Hierarchical extraction of farmland from high-resolution remote sensing imagery

Abstract: Farmland is the material base of human survival and development. Currently, China faces the serious situation that a large population corresponds to less average arable land during a long term. As Chinese urbanization process has accelerated in recent years, farmland in particular area - suburb is often developed to construction land, and even suffers the risk of being illegally occupied. With the implementation of geography national condition monitoring plan on a national scale, China is in urgent need of the development of efficient extraction and monitoring method of farmland for the protection and rational utilization of farmland. High-resolution remote sensing image contains rich and detailed ground information, and it can accurately reflect the suburb terrain types and their spatial distribution. However, house, road, drainage, tree are mixed with farmland in the high-resolution remote sensing image of suburb, and the suburban grounds' features are very similar in the spectrum, shape and texture characteristics, which leads the extraction of farmland to become very difficult. It is more feasible to extract the farmland from the non-construction area, therefore, the construction area is separated out from the image in the first place. The best segmentation scale suitable for farmland is determined according to the multi-scale segmentation in order to accomplish the extraction of farmland in an object-based approach, and then the whole image is segment in this best scale. Furthermore, the typical samples of farmland and non-farmland are selected to train the support vector machine (SVM) model. After the farmland has been classified via SVM, the spatial distribution relationship between segmented objects is taken into consideration to remove the false alarm objects and offset the omitted objects. Specifically, the proposed method consists of four steps: construction area removing, hierarchical farmland extraction, classification via SVM, and judgment by spatial distribution relationship. As a result, according to the characteristics of the suburban farmland in high-resolution remote sensing image, an automatic farmland extraction method combining multi-scale segmentation and hierarchical recognition is presented in this paper. Firstly, an improved algorithm of Harris corner detection constrained by NDVI is developed to extract the corner, and based on the probability density map of built-up areas, the complex construction areas are separated by using Otsu algorithm (OTSU). Secondly, the estimation of scale parameter (ESP) is used to analyze the parameters for generating the multi-scale segmentation of the non-construction area, and then the optimal ones suitable for extracting farmland are selected to segment the image as a whole. Thirdly, based on three rules for identifying farmland objects, shape and spectrum are intergraded to extract the typical farmland objects from a mass of segmented objects; after the SVM model is trained based on the intact farmland samples and the non-farmland samples in construction areas, this model is used to classify the remaining uncertain ground objects. Finally, the spatial distribution relationship is taken into account to refine the classification results produced by SVM. In this study, high-resolution remote sensing image from QuickBird is used to precisely extract farmland in suburb. In the farmland extraction experiments, correct rate of the proposed method is 80.09% which is 17.88% higher than the object-oriented SVM classification method, and error rate of the proposed method is 12.26% which is 1.30% lower than the object-oriented SVM classification method. The final results indicate that the proposed method can effectively extract farmland with different structures and spectral features from high-resolution remote sensing image of the complex environment of suburban area.