Crop Classification Using Time Series Remote Sensing Data

Timely and accurate crop type distribution maps could assist crop growth monitoring.Satellite images could effectively observe the land surface and have been widely used to identify crop types.However,there are still some drawbacks(such as the limitation of ground reference data and low temporal frequency of the 30 m remote sensing data),so that it is difficult to generate 30 m crop type distribution map at large scale.This paper tried to solve these problems on three aspects:(1)to merge Landsat-5 TM data and HJ-CCD data and improve the temporal frequency at 30 m resolution;(2)to use short time series to identify crop types early;and(3)to propose new methods to identify crop types when none or limited training samples of the classification year were acquired.The experiments were conducted in Xinjiang(China)and Kansas State(USA),and we then derived the following conclusions based on these experiments:(1)Results showed that Landsat-5 TM and HJ CCD images had similar radiometric performances in multi-spectral bands.So that we acquired a 15-day image time series with 30 m spatial resolution by merging Landsat-5 TM and HJ-1 CCD data.Subsequently,optimal temporal windows for accurate crop mapping were evaluated using JBh distance.Overall classification accuracy using optimal temporal windows were 84.67% and 91.00% in Bole and Manas,respectively;and entire image time series cannot improve classification accuracy significantly.In addition,merged time series had higher temporal resolution,which were more likely to comprise the optimal temporal periods than single-sensor time series.Therefore,the use of merged time series increased the possibility of precise crop classification.(2)This paper conducted three experiments to analyze the effect of time series length on crop classification accuracy and evaluate the potential of early crop type classification.The result of Kashgar showed that April~June image time series could identify crop types.Pair-wise JM distance of the major crops were higher than 1.6,and the overall classification accuracy were 85.16%.The result of Bole showed that classification accuracies of cotton,grape,maize and watermelon saturated with 160-day,100-day,160-day and 130-day time series.The optimal time phrases for crop identification were crop greening and NDVI peak periods.Furthermore,experiments in Kansas showed that classification accuracy and certainty saturated with five-month long time series(April ~ August),and NDVI had the most contribution to crop identification.(3)This paper proposed M-Voting and P-Fusion methods to fuse the classification result of multiple classifiers,and then compared the quantity demand of training samples of M-Voting,P-Fusion and single classifiers.Result showed that when the training sample size was low(less than 500),classification accuracy of M-Voting and P-Fusin were higher than single classifiers.While,if training sample size was large(for example,larger than 4,000),single classifiers could achieve high classification accuracy,M-Voting and P-Fusion cannot improve classification accuracy.Additionally,pixel-based classification and object-based classification had similar classification accuracy,but object-based classification results had less salt-and-pepper noise,which were more similar to farmlands.(4)This paper proposed a new method to identify crop types using reference image time series when the ground reference cannot be acquired.In Bole and Manas,we used historical MODIS NDVI time series and ground reference data to generate reference NDVI time series for each crop type.Then,we used the reference NDVI time series to identify crop types at 30 m resolution.The overall accuracies were 87.13% and 83.38% in Bole and Manas,respectively.However,we had to transform the reference NDVI to Landsat/HJ NDVI before the classification,and this transformation may cause misclassification.We then improved this method and used historical samples between 2006 and 2013 to generate training samples of 2014.Among the 5,412 samples acquired from this method,the crop labels of 5,259 samples were same to the CDL labels.Then,we used these training samples to identify crop types in 2014 and the classification accuracy were higher than 90%,the crop type distribution was similar to corresponding CDL data.