A new approach for crop identification with wavelet variance and JM distance

This paper develops a new crop mapping method through combined utilization of both time and frequency information based on wavelet variance and Jeffries-Matusita (JM) distance (CIWJ for short). A two-dimensional wavelet spectrum was obtained from datasets of daily continuous vegetation indices through a continuous wavelet transform using the Mexican hat and the Morlet mother wavelets. The time-average wavelet variance (TAWV) and the scale-average wavelet variance (SAWV) were then calculated based on the wavelet spectrum of the Mexican hat and the Morlet wavelet, respectively. The class separability based on the JM distance was evaluated to discriminate the proper period or scale range applied. Finally, a procedure for criteria quantification was developed using the TAWV and SAWV as the major metrics, and the similarity between unclassified pixels and established land use/cover types was calculated. The proposed CIWJ method was applied to the middle Hexi Corridor in northwest China using 250-m 8-day composite moderate-resolution imaging spectroradiometer (MODIS) enhanced vegetation index (EVI) time series datasets in 2012. The CIWJ method was shown to be efficient in crop field mapping, with an overall accuracy of 83.6 % and kappa coefficient of 0.7009, assessed with 30 m Chinese Environmental Disaster Reduction Satellite (HJ-1)-derived data. Compared with methods utilizing information on either frequency or time, the CIWJ method demonstrates tremendous potential for efficient crop mapping and for further applications. This method could be applied to either coarse or high spatial resolution images for agricultural crop identification, as well as other more general or specific land use classifications.