Adaptive algorithm for parameterization of feature extraction techniques in remote sensing images
MetadataShow full item record
Remote Sensing Images (RSI) involves identification of a Region of Interest (Roi) from a suitable Feature Extraction Algorithm (FEA) that best extract the identified Roi, a process known as Dimensionality Reduction. This process can yield better results if the FET's parameter selection strategy specifies an optimum parameter value combination. As a result, parameterization plays an important role in the use of FET algorithms. Most research work on improving parameter selection strategies are based on empirical experiments, which are often characterized by short runs that in most cases result in invalid conclusions. Although empirical experiments have produced acceptable results, this strategy tends to bias the parameter distribution to a narrow range thereby reducing the capability to discriminate the modeled objects. Therefore, the research study presented a novel adaptive heuristic algorithm based on the Gabor Filter (GF) to generate useful solutions to optimization of parameter selection strategies for FET in RSI. The adaptive heuristic named "GenApp" implements a genetic approach that is premised on Genetic Algorithms (GA). The GenApp uses a 16bits random generator to generate phenotype values that are feed into a genAid simulator to generate genotypes values. GenApp was evaluated against existing algorithms namely the GF and the MGF for efficiency, fitness of value and the worst case scenarios. The results presented in our simulations show the GenApp algorithm having the best fitness value in five (5) random simulation runs. On the other hand, the GenApp posted the highest complexity values of 0.63 and 0.03 points above the existing algorithms. This lapse is attributed the initialization steps of the GenApp that screen the initial data of population. The screening process eliminates the worst-case scenario of the GenApp algorithm; hence this could be seen as a trade-off in its overall performance as compared to other existing algorithms that fail to screen the initial population.