Object-Based Farmland Recognition And Extraction From High Resolution Remotely Sensed Imagery

The spatial resolution of remotely sensed imagery has increased rapidly and entered the sub-meter times due to the development of satellite sensors. While providing spatial and structural information in details, the high-resolution remotely sensed imagery brings much more features and complicated textures, which increase the complexity of image processing. Pixel-based approaches cannot meet the high-resolution remote sensing image processing and applications. Image analysis is transferring the strategy of pixel-category to the object-extraction mode.Taking Yuhang District of Hangzhou City as the research area, and QuickBird high-resolution Remotely Sensed Imagery as the experimental data, this dissertation investigated the object-oriented model of farmland recognition and extraction in accordance with the idea of "spectrum recognition—image segmentation—object extraction". Firstly, the classification decision tree was constructed based on the vegetation index calculated through multi-spectral data of QuickBird imagery to obtain the vegetation information, and then the morphological opening/closing operation was used to eliminate small fragments of the vegetation information. Secondly, a marker-based watershed algorithm segmented the panchromatic band data of QuickBird imagery to obtain the object image. In the end, the binary image of vegetation and the object image were superimposed together to extract farmland object according to identification of vegetation pixels as well as the area feature of the farmland. The main contents and results of the research were presented as follows.(1) Analysis and identification of spectral information. The values of spectral response in all bands of multi-spectral QuickBird imagery was counted and analyzed. According to the difference with other surface features, a vegetation index image could be generated by computation and combination among multiple bands. The threshold was determined by vegetation characteristics. Then the classification decision tree was constructed to obtain the binary image of vegetation. The experiments showed that the threshold (NDVI=0.095) calculated by the method of minimum error could obtain good vegetation extraction results. In order to facilitate follow-up treatment, the morphological opening/closing operation was used to eliminate small fragments in the binary image.(2) Image segmentation and object generation. Following the process "gradient computation—low-pass filter—Watershed segmentation—Object generation" the result image with the object as the unit could be obtained. Butterworth low-pass filter was constructed to smooth the gradient image generated via computing the gradient of panchromatic image in order to remove part of the influence of local minimum on the following segmentation. Afterwards, the marker-based watershed algorithm was applied to the image, combined with the method of minima imposition which could effectively control the over-segmentation. Finally, the object image was formed by marking each region. This improved watershed algorithm kept good boundary information and leaded to closed and connected regions.(3) Farmland extraction. The method of superimposing the vegetation extraction binary image with the object image to extract farmland objects was put forward. The main idea included two steps. Step one was to superimpose these two images together and accomplished the first extraction by judging whether every object intersected the vegetation pixels. Step two was to complete an advanced extraction using thresholds of the extent of vegetation cover as well as the size of the object itself, according to the area feature of farmland. After the above operations, non-vegetation surface features such as residential areas and water as well as some non-farmland vegetations such as sparse plant canopies, could be removed effectively. Meanwhile, the edge profile information remained basically completely as long as the farmland was extracted from the background. At last, a pixel-based and an object-based confusion matrix were generated separately to evaluate the segmentation results from different aspects. For the pixel-based confusion matrix, the evaluation result showed that the overall classification accuracy was 87.50%, with a Kappa coefficient of 0.7496. And for the object-based one, the overall classification accuracy was up to 86.98%, with a Kappa coefficient of 0.7397. Therefore the approach presented in the paper could meet the demand of feature recognition and extraction of high-resolution remotely sensed imagery.Currently there are still no standard approaches and reference criteria on object-oriented extraction of high-resolution remotely sensed imagery. For further research, there are two main directions worthy of efforts. The first is to search and develop more appropriate image segmentation approaches in order to improve the accuracy of image following processing. The second is to consider selectively adding other features of farmland such as texture and shape in the farmland object extracting process in order to lower the misclassification rate.