基于时间序列叶面积指数稀疏表示的作物种植区域提取

以华北平原黄河以北地区为研究区域,以时间序列叶面积指数LAI(Leaf Area Index)傅里叶变换的谐波特征作为不同作物识别的数据源,利用稀疏表示的分类方法识别2007年—2016年冬小麦、春玉米、夏玉米等主要农作物种植区域。首先利用上包络线Savitzky-Golay滤波分别对2007年—2016年的时间序列MODIS LAI曲线进行重构,进而对重构的年时间序列LAI进行傅里叶变换,以0—5级谐波振幅、1—5级谐波相位作为作物识别的依据,基于各类地物的训练样本,通过在线字典学习算法构建稀疏表示方法的判别字典,对每个待测样本利用正交匹配追踪算法求解稀疏系数,从而计算对应于各类地物的重构误差,根据最小重构误差判定待测样本的作物类型,并对作物识别结果的位置精度进行验证。结果表明,2007年—2016年作物识别的总体精度为77.97%,Kappa系数为0.74,表明本文提出的方法可以用于研究区域主要作物种植区域的提取。

Extraction of planting areas of main crops based on sparse representation of time-series leaf area index

Crop mapping is an important component of agriculture monitoring. Accurate information on crop area coverage is vital for food security and the agricultural industry, and the demand for timely crop mapping is high. Previous research indicated that remote sensing technology is a practical and feasible method for agricultural crop area extraction.
In this study, the north area of the Yellow River in the North China Plain is chosen as the study area, where the main crops are winter wheat, maize, cotton, and soybean. To obtain the distribution information of crops, the yearly four-day composite MODIS time-series Leaf Area Index (LAI) with 500 m spatial resolution is collected. A total of 92 MODIS LAI images obtained yearly from 2007 to 2016 are used to build time-series LAI curves. To avoid the edge effect of the time-series LAI caused by the Savitzky-Golay filter, the last two phases of LAI images in the last year and the first two phases of LAI images in the next year are added to build the time-series LAI in a year. The Savitzky-Golay filter is then applied on the yearly time-series LAI pixel by pixel to minimize effects of anomalous values caused by atmospheric haze, cloud contamination, and so on. Fourier transform method based on reconstructed LAI is further employed to extract the key parameters. The 11 parameters, including the amplitudes of 0-5 terms and the phases of 1-5 terms, are taken as the features for crop identification. The training samples and verification samples of various crops are obtained through ground investigation and Google Earth images. On the basis of the training samples of various crops, online dictionary learning algorithm is applied to construct the dictionary used to identify the crops. With the dictionary, the orthogonal matching pursuit algorithm is further applied on samples under testing to obtain the sparse representation coefficient. Then the crops are identified according to the minimum reconstruction error, which can be calculated by the dictionary and the coefficient. Therefore, the areas planting winter wheat, spring maize, summer maize, cotton, and orchard from 2007 to 2016 are extracted in the study area. Lastly, the accuracy of the identification results is evaluated yearly by a confusion matrix.
Results show that the reconstructed time-series LAI curves are smooth and consistent with crop growth and development characteristics. Overall identification accuracy reaches 77.97% with a Kappa coefficient of 0.74 from 2007 to 2016. User accuracies for individual crops are as follows:winter wheat and summer maize, 90.60%; spring maize, 73.40%; early summer maize, 81.80%; cotton, 69.40%; and orchard, 81.60%. Annual overall accuracies from 2007 to 2016 range between 70.57% and 83.71% and Kappa coefficients range from 0.66 to 0.81.
In conclusion, combining the harmonic characteristics of the time-series LAI with the sparse representation can effectively identify the areas for planting different crops. The approach developed in this study is feasible for extracting information on main crop distribution in the study area.