Identification Of Typical Crops In Northeast China Using MODIS And Landsat Time Series Data

The thematic information on the refined crop types is the basic data for the dynamic monitoring of crops.Rapid and accurate getting planting area and spatial distribution of crops at regional scale are essential for estimating crop yields,building agricultural models,sustainably managing natural resources,and making agricultural policy.Traditional products only provide the classification thematic information of primary land types such as farmland and forest,while crops are classified into one type.With the improvement of understanding of the phenology feature of major crops and the accumulation of satellite-based observations,there is a chance to capture the typical phenological characteristics of crops and distinguish detailed crop clusters with elevated accuracy.How to make full use of time-series remote sensing information,and improve the understanding of crop phenological characteristics at pixel scale,to improve the accuracy and reliability of crop recognition remains to be studied.In this study,we take Heilongjiang(HLJ)Province and Shenyang City,Liaoning Province,which are important grain bases in China,as the research areas.Using the time-series-based MODIS data and Landsat data,which cover the whole crop growth period,combined with the ground measured data,statistical data,and multi-source remote sensing thematic products,this study researches the automatic remote sensing identification method of crops at muti-scales.The main research contents and conclusions of this paper are as follows:(1)Analysis of remote sensing time series features of typical crops in HLJ Province.Based on ground-measured samples and high-resolution remote sensing images,a total of 3880 homogeneous MODIS samples of four types of crops(soybean,wheat,corn,and rice)were selected from typical crop areas in Heilongjiang Province from 2005 to 2018.Using MODIS surface reflectance products(MCD43A4)and phenology-related indices,7 spectral reflectance bands and 9 vegetation indices related to crop growth and development.The study was analyzed from three aspects: time series spectral features,separability of time series features among crops,and distribution of time series features.The results show that the time series spectral features of four typical crops can effectively capture the unique phenology features in the process of crop growth and development,especially the NIR-and SWIR-related indices,which are respectively linked with crop biomasses and the water content.Therefore,the features with high separation index(SI)between crops are mostly concentrated at the green-up and maturing stages of crops,and the irrigation stage of paddy.The temporal features distribution of typical crops in HLJ Province is close to the normal distribution,therefore the MODIS temporal feature data set can be used as a priori knowledge for crop identification.(2)We developed a classification model(TSAW)based on automatic weighting time-series features by Separability Index(SI)and classified four major crops(soybean,wheat,maize,and paddy)in HLJ Province,China.The phenology features of the different crops are automatically described by the time series feature vector of each crop phenology.A 2-dimension vector of feature as prior knowledge is adopted to describe the special growth rhythm of each crop,and as well serving for the sequence samples over one life-cycle of growth development.To emphasize the distinctive features,we assigned high weights to the related indices at certain time phases of the identifying crop pairs,measured the degree of matching by the rule of “fuzzy membership”and calculated the possibility of the determined class by a multi-stage weighting strategy.There are 3880 crop samples from 2005 to 2018 in HLJ Province,of which 20% of the samples are randomly selected for model construction,and the remaining 80% are used for model verification 10 times.The results show that the model can effectively identify four typical crop types.The average overall accuracy of 10-times verification is 0.97 and the kappa coefficient is0.96.(3)Application of crop TSAW model.To further verify the applicability and reliability of the TSAW model in multi-year,multi-regional,and multi-scale,we started the typical crop identification and spatial mapping under the MODIS scale in HLJ Province and Landsat scale in Shenyang in 2019 and verified the accuracy of the extracted area and spatial distribution.The results show that the TSAW model has high accuracy and reliability in the cross-year application of HLJ.The overall accuracy and kappa coefficient of four types of crop classification under the MODIS scale in HLJ Province is higher than 0.95.The model performed good on the Landsat scale without ground measured samples.Based on the MODIS time series feature reference data set and TSAW model,the relative accuracy of corn paddy identification in Shenyang is more than 95% and 80% respectively.The preliminary application of the presented approach performs well via the capture of valid phenology features of major crops within the dominant agriculture region of Heilongjiang,with the potential to serve the extraction of fine crop types over wide agriculture regions.This paper has 44 figures,17 tables,and 90 references.