Object-oriented Land Cover Classification Techiniques For Multitemporal Satellite Images

With the continuous development of satellite remote sensing technology and the increase of application requirements, this technology is applied more and more extensively. Pure research on mono-temporal satellite remote sensing can’t satisfy people’s needs. Therefore, for land cover classification of mono-temporal satellite images, not only the spatial resolution is relatively low, but also spectral information is generally relatively less. So salt and pepper phenomenon will occur and the final classification accuracy cannot meet human demands. In this paper, starting from the characteristics of multi-temporal satellite image data, it is aimed to dig multi-temporal satellite image information at depth to enhance the applications capacity of satellite image data. We firstly research the data preprocessing of satellite image to improve the classification accuracy, for example, radiometric calibration, atmospheric correction, and registration. Then the segmentation based on graph is studied. And the time series features are studied in depth. Finally, we introduce incremental learning to ensemble learning for multi-temporal satellite imagery. Compared with the conventional multi-temporal classification algorithms, incremental ensemble learning algorithm achieves better results.The main work of this paper is to study the satellite imagery data preprocessing, time series index feature extraction and object-oriented multi-temporal land cover classification of satellite images, including the following three aspects:First of all, based on the principle of acquiring multi-spectral satellite imagery data, the spatial resolution and spectral resolution of satellite image data are relatively low, and the amount of data is larger. Based on the preprocessing, we studied the cut based on graph theory. Because the deficiencies of minimum cut algorithm, we studied the normalized cut based on graph, and achieved good performance. Besides, it is also the basis of object-oriented classification.And then from the time series feature of the multi-temporal multi-spectral satellite image data, we studied the index feature, time se ries index feature that benefit land cover classification. To further reflect the different land cover has different sensitivity to different time series index features, the first order and second-order differential time series index feature are extracted. We also study object-oriented classification method. Compared with traditional pixel-based classification, the experiment result shows that the object-oriented classification results are better.Finally, based on ensemble learning, we introduce the incremental learning for multi-temporal satellite image classification. By learning the new sample, and then constantly updated classifier, then integrate a new strong classifiers that can predict test samples more accurate. The multi-temporal satellite images are used in the experiments to verify the effectiveness of multi-temporal classifier. And compared with traditional composite kernels, ensemble learning method, incremental ensemble learning algorithm performs better classification performance for multi-temporal satellite image classification.