Remote Sensing Classification Study Of Wetlands Based On Tolerant Rough Set And Neural Network

Wetlands, as one of the important parts of the ecosystem, have great significance in flood regulation, pollutants degradation, groundwater recharge, climate regulation, the maintenance of balancing the regional water and protecting biodiversity and other aspects. However, human activities have intensified greatly, which directly resulted in a sharp decrease of wetland in a large area, a gradual degradation of ecological functions and a serious threat to the ecological balance of the regional area.Therefore, in order to maintain the sustainable development of wetland ecosystem, it is necessary and important to conduct the investigation of wetlands in a timely, accurate and efficient way. Consider that the development of remote sensing technology provides great convenience for wetland investigation. With a background in Shuangtaizi estuarine wetlands this study uses remote sensing images as data source to explore the new methods of extracting wetlands cover information based on tolerant rough set theory and BP neural network. The main contents are given as follows:(1) Concerning the classification model based on BP neural network and tolerant rough set theory. In light of the defects of BP algorithm, it is convenient to begin with the optimization for the input of the network. Due to the merits of tolerant rough set theory in dealing with uncertain data and noise data, a new classification model for wetland remote sensing is established by the combination of tolerant rough sets with neural networks, where tolerant rough set theory is used to delete the noise hidden in the the training sample of neural network in order to purify the input data and then improve the ability of classifying the remote sensing image by BP neural network.(2) Regarding data prepossessing and the suitability analysis of Landsat-8 OLI image data in wetland remote sensing classification. The original image needs to be pretreated for use because of various factors. In this study, the procedures about radiation calibration and atmospheric correction for image data are first conducted, then according to the extracted boundary of the study area, the image fusion and mosaic and trimming process are carried out. In this way, the Shuangtaizi estuarine wetlands are divided into eight categories including water, aquaculture pond, Suaeda, reeds, paddy fields, beaches, residential and mixed vegetation. Apart from that, sample data collection is also done by choosing the training sample points according to the needs of this study and validating sample points through field acquisition. Next, the applicability of OLI images in wetland remote sensing classification is investigated. The maximum likelihood method is used to OLI image classification. Compared with the analysis of common ETM+data, the classification accuracy of OLI data is relatively high and can be used for the classification research on wetland remote sensing, providing data reference for the following neural network classification experiment.(3) Relating to the classification implementation of tolerant rough set-BP neural network. The first step is to extract the gray value of each band for the training sample, and calculate the optimal similarity thresholds based on rough entropy definition so as to get the tolerant rough set for each training sample. Meanwhile, identify the decision attribute to get a new training sample. Then establish the classification model based on the BP neural network using the Matlab software platform and extract land-cover information. To end this section, the accuracy assessment for classification results has been researched together with the comparison of the obtained results with the result of a simple BP neural network and the result of the pretreatment of attribute reduction on sample data. It is shown that the classification method of BP neural network based on tolerant rough sets can eliminate the noise in the training sample data, improve the success rate of the network training and shorten the network training time, performing better in the classification effect with an overall accuracy of 91.25% and a Kappa coefficient of 0.8969, which were increased by 7.92% and 0.0926 over the traditional BP neural network classification method, and 3.03% and 0.0357 over the pretreatment of attribute reduction. Therefore, the BP neural network classification based on tolerant rough sets is a preferable land-cover classification method for wetlands, providing a new technology for the environmental monitoring of wetland resources.