Land Cover Classification Of HJ-1A/B Images For Ecosystem Carbon Budget Estimation

The quantitative esitmation of terrestrial ecosystem carbon budgets is important in ecosystem and global climate change research. It supports climate change forecast and regional carbon management, which serves for mitigating and adapting climate change. Different types of ecosystems need different carbon budget esitmation methods. Studing the distribution or area constitution of various ecosystems therefore is the most fundamental and critical task in the esitmation of terrestrial ecosystem carbon budgets.The existing research on terrestrial ecosystem carbon budget esitmation often lacks large-scale, time continuous and comparable land cover data. The traditional ground survey methods are barely able to meet the demand. Remote sensing technology is advanced in sensing large areas, fast access to information, short period, and less restrictive to ground conditions, so remote sensing image classification becomes one of the most important techniques to obtain large area land cover information. The object-oriented classification method is a widely applied one in remoge sensing image classification. It uses several adjacent pixels with more semantic information as a processing unit to achieve high-level remote sensing image classification and target objects extraction.However, terrestrial ecosystem carbon budget esitmation has some special requirements for remote sensing land cover classification. For example, it needs to develop a land cover classification system, which is compatible with the terrestrial ecosystem carbon budget and able to assess the amount and potential of carbon budgets. It then requires selecting the appropriate remote sensing data for the new classification system, and carrying out data processing, information extraction and classification accordingly. Therefore, existing classification systems, image processing methods and classification methods can not be directly used for land cover classification in carbon budget esitmation. There is need to study the key techniques involved in land cover classification in carbon budget esitmation based on its specific data sources, special application purposes and service targets. This dissertation study conducted several specific studies to meet such demand.(1) Developed a land cover classification system with the capability of assessing the amount and potential of carbon budget for carbon budget estimation. Based on the different CO2assimilation efficiency of vegetation and the classification rules of the land cover classification system, this dissertation study developed a land cover classification system for carbon budget estimation. (2) Studied the radiometric normalization methods of remote sensing data from multiple sensors or/and multiple time for carbon budget estimation, and built a standardized data set with relatively consistent spectral feature. To provide high quality data source, this dissertation study compared existing radiometric normalization methods, and selected appropriate radiometric normalization methods to process multi-temporal HJ-1A/B data and ETM+data.(3) Studied the remote sensing image transformation methods for carbon budget estimation. To highlight the remote sensing radiometric feature of various target objects and improve the image interpretation accuracy, based on the analysis of a large number of objects’ spectral characteristics in HJ images, this dissertation study formulated the LBV data transformation equations for HJ multi-spectral images, and discussed the impact of image transformation on carbon budget classification.(4) Studied the feature selection methods of object-oriented classification to optimize feature space and improve the classification efficiency. To solve the feature redundancy problem during an object-oriented classification process, this dissertation study first evaluated the objects’features based on feature correlation, inter-class distance and intra-class distance, and then retrieved the optimal feature set for carbon budget classification by comprehensive analysis of the multiple esitmation results.(5) Studied the highly automated classification methods for carbon budget estimation. To reduce human intervention in a classification process and obtain land cover information quickly and timely, this dissertation study provided an automated classification method to offer stable, and reliable data and robust technical support for a large-scale and periodic terrestrial ecosystem carbon budget esitmation.Experimental results demonstrate that the multi-sensor and multi-temporal images processed by radiometric normalization and LBV transformation have more similar radiometric characteristics, which is helpful for clustering different object. Feature space optimization and multi-layer analysis methods can significantly improve the classification efficiency and classification accuracy of carbon sequestration body, and provide stable and reliable classification products.The major contributions of this dissertation include:(1) Devloping a land cover classification system particularly for carbon budget esitmation. This dissertation study also selected HJ satellite images made in China as a data source after analyzing the feasibility of remote sensing identification for every land cover type in classification system.(2) Adjusting the radiometric normalization and image transformation methods to HJ data made in China for carbon budget esitmation. It not only provides a high-quality data source for carbon budget classification, but also lays the foundation for image processing research of the H J data made in China.(3) Proposing a new object-oriented classification algorithm based on the existing SEaTH algorithm for feature selection, to solve the feature redundancy problem during a classification process.(4) Proposing a portable automatic classification method for carbon budget esitmation, and exploring the application potential of HJ data made in China in land cover classification for carbon budget esitmation.