Improvement of Building's Roof Detection via Intelligent Fusion of Current Algorithms in Hyperspectral Images

Today, in order to achieve more complete and precise information extracted from the complex and heavy hyperspectral data, there have been introduced more specialized and advanced methods for analyzing data. Among the effective methods is the spectral target detection algorithms, at the first step 14 algorithms have been evaluated in three groups, including deterministic measures, statistical measures and anomaly detection. Then the higher accuracy algorithms in each set have been selected as follows: Spectral Correlation Similarity (SCS) and Spectral Angle Measure (SAM) from deterministic measures set, Spectral Information Divergence (SID) and Jeffries-Matusita Distance (JMD) from statistical measures set, and Covariance-based Matched Filter Measure (CMFM), Constrained Energy minimizing (CEM) and Correlation-based Matched Filter Measure (RMFM) from anomaly detection set. These seven algorithms are fused by different strategies and methods. There were two strategies, including sequential and simultaneous fusion too. In the sequential fusion strategy, the mentioned selected algorithms from each set were fused in the first step and their results were fused in the second step by using these four Boolean, Euclidian, Fuzzy Inference System (FIS) and Adaptive Neuro-Fuzzy Inference System (ANFIS) fusion. In the simultaneous fusion strategy, the results of the mentioned seven algorithms were fused at once by ANFIS method. The experiments on the roof detection application, through a compact Airborne Spectrographic Imager (CASI) hyperspectral imagery, taken from one of the Toulouse urban areas in southern France, showed that the simultaneous fusion strategy via ANFIS was the most precise automatic one among other primary and synthetic target detection methods.