Research On Forest Type Recognition For Hyperspectral Remote Sensing Based On Cat Swarm Optimization Algorithm

Protection and rational utilization of forest resources have an important significance for the stability of the earth’s ecological system.Correct classification of forest tree species group is the important foundation and basis. The rise and development of hyperspectral remote sensing laid a solid foundation for obtaining forest resources information. Domestic and foreign experts use hyperspectral remote sensing data and different classification methods for identifying forest types, in order to find a more efficient and rational manner and get more accurate results.In this study, hyperspectral remote sensing data of the domestic environment and disaster monitoring small satellite and spectral information are used for fine recognition of forest types based on cat swarm intelligence optimization algorithm in the study area of Jilin Wangqing. For this purpose, the dissertation systematically explores the cat swarm optimization algorithm and its improved algorithm for recognition of forest type, and establishes one kind of high spectral dimensionality reduction model by searching and selecting bands, and establishes four kinds of forest classification model based on cat swarm algorithm.The dissertation does some systematic study on hyperspectral dimensionality reduction, and on clustering for identifying forest types, and on mining spectral information rules for identifying tree species(group) based on cat swarm intelligence algorithm.First, the environment and disaster monitoring small satellites (HJ-1A) can obtain the data at a low cost, but it is not a long time to provide data in China. So the data will be rarely applied to identify the tree species. Mining the utilization potentiality in forest area and the usability of forest fine identification on H.T-1A Hyperspectral data, has the profound practical significance. Secondly, because the impact of remote sensing image selection, preprocessing methods and different classification methods, hyperspectral remote sensing image classification still has many challenges and a greater room for improvement. For example, in the study, accuracy obtained by existing classification method is restricted because of the complex terrain or higher spatial heterogeneity. Different classification methods have advantages and disadvantages for different circumstances, the classification results will be very different. Integrating and improving the existing algorithms, seeking the new algorithm become the focus of research for hyperspectral image classification. The thesis work around these problems:Firstly, HJ-1A remote sensing images of the study area are preprocessed that can form the basis data for classification; then by selecting emerging and relatively good cat swarm intelligence algorithm, based on the study of the mechanism of the algorithm, and based on full use of information technology, the thesis proposes discrete binary cat swarm optimization (BCSO), establishes band search model for automatically searching out good separability band combinations and laying a foundation for better follow-up study of Mongolian oak, birch, larch, poplar, spruce and other species classification; then the dissertation puts the clustering model based on CSO algorithm and semi-supervised fuzzy C cat swarm optimization algorithm (CSO-SMFC), coniferous forest, broadleaf and mixed forests and other forest types of the study area to be classified, and compares the results; and finally, in the above study, the basic classification rule mining algorithm based on CSO(CSO-Miner) and improved cat swarm optimization and support vector machines (ACSO-SVM), which the two ways are proposed, and then builds a classifier to form a high spectral species classification model and uses the model to recognize Mongolia oak, birch, larch, poplar, spruce and other dominant species group in the study area, and the thesis has compared, validated and evaluated the classification results. This thesis aims to explore the forest classification methods, to provide reference and basis for subsequent application to other forestry research.The results show that:(1) during the band selection, when the subspace is 3, BCSO extracted band combination is 21-43-109; (2) in the coniferous forest, broadleaf forest and mixed forest types recognition, based on the basic cat swarm clustering algorithm, the optimal accuracy is 83.5%, and the optimal precision of clustering model based on CSO-SMFC is 85%; (3) during fine recognition of the tree species group, the best overall classification accuracy based on CSO-Miner model is 80.83%, Kappa coefficient is 0.77; ACSO-SVM overall classification accuracy is 84.16%, Kappa coefficient is 0.81. Combining the study area information, applicating classification method based on CSO, using the environment and disaster monitoring small satellite hyperspectral remote sensing images, Mongolian oak, birch, larch, poplar, spruce are better recognized in Wangqing Forest Management Area.This dissertation has carried on the exploratory study in the field of hyperspectral remote sensing classification, and combines the computer intelligent algorithm with remote sensing technology, and provides a new way of thinking for the forest type identification. The results of the research can provide the theoretical guidance and technical support for basic logging enterprises for forest investigation and the national forestry planning, and can lay the foundation for further studies in many aspects of forestry fields. It has a certain role to our country forestry development.