Solving Optimization and Inverse Problems in Remote Sensing by using Evolutionary Algorithms

Kurzfassung

This thesis objective is the solving of combinatorical and inverse problems in Remote Sensing by using genetic algorithms. The first part introduces optimization theory. Four different deterministic local search algorithms are reviewed. Differences and similarities between these algorithms are examined, also their behavior in a representative test domain. Then the theory of evolutionary computing is explained. It is shown, that evolutionary algorithms are in contrast to the previous discussed search algorithms not deterministic, but heuristic. Furthermore the difference between local and global search is pointed out. An genetic algorithm, which is inspired by evolutionary algorithms, is developed. The program is written in an object oriented style using C++. This program is tested with several test functions which are common in global optimization literature. But besides of forward problems, also an inverse problem is solved in this methodology part. It is shown that the algorithm delivers reasonable results. The algorithm is enhanced with local search and parallel computation. The proof is made that by merging local and global search, a signifcant reduction in the number of function calls can be reached. Moreover by doing hybridization more robust results are gained. At the end of the methodology part the reader has an overview about the developed genetic algorithm and the different search strategies. In the second part two problems in the field of Remote Sensing are solved using genetic algorithms. The first one is a combinatorical task, which arises in the field of an ozone retrieval algorithm. The parallelized genetic algorithm is adapted to the specific problem domain. The fitness function is formulated according to the combinatorical problem, methods are written for the specific tasks like reading HDF files and starting external processes. Then under different conditions the program is applied and the results are discussed. A second problem deals with the retrieval of cloud parameters. This task is an inverse problem and the genetic algorithm is enhanced with an local search operator. The task is about funding input parameters that correspond to given measurements. Because of this, a total least squares approach is selected for the local search. As a result we see that the hybrid approach provides more accurate results then the pure genetic algorithm.