Identifying crop planting areas using Fourier-transformed feature of time series MODIS leaf area index and sparse-representation-based classification in the North China Plain

Abstract

ABSTRACTAccurate information of crop distribution is vital for agricultural industry and food security, which gives rise to a strong demand for timely crop mapping. This study applies a method that combines sparse representation and harmonic characteristics derived from time series Moderate Resolution Imaging Spectroradiometer (MODIS) leaf area index (LAI) to identify crop planting areas to the north of the Yellow River in the North China Plain, where winter wheat and maize are widely cultivated. The upper envelope Savitzky–Golay filter was applied on the yearly time series MODIS LAI pixel by pixel to minimize the effects of anomalous values caused by atmospheric variability and cloud contamination. The Fourier transform method was then employed to extract the key parameters from the Savitzky–Golay filtered LAI. Totally about 11 parameters were extracted, including the amplitudes of the terms 0–5 and the phases of the terms 1–5 of the Savitzky–Golay filtered LAI, and were taken as the features for the crop identification. Based on the training samples of the identified crops, which were obtained through field campaigns and Google Earth images, the online dictionary learning algorithm was applied to construct the dictionary for identifying the crops. With the dictionary, orthogonal matching pursuit algorithm was applied to obtain sparse representation coefficients of the identified samples. Then crops were identified using the minimum reconstruction errors which were calculated by the dictionary and the sparse representation coefficients of each category. Winter wheat, spring maize, summer maize, cotton, and orchard planting areas from 2003 to 2016 in the study area were identified. The accuracy of the identification was evaluated by the confusion matrix. Average overall identification accuracy in the 14 years was 78.67%, with a kappa coefficient (κ) of 0.75. Annual overall accuracy from 2003 to 2016 was from 70.57% to 83.71%, and κ were from 0.66 to 0.81. These results indicate that the accuracy of crop identification is high. Compared with traditional classification methods, such as supervised classification, the approach developed in this study is feasible and efficient for identifying crop distribution in the study area. Repeated and expensive ground sampling year after year is not needed, which effectively reduces computation cost. The findings of this study can provide valuable implications for crop condition monitoring and yield estimation in the region.