基于时间序列NDVI的灌溉耕地空间分布提取

灌溉耕地空间分布地图是农业和粮食政策制定、水资源管理和环境变化研究的基础数据,及时、准确和适用性强是区域尺度灌溉耕地分布提取方法的重要要求。相比于非遥感数据融合和遥感影像分类方法,利用遥感参量对统计数据进行空间重建的方法具备少受样本限制、运算速度快、与统计数据匹配水平高和产品分辨率适宜的优势,但目前该方面的研究较为薄弱。该文以中国为研究区域,应用国外提出的基于时间序列NDVI的灌溉面积统计数据空间化方法,研制中国2010年、空间分辨率250 m的灌溉耕地空间分布地图,深入分析了该方法的应用效果及其影响精度的主要原因。结果表明,利用该方法获得的灌溉耕地空间分布数据的空间位置精度与同类遥感产品相当,在数量精度上具有明显优势,不同区域的制图效果具有差异性。全国总体精度64.20%,各省精度极差为48.35%。制图误差主要来源为耕地底图、方法假设和分类参量,未来方法的优化应重点加强耕地分布制图、改进方法假设、进行不同类别和时间点的特征参量的筛选和利用。

Extracting irrigated cropland spatial distribution in China based on time-series NDVI

Abstract: Geospatial information of irrigated cropland is necessary for the formulation of food policy, water management and climate change studies. In addition to those methods based on pure image classification or non-remote sensing data, spatial reconstruction of statistics by using remote sensing features, a branch of multi-data fusion, with the advantages of less relying on the sampling points with a good consistency with the statistical data, has played an important role in land cover mapping. However, it gains less attention in regional irrigated cropland extraction, which makes it unclear about its applicability in different regions. In this paper, we firstly tested a fusion method based on NDVI data and statistical data of spatial distribution of irrigated cropland in China. Then, quantitative and spatial accuracy assessment and comparisons with other datasets were also carried out for the sake of discussing the availability of the map. Finally, the possible factors reducing the accuracy of classification were discussed. The results showed that the ratio of irrigation farming decreased and the fragmentation of irrigated croplands increased gradually from east to west. Huang-Huai-Hai and Yangtze River plant regions were the places with the most concentrated irrigation. While in the locations with low precipitation such as northeastern and northwestern areas, irrigation farming was distributed along local water resources. Those irrigation areas were all consistent with the recognized irrigation areas. Quantitatively, the relative errors of more than 90% counties were within 5%, and most of the counties with high relative error (>30%) belonged to Shanxi while the rest were shared by several other provinces. From the view of absolute error, the number of negative ones was much less than positive ones, and this rule was also appropriate on province scale. The total spatial accuracy of the new map was 64.20%, but the values ranged from 31.21% to 90.64% on province scale. Provinces with the accuracies higher than average level were mostly distributed in the eastern areas of the country, and the precision level went lower from north to south. Meanwhile, there was no apparent geographical rule in west. Referring to the comparisons with similar datasets, this fusion method of statistic and remote sensing data could not only perform better quantitatively, but also provide more spatial details than data fusion method without satellite images. In addition, it maintained a same spatial accuracy level with the image classification but accelerated the operating process. These indicated that, the output of the method was both quantitatively and qualitatively comparable to that of similar method in China, yet there was a certain distance with its first application in America. Analysis suggested that, the cropland mask, the method hypothesis and the selected features are the major factors which largely influence the mapping accuracy, so the improvement of the method relies on better cropland maps, and optimization of geographical and spectral features.