基于FastICA算法和MODIS数据的水稻面积提取

以苏、皖、赣三省为研究区域,采用FastICA算法从MODIS数据中提取2010年水稻种植面积,并验证该算法在混合像元分解中的有效性。在对2010年46景8 d合成地表反射率产品数据进行预处理的基础上,结合MODIS土地利用产品和平滑滤波算法,构建耕地类型像元的I_LSW和I_NDV时相变化曲线。依据I_LSW和I_NDV曲线在水稻移栽期前后的变化规律,并根据由各地区水稻I_NDV时相曲线计算得到水稻相似性指数,从MODIS影像中提取水稻像元。采用FastICA算法对潜在水稻像元水稻生长期内的I_NDV时相曲线进行分解,计算每个像元的水稻丰度,绘制水稻丰度图,获取研究区各省水稻分布和种植面积。利用统计年鉴数据和样方资料对FastICA算法提取的水稻面积进行了验证。结果显示:采用水稻相似性曲线有利于提高稻田识别效率,所获取的水稻分布与实际情况吻合;FastICA算法能够分解不同地区水稻I_NDV时相曲线;与统计资料比较,江苏、安徽、江西三省水稻面积的提取精度分别为86.4%、87.9%、51.5%。江西水稻面积提取误差主要出现在地形起伏较大的山区。

Extracting the paddy rice area from MODIS imagery by FastICA algorithm

This paper is aimed to extract the paddy rice area of Jiangsu,Anhui and Jiangxi provinces in 2010 by using multi-temporal MODIS imagery,and to investigate the efficiency of Fast Independent Component Analysis(FastICA) algorithm in dealing with mixed pixels.In order to map the paddy rice area,the time series of Land Surface Water Content Index(I_LSW) and Normalized Difference Vegetation Index(I_NDV) are first produced from MODIS data.Then they are smoothed in order to reduce contamination by clouds.Meanwhile,the pixels with non-crop classes are removed from the index-images using the Land Cover Type data,which is one of the public-distributed MODIS products.The mapping accuracy of the paddy rice area is always affected by the spatial variation of rice planting schedules and rice species.To solve this issue,Rice Similarity Index(RSI) is calculated from I_NDV and used to establish standard rice growing curves for different parts of the study area according to the field survey.Combining the standard rice growing curves and the time series of index-images,a rice mapping algorithm is developed and performed to extract the rice planting area.The results show that the obtained rice distribution map is well consistent with the actual situation.However,the retrieved rice planting area is overestimated in most part of the study area,which can be primarily ascribed to the mixed pixels.In this paper,the FastICA algorithm is adopted to decompose the mixed pixels using time series of I_NDV.Before applying the FastICA algorithm,the images of times series of I_NDV are masked by the obtained paddy rice map.Menwhile,several parameters of the FastICA are optimized by comparisons.As a result,a map of paddy rice abundance fraction is retrieved for each province by FastICA.According to the analysis and comparison with the statistic data,the average accuracies of the obtained rice area are 86.4%,87.9% and 51.5% in Jiangsu,Anhui and Jiangxi provinces,respectively.The larger area error in Jiangxi province can be ascribed to the great proportion of the paddy fields located in the mountainous area.