Knowledge Extraction from an Adaptive Neuro-Fuzzy Inference System for Detecting Urban Objects, Case Study: Buildings and Trees

Nowadays, powerful detection systems, which the learning procedure of them is black box and is not available, have been widely used to classify data. However, the understandability of the acquired knowledge from these detection systems can significantly help operator in carrying out classification performance with high accuracy and precision. Hence, knowledge acquisition in a form of fuzzy rule set is an important issue in the image processing that causes to comprehend the classification methods appropriately and to improve them subsequently. The purpose of this paper is proposing a method to extract fuzzy rules in IF-THEN form via an Adaptive Neuro-Fuzzy Inference System (ANFIS) for classification LiDAR data and digital aerial images. Detection of building and tree in urban areas needs to determine some features to perform the detection procedure; because classification algorithms decide about pixel entity based on its feature vector. These features can make the object separation possible by the textural, the spectral, and the structural characteristics. Nowadays, by increasing the number of the active and passive sensors, it is possible to record the textural, the spectral, and the structural characteristics of objects in different wavelengths by various approaches. In this paper, some potentially features were generated, and then optimal features were selected using the genetic algorithm. Using the selected optimum features, an ANFIS was used to recognize the objects accurately. In this regard, at first, the prepared training data was utilized as inputs of grid partitioning algorithm and a Sugeno fuzzy inference system with one output was generated by determining the type and the number of input membership functions, as well as theS type of output membership functions. Then, the grid partitioning algorithm figured out the best state of the membership functions after investigating the whole of the possible states. Afterwards, the training and checking data were entered into the generated ANFIS and during the training procedure, the final classifier was concluded to detect buildings and trees. Finally, by proposing a different fuzzy-based method and using the selected training data, as well as the output membership functions of the proposed ANFIS, a set of effective fuzzy rules were extracted. The proposed method has three main steps. In the first step, the tuned premise parameters (after training process) of inputs training data of ANFIS were extracted according to the mean values of the membership functions. In the second step, firstly, based on the number of membership functions of each feature, the total number of feasible fuzzy rules was determined. Then, for each training data, the fired values for all rules were computed. The rule that had the most effect in the process was chosen as the fired fuzzy rule of each training data related to the desired object class. In the third step, the fuzzy rules which has the importance more than a specified threshold in the classification procedure were extracted. The extracted fuzzy rules were considered and analyzed logically regarding the feature layers, and the results show the high capability of the fuzzy-based proposed method in extracting rules from the objects detection procedure.