Building Detection from High-resolution Satellite Imagery Using Adaptive Fuzzy-genetic Approach

We present a technique for extracting the buildings from high resolution satellite imagery using adaptive fuzzy-genetic approach. The technique was inspired from the genetic image exploitation system, GENIE PRO, conducted by Perkins et al., (2005) but brings an important novelty, which is an adaptive-fuzzy module that fine-tunes the genetic algorithm parameters aiming to improve the feature extraction performance. The technique integrates the well known genetic algorithm concepts such as population, chromosome, gene, crossover and mutation into the fundamental image processing concepts. The population is defined as the set of chromosomes, which consists of a predetermined number of image processing operations (genes). The genes are comprised the basic image processing operations. The algorithm is initiated by selecting the training samples for the building and non-building areas from the imagery. The image processing operations are applied in a chromosome-by-chromosome basis to obtain specific attribute planes. These planes are then fed into Fisher Linear Discriminant (FLD) module, which finds an optimal discriminating hyper plane between the building and nonbuilding features. Next, for each chromosome, the fitness values are calculated by analyzing the detection and mis-detection rates. After that the crossover and mutation operations are applied to arbitrary chromosome(s) to create a better population in the next generation by diversifying the current population. At the end of each generation cycle, the crossover and the mutation probabilities are adjusted by the adaptive-fuzzy module for the next generation. The evolutionary process is repeated until a satisfactory level of iteration is reached. Finally, a post-processing operation is performed in order to enhance the extracted building polygons by means of the morphological image processing operations. The approach was implemented on a selected urban area of the city of Ankara, Turkey using the 1-m resolution pan-sharpened IKONOS imagery. The study was found to be quite promising since the building regions were successfully extracted with an approximate detection rate of 90%.