Adaptive neuro-fuzzy ( anfis ) building detection in high resolution optical images

Importance: The building detection is an important issue for the interpretation of remote sensing data which includes images and map layers in geographic information databases. The verification and creation of maps and GIS data, automatic analysis of land use, sealed areas measurement for public organization uses, etc. are amongst the most common applications for automatic building detection. Buildings are of substantial objects in remote sensing images and directly points to residential areas. In most practical cases, buildings are well identifiable objects by a human perception. Thus, an automatic system that is able to mimic human operator cognition is in fact desired. Introducing of the problem along with data used: Detection of building blocks in spectral images is a complex and non-trivial task as the overall spectral characteristics of buildings are usually not constant and instead are very varied and not easily recognizable from the surrounding area. The Neuro-Fuzzy system incorporates the human-like logic style of fuzzy computation systems through the use of fuzzy sets and a linguistic model comprised of multi sets of IF-THEN fuzzy rules. A Neuro-Fuzzy system approximates an N dimensional function that is partially defined by the provided training data. The fuzzy rules encoded within this system play as non-exact samples, and can be considered as prototypes of the imported training data. Method used: In this study, we propose an approach for building detection procedure, using high-resolution satellite imagery by the means of an intelligent hybridization system namely ANFIS (the adaptive neuro fuzzy inference system). This approach improves object detection accuracy by reducing the premature convergence issue and local minima problems encountered when using neural network algorithms. A feature vector for possible building blocks is selected manually and calculated so the buildings are extracted using neuro-fuzzy classifier. Fuzzy membership functions were selected symmetrically triangular. The averaging defuzzification method was considered in the output section of fuzzy system. Parameters of neural network module were iteratively updated using the fuzzy module and by this the overall performance of the algorithm improved both in theory and practice. To initiate the approach, training samples are selected that represent the specified two feature classes, in this case “building” and “non-building”. This process is supervised so we have to introduce the algorithm with proper training data. At the end of each training cycle, the adaptive-fuzzy module determines the new (adjusted) parameters of the network. This evolutionary process repeats until a specified number of cycles have been reached. To enhance the detected building patches, morphological image processing operations are applied. Analysis and conclusion: The detection performance of the algorithm was much better for simple and rigid buildings than for complex and high dimension buildings. The approach was tested on a test scene using 1 m resolution pan-sharpened IKONOS image. The calculation and analysis results, performed by MATLAB/Simulink, represents that the proposed method has a small non-detection region. Also, this method is capable of detecting specified objects accurately within the minimum standard time. The kappa analysis for the proposed adaptive Neuro-Fuzzy algorithm approach was clearly promising. The runtime of the present approach depends on the image content. The low-level processing requires the major part of the program runtime.