urban classification using scikit learn python libraries and QGIS. In this study, multi-source image fusion techniques was used. The current study includes two study areas. The first study area extracted buildings from Pleadeis images. The second study area extracted buildings from World View-2 and Spot-5 imagery, height were combined to detect the urban information.
For the first study area, classification algorithms such as the conventional classifier (maximum likelihood) and two machine learning classifiers (support vector machine, and backpropagation) were compared to the output of the Random Forest algorithm. The overall accuracy for Random Forest, maximum likelihood, support vector machine, neural network was found 97%, 95%, 93% and 92% respectively. The results showed that random forest was the best.
For the second study area, World View-2 and Spot-5 data were fused using three image fusion techniques (Modified-IHS, WAVELET PCA, and WAVELET-IHS). The Grey-Level Co-occurrence Matrix (GLCM) approach was also applied to determine which attributes are important in detecting and extracting urban areas. The urban mapping performance was analysed using deep learning techniques; Random Forest approach and Support Vector Machines. Normalized Digital Surface Model (DSM) was also generated with 0.5m resolution. Normalized DSM is also used with fused image as input for the classifications. The spectral and textural features extracted from fused imageries, height data and the fused image were used as input for Random Forest classification as well as for Support Vector Machines Classification. The performance of different combinations were evaluated based on their classification accuracies. The results demonstrated that when textural features of colour images were introduced as the input of the classifiers, the overall accuracy, F1 measure and kappa index were improved in the most cases. This followed by integrating Normalized DSM as a layer with the fused image comparing to using the fused image only. Using textural features improved the accuracy of classification, and the overall accuracy reached 0.977, 0.972, 0.977 of Wavelet-IHS, Wavelet-PCA and Modified-IHS respectively using Random Forest classification and overall accuracy reached 0.968, 0.973 and 0.967 using Support Vector Machines classification.
Team Work of Project
Rania Elsayed Ibrahim,
Division : Aviation and Aerial Photography