基于时间序列谐波分析的东北地区耕地资源提取

耕地是人类社会赖以生存发展最重要的资源之一,及时获取其空间分布是国家农业决策的基础。论文利用2007年多时相的SPOT/VGT NDVI数据提取东北地区耕地资源信息。以NDVI时间序列数据年内变化振幅和周期差异性作为分类的依据,采用时间序列谐波分析法对全年时间谱NDVI数据进行重构,减少高频噪声对信息提取的影响,获得研究区地物信息在时间维度上的振幅、相位以及年均NDVI值影像图,然后将三者合成。应用神经网络分类方法,对合成后的影像选择训练样本,获取东北地区耕地资源的空间分布。实验中提取耕地的精度为83.26%,Kappa系数为0.732 4;该方法获取耕地资源空间分布的精度均高于GLC2000、UMD、IGBP和中科院1∶100万土地利用数据4种分类产品。研究表明,基于时间序列谐波分析法对NDVI数据重建,利用不同类型植被NDVI曲线在一年内振幅、相位特征的差异,采用神经网络分类的方法,可以精确地提取耕地资源信息,及时为农业和土地管理部门管理决策提供科学依据。

Application of Harmonic Analysis of Time Series to Extracting the Cropland Resource in Northeast China

As one of the most important agricultural resources, the cropland is the basic survival condition for human being. Accurate information on cropland area is of critical importance for assessing food security. The Northeast China includes provinces of Liaoning, Jilin, Heilongjiang and eastern part of Inner Mongolia Autonomous Region. It is one of the most important marketable production bases and output regions with rich water resources, fertile soil and vast cultivated land. With the unprecedented combination of economic and population growth, a dramatic land transformation has caused across this region, and the cropland degradation is increasingly serious. In order to preserve and manage cropland resources, it is essential to investigate and monitor cropland dynamics. Compared to traditional observations in the field, the principal advantage of remote sensing data is the possibility that they offer to gather synoptic information at regular time intervals over large areas. Especially for the muti-temporal images, repeated observations can be used to monitor characteristics of phonological dynamics at regional level. The normalized difference vegetation index (NDVI) which derived from the remote sensed data, is one of the most important parameters for the vegetation growth and was widely used in the land cover classification. In recent years, with the development of the theory about Artificial Neural Network (ANN) system, the neural network technology is becoming increasingly an effective means of classification processing of remote sensor digital images. Therefore, on the basis of the muti-period NDVI, the cropland can be identified and separated from the other land cover types by means of the neural network technology.In this paper, Harmonic Analysis of a Time Series of SPOT/VGT NDVI data was used to develop an innovative technique for cropland identification in Northeast China based on temporal variations of NDVI values during 2007. Different vegetation classes (forest, cropland, grassland, water body) exhibiting distinctive seasonal patterns of NDVI variation have strong periodic characteristics. A Discrete Fourier Filter was applied to NDVI time-series data in order to minimize the influence of high-frequency noise on class assignment. Because of the different phonology and periodicity in various land use types, the amplitude, phase and annual NDVI mean value in the studying area are acquired and integrated for an image. According to the training sample size, the neural network classification measure is introduced to extract cropland. The total accuracy is 83.26% while Kappa coefficient is 0.7324. The accuracy of measurement to extract cropland information is much higher by comparing the four current products (GLC2000 land cover data, UMD land cover data, IGBP land cover data and CAS land cover data). The study indicates that it is feasible for cropland extraction utilization of time-series analysis and neural network classification, and can provide accurate, scientific cropland information for agricultural administrators and land management decision making.