河套灌区沈乌灌域GF-1/WFV遥感耕地提取

为提高基于遥感影像的灌区耕地自动快速提取,该文针对河套灌区沈乌灌域种植结构特点,利用实地调查结果、Google earth和GF1-WFV遥感影像构建了研究区主要作物及土地利用类型的NDVI时间序列,并利用HANTS滤波法对NDVI时间序列进行了平滑处理。分别采用基于遥感与Google earth的目视解译、监督分类(支持向量机)、基于NDVI时间序列的决策树分类与监督分类相结合的方法、基于HANTS滤波法平滑处理后的NDVI时间序列决策树分类与监督分类相结合的方法对灌区耕地进行提取。利用基于Google earth与目视解译的10 000个随机验证点以及正确率(用户精度)、完整率(生产者精度)和整体精度(提取耕地面积与实际面积的比值)3个指标对提取结果进行了评价。验证结果表明:监督分类(支持向量机)提取结果的正确率、完整率和总体精度仅为84.82%、64.4%和75.68%;基于NDVI时间序列的决策树分类与监督分类相结合的方法提取精度分别为94.28%、84.21%和89.1%;基于HANTS滤波法平滑处理后的NDVI时间序列决策树分类与监督分类相结合的方法提取精度进一步提高,3个指标分别达到94.47%、87.32%和92.24%。在作物种类繁多的大型灌区,时空分辨率优异的GF1-WFV数据在耕地面积提取上具有很强的实用性;结合作物生长规律与遥感信息的联合方法能够有效提高耕地面积的提取精度。

Cultivated land extraction based on GF-1/WFV remote sensing in Shenwu irrigation area of Hetao Irrigation District

Abstract: In order to improve the automatic extraction of cultivated land in irrigation area in remote sensing images, according to the planting structure characteristics in Shenwu irrigation area, Hetao Irrigation District, the NDVI (normalized difference vegetation index) time series of main crops in the study area were constructed based on field survey results, Google earth and GF1-WFV remote sensing images. OIF index was used to select the best band combination. Furthermore, the harmonic analysis of time series (HANTS: An improved algorithm based on Fourier transform, which can flexibly deal with the problem of unequal intervals of data that constitute the time series) method was employed to smooth the NDVI time series. Visual interpretation based on remote sensing and Google earth, supervised classification (support vector machine), and the combination method of supervised classification and decision tree classification based on NDVI time series (before and after smoothed by HANTS filtering method) were used to extract the cultivated land area of the irrigation area. The extraction errors of different methods were verified by visual interpretation and 100 000 000 random verification points whose attributes were given by the means of Google earth and visual interpretation. Moreover, 3 indices, i.e. accuracy (equivalent to the user precision in the confusion matrix), integrity rate (equivalent to the producer accuracy in the confusion matrix) and overall accuracy (ratio of extracted land area to actual area) were used to evaluate the results. The results demonstrated that the accuracy, integrity rate and overall accuracy of supervised classification (support vector machine) were only 84.82%, 64.4% and 75.68%, respectively; for the combination method of supervised classification with decision tree classification based on NDVI time series (unsmoothed), the 3 indices were 94.28%, 84.21% and 89.1%, respectively; the combination method of supervised classification with decision tree classification based on NDVI time series (smoothed) was further improved, and the 3 indices reached 94.47%, 87.32% and 92.24%, respectively. The GF1-WFV data can be used for extraction of cultivated land area, which has better spatial and temporal resolution, and has stronger ground identification ability in the irrigation area with more complex underlying surface. The NDVI time series based on the GF1-WFV data can describe the crop growth law in the study area completely, and can be used to extract the crop spatial information accurately and efficiently through the difference in the amplitude and the phase of the NDVI curve between different crops. It avoids the phenomenon of pixel-based traditional classification, for example, different objects have the same spectrum and the same objects have different spectrum, and overcomes the limitations of single image data. Compared to the results of supervised classification, the accuracy is greatly improved. After smoothing by HANTS method, the NDVI time series keep the basic shape of the original curve, and effectively eliminate the influence of outliers and noise, which more tally with the actual growth law of crops. Through the combination of supervised classification with decision tree classification based on NDVI time series (smoothed), the extraction precision of cultivated land is improved effectively. The method combining crop growth law and remote sensing information can improve the extraction accuracy of cultivated land area effectively.