Early-season crop type mapping using 30-m reference time series
2020-06-01HAOPengyuTANGHuajunCHENZhongxinMENGQingyanKANGYupeng
HAO Peng-yu ,TANG Hua-junCHEN Zhong-xinMENG Qing-yan,KANG Yu-peng
1 Key Laboratory of Agricultural Remote Sensing,Ministry of Agriculture and Rural Affairs/Institute of Agricultural Resources and Regional Planning,Chinese Academy of Agricultural Sciences,Beijing 100081,P.R.China
2 Shenzhen Key Laboratory for Geo-Environmental Monitoring of Coastal Zone of the National Administration of Surveying,Mapping and Geo-Information & Shenzhen Key Laboratory of Spatial Smart Sensing and Services,Shenzhen University,Shenzhen 518060,P.R.China
3 Institute of Remote Sensing and Digital Earth,Chinese Academy of Sciences,Beijing 100101,P.R.China
Abstract Early-season crop type mapping could provide important information for crop growth monitoring and yield prediction,but the lack of ground-surveyed training samples is the main challenge for crop type identification.Although reference time series based method (RBM) has been proposed to identify crop types without the use of ground-surveyed training samples,the methods are not suitable for study regions with small field size because the reference time series are mainly generated using data set with low spatial resolution.As the combination of Landsat data and Sentinel-2 data could increase the temporal resolution of 30-m image time series,we improved the RBM by generating reference normalized difference vegetation index(NDVI)/enhanced vegetation index (EVI) time series at 30-m resolution (30-m RBM) using both Landsat and Sentinel-2 data,then tried to estimate the potential of the reference NDVI/EVI time series for crop identification at early season.As a test case,we tried to use the 30-m RBM to identify major crop types in Hengshui,China at early season of 2018,the results showed that when the time series of the entire growing season were used for classification,overall classification accuracies of the 30-m RBM were higher than 95%,which were similar to the accuracies acquired using the ground-surveyed training samples.In addition,cotton,spring maize and summer maize distribution could be accurately generated 8,6 and 8 weeks before their harvest using the 30-m RBM;but winter wheat can only be accurately identified around the harvest time phase.Finally,NDVI outperformed EVI for crop type classification as NDVI had better separability for distinguishing crops at the green-up time phases.Comparing with the previous RBM,advantage of 30-m RBM is that the method could use the samples of the small fields to generate reference time series and process image time series with missing value for early-season crop classification;while,samples collected from multiple years should be further used so that the reference time series could contain more crop growth conditions.
Keywords:early season,Landsat,Sentinel-2,reference time series,crop classification,Hengshui
1.Introduction
Timely and accurate crop type distribution maps could provide valuable information for crop monitoring and yield prediction,these early-season information could supply a variety of decision-making and private-sector applications,such as crop insurance and land rental decisions (Potgieteret al.2010;Torres-Sanchezet al.2014;Vaudouret al.2015;Skakunet al.2017;Mondal and Jeganathan 2018).As different crops have unique phenological parameters and image time series could describe the phenological differences among the crops,the time series data have been widely used to identify crop types (Chockalingam and Mondal 2017;Zhonget al.2019;Luoet al.2020).
There have been several studies trying to identify crop types at early season,Wardlowet al.(2007) found that the crops have relatively high separability at the green-up time phases;furthermore,Haoet al.(2015) evaluated the effect of time series length on crop identification accuracies,and the results from multiple study regions concluded that the crops could be accurately identified 4-6 weeks before the harvest (Potgieteret al.2010;Haoet al.2015,2018b;Caiet al.2018).But the remote sensing features of some crops are still confused,for example,cotton and maize have similar growing seasons (Forkuoret al.2014).
One challenge of early-season crop type mapping is the collection of ground-surveyed samples.Generally,the ground samples should be acquired annually for crop type mapping,but ground survey is a labor-and financeconsuming work so that acquiring ground-surveyed samples annually is difficult (Zhonget al.2012).There have been several studies trying to identify crop types without the use of ground-surveyed training samples and the works could be summarized as three approaches:(1) unsupervised classification,(2) rule-based methods for specific crop type identification and (3) supervised classification with training samples collected in other years.For the unsupervised method,Wanget al.(2019) generated 30-m resolution crop type maps in midwestern United States using unsupervised learning in conjunction with aggregate crop statistics.Towards the rule-based methods,Donget al.(2016)proposed a phenology-and pixel-based method to use the special phenological characters of single-season rice to identify paddy rice in Northeast Asia,the limitation of this method is that it can only be used to identify single crop type.An alternative method is to use training samples collected in other years to train the supervised classifier based on the fact that phenological characteristics of the same crop among multiple years are more similar than those of different crops.Zhonget al.(2014) proposed a phenologybased method to identify corn and soybean in Kansas using Landsat data and the classification accuracy was about 88%.Haoet al.(2016a) proposed a reference time series based method (RBM),which generated reference time series for each crop from multi-year MODIS data and ground-surveyed samples,and then used the reference time series to identify crop types.An improved version of RBM acquired training samples in the classification year by comparing the MODIS time series of the “possible training samples” with the reference time series (Haoet al.2016b).The limitation of these studies is that the reference time series are generated by MODIS data,and in the regions where the crop fields are characterized by small field size,MODIS data are mainly mixed pixels,so that it is difficult to generate the reference time series in these regions using MODIS data,there is a need of image time series at finer resolution such as 30-m or even higher spatial resolution.
Generally,image time series with high temporal resolution could better describe the difference among the crops (Haoet al.2018c) because the high temporal resolution image time series could better capture the tiny difference between crops with similar growing season (Wardlow and Egbert 2010).However,most sensors providing high-temporal resolution image time series are of coarse spatial resolution,such as NOAA-AVHRR data and MODIS,these images always contain a lot of mixed pixels in the regions where the crop fields are characterized by narrow and long,like North China Plain.The high spatial resolution data source has low temporal resolution (for example,the revisit period of Landsat data is 16 days),and the image time series contain missing value pixels due to the cloud and scan line corrector failure (SCL-off data) (Knaueret al.2017;Löwet al.2018;Haoet al.2018a).Furthermore,most classifiers cannot process the image time series with missing value which becomes a challenge for crop type mapping.
Given the above discussion,the objective of this study is to identify crop types at 30-m resolution in early season without the use of ground-surveyed training samples in the classification year.To achieve this objective,we improved the RBM by generating reference time series using 30-m image time series (30-m RBM),and used the 30-m RBM to process the incomplete image time series with missing value for early-season crop type mapping.The method was tested in Hengshui City,China,which is a representative cropland-dominated region in North China Plain,and we then evaluated how early we can generate the accurate crop type maps using the 30-m RBM in Hengshui.
2.Data source and methods
2.1.Study region
The study region is Hengshui City (115°10´-116°34´E and 37°03´-38°23´N),it is located in North China Plain(Fig.1),and the acreage of farmland is 5.71×105ha(approximately 70.3%) (Liuet al.2015).The study region has warm temperate continental monsoon climate which is characterized by hot,rainy summers and cold,dry winters.The main land cover types within this area are cropland,grassland,residential land,industrial land and water areas(Yuet al.2013).The study area is a representative region for rotate crops of winter wheat and summer maize.The major crops in the study region are winter wheat,summer maize,spring maize and cotton.Winter wheat is harvested in June,while spring maize and cotton are mature in late August and September-October,respectively.Summer maize is planted after the harvest of winter wheat and harvested in late September (Haoet al.2018b).As the shortages of water affect crop production seriously (Liuet al.2015;Xieet al.2017),early season crop type maps are urgently needed to improve crop growth monitoring and support decision making for irrigation management.
2.2.Satellite data
We used all available Landsat-7/8 and Sentinel-2 images within the study region between March and October in 2017 and 2018,and processed the data on Google Earth Engine (GEE) platform (Google 2015).The blue,red and near-infrared (NIR) bands of top of atmosphere (TOA)reflectance of Landsat 7/8 and Sentinel-2 data were used in this study.The surface reflectance (SR) data were not used in this study because the Sentinel-2 SR data of 2017 and 2018 are currently not available on GEE platform.There are radiometric differences between Landsat and Sentinel-2 data set (Xionget al.2017),and Harald and Freek(2016) found that the correlation of Landsat and Sentinel-2 TOA reflectance data was higher than their SR data.In addition,Haoet al.(2014) also found that acquiring image at optimal time phase could increase the crop separability,and the radiometric difference among different data sets had tiny effect for crop separability.Landsat-7,Landsat-8 and Sentinel-2 data are available from GEE as “LANDSAT/LE07/C01/T1_TOA”,“LANDSAT/LC08/C01/T1_TOA” and“COPERNICUS/S2” image collection (Google 2015),177 and 509 scenes of images were collected in this study for 2017 and 2018,respectively.Next,all data were re-sampled to 30-m,and we used the quality assurance (QA) band to mask the cloud pixels (ESA 2016).For Landsat-7 images,QA values equal to 66 and 68 were remained as non-cloud pixels;for Landsat-8 images,QA values equal to 322,324,386 and 388 were remained as non-cloud pixels;for Sentinel-2 images,QA value equal to 0 was remained as non-cloud pixels.Afterwards,15-day image time series were generated using the 15-day median composition for the multi-spectral bands.We used 15-day composited time series in this study because previous studies have proved that the 15-day composition is the best choice for early-season crop type mapping when cloud-free images cannot be acquired daily (Haoet al.2018c).Finally,We acquired normalized difference vegetation index (NDVI)and enhanced vegetation index (EVI) time series from the 15-day composited images,and used both NDVI and EVI time series to identify crop type in this study (Rouseet al.1974;Hueteet al.2002),as NDVI and EVI had shown high contribution to separating crop types (Wardlowet al.2007;Haoet al.2015,2018a).We compared the 15-day composited NDVI and EVI data generated from Landsat and Sentinel-2 data separately,and found that the NDVI and EVI calculated from 15-day composited Landsat and Sentinel-2 data had good correlation (Fig.2),theR2wasnearly to 0.9 and the fitted lines were near to 1:1 line.Thus,although the simply combination of Landsat and Sentinel-2 data without radiometric correction was not perfect,the 15-day composited image time series at 30-m spatial resolution could discriminate the difference among crops and generate reliable early-season crop type mapping.NDVI and EVI are calculated using eqs.(1) and (2):

Fig.1 Study region of Hengshui City,Hebei Province,China.A,location of the study region.B,spatial distribution of groundsurveyed samples.

whereρ(NIR),ρ(Red) andρ(Blue) denote the TOA reflectance of NIR,Red and Blue bands,respectively,which are Band 4,Band 3 and Band 2 of Landsat-7 data;Band 5,Band 4 and Band 3 of Landsat-8 data;Band 8,Band 4 and Band 2 of Sentinel-2 data.For eq.(2),L denotes the canopy background adjustment that addresses nonlinear differential NIR and red radiant transfer through the canopy;C1andC2denote the coefficients of the aerosol resistance term;the coefficients adopted in the EVI algorithm are:L=1,C1=6,C2=7.5,and G (gain factor)=2.5 (Hueteet al.2002).
2.3.Flowchart
An overview of the 30-m RBM methodology is shown in Fig.3.Firstly,ground-surveyed samples were collected in both 2017 and 2018 for training and validation.And then the training samples and 15-day composited NDVI and EVI time series were used to generate reference time series for each crop.The reference time series were used to process the incomplete image time series for early-season crop type mapping,and effect of time series length on crop classification accuracy were then evaluated.Finally,the accuracies of early-season crop type maps were assessed by validation samples collected in 2018 (Section 2.7),the training samples collected in 2018 were also used for early-season crop type mapping,and the training samplesderived crop type maps were compared with the reference time series-derived crop type maps to further evaluate the performance of reference NDVI and EVI time series for crop type mapping.

Fig.2 Correlation between NDVI and EVI calculated from 15-day composited Landsat and Sentinel-2 data.We collected NDVI and EVI from Landsat-8 and Sentinel-2 images at different growing stages for this correlation comparison.The dashed line in the figure is the 1:1 line.
2.4.Collecting ground reference data
Ground reference data were collected by three field surveys of Hengshui in August 2017,May and June 2018,the spatial distribution of the field-surveyed samples acquired were shown in Fig.1.The surveyed fields were selected to include the major crops across the entire study areas,and the sizes of all surveyed crop fields were larger than 40 m×40 m to ensure that these surveyed fields contain pure 30-m pixels.For the surveyed fields,the crop types and geographic coordinates were recorded.Afterwards,the central points of the surveyed crop fields were digitized as points using ArcGIS,and these points were used as training and validation samples.All samples acquired in 2017 were used to generate the reference NDVI/EVI time series,the NDVI and EVI of all samples among the entire time series were checked and the samples with missing value in the time series were removed.All samples acquired in May 2018 and a part of samples acquired in June 2018 were used as training samples,the other samples acquired in June 2018 were used as validation samples (Table 1).
2.5.Generating reference NDVI and EVI time series for each crop

Fig.3 Workflow of this study.

Table 1 Number of ground-surveyed samples acquired
In this study,we used artificial immune network (AIN)algorithms (i.e.,antibody network (ABNet)) to generate reference NDVI and EVI time series (Zhong and Zhang 2012).ABNet algorithm is inspired by the human immune system,and these algorithms have shown good potential to generate land cover maps (Imet al.2012a,b;Changet al.2016).The units of ABNet algorithms are “antigens” and“antibodies” (Fig.4-A),the training samples are used as“input antigens” and images for classification are used as“new antigens”.In the training procedure,the “antibodies”are generated from “input antigens”,and then in the classification procedure,“antibodies” are used to identify“new antigens”.The entire procedure is similar to the human immune system.
The training procedure of the ABNet algorithms is to generate the “antibodies” using the training samples (“input antigens”).Generally,the “antibodies” have three attributes:the crop type,central vector,and the recognizing radius,and the “antigens” contain two attributes:the crop type and centre vector (Fig.4-A).In this study,the central vector of each “antibody”/“antigen” was defined as the NDVI/EVI time series.For each crop type,the training procedures contained five steps (Fig.4-B):pre-selection,cloning,mutation,adaptive calculation of new “antibodies”,and antibody reorganization (Zhong and Zhang 2012).The first step (pre-selection) was to select an antigen which could best represent the unrecognized antigen population,the following four steps were conducted based on the selected antigen.The second step (cloning) was to clone a large number of the selected antigens.The third step (mutation)was to randomly mutate the cloned antigens and generate possible “antibodies”.The fourth step (adaptive calculating new artificial antibody) was to use all “possible antibodies” to recognize all “antigens”,then the “possible “antibody” which could identify the most recognized “antigens” was defined as an “antibody”,and the “antigens” which could be identified by the “antibodies” were labeled as recognized “antigens”.Finally,repeating steps 1-4 until all “antigens” of the crop type are identified by “antibodies”.Central vector of each antibody was defined as the reference NDVI/EVI time series.Similarity between NDVI/EVI time series of “antibodies” and“antigens” was calculated by Euclidean distance (ED) as described in eq.(3) (Lhermitteet al.2011):

Fig.4 Diagram of antibody network (ABNet) algorithm.A,the construction of ABNet and “antibody”.B,the workflow of generating a new “antibody” from the “antigens”.

whereatandbtdenote the NDVI/EVI of time seriesaandbat momentt,respectively,andndenotes the number of samples in the time series.The “antibodies” can be generated by the training procedure adaptively and there is no need to optimize any parameter during the training procedure.
ABNet is employed in this study to generate reference NDVI/EVI time series because ABNet can adaptively build up the network and guarantee its convergence without any parameters.Additionally,each crop may have multiple growth situations,and the ABNet could recognize different distinctive patterns which are exactly suitable for identifying the same crop in different situations with variable NDVI/EVI time series.
2.6.Early-season crop type mapping
We need to process incomplete image time series when trying to identify crop types at early season.For example,if we try to identify crop types at June 30th,we need to use the images acquired on March 1st and June 30th to identify crops,the image time series donot cover the entire growing season and there are missing value pixels in the image time series.
Firstly,we tried to generate a non-vegetation mask to remove the residential and water surface regions.The vegetation land had high vegetation fraction during the growing season,so that we need to set a threshold to exclude the non-vegetation region.Based on our previous work in the same study region (Haoet al.2019),if the maximum NDVI of a pixel during the growing season is lower than 0.4,the pixel could be identified as nonvegetation.In this study,we collected all images we could acquire and calculated the maximum NDVI of each pixel for each time series length (for example,March-June).Then,if the maximum NDVI of a pixel was lower than 0.4,the pixel was removed as non-vegetation pixels at this time phase.
Next,we improved the classification procedure of ABNet to process the incomplete time series for early-season crop type mapping.The original ABNet classification procedure calculates the Euclidean distance between the NDVI/EVI time series of a pixel and the NDVI/EVI time series of each reference time series,the pixel is labeled as the crop type of the reference time series with the minimum distance.We improved this classification procedure by adding an adaptive “missing value removing” operation before the classification (Fig.5).For example,among the 10 time phases of the entire growing season,if a pixel had missing value in time phases 3 and 5,we extracted the sub-time series excluding the two time phases with missing value from the reference time series,and then used the sub-time series for classification.
2.7.Accuracy assessment
To evaluate the performance of 30-m RBM for early-season crop classification,we used the 30-m RBM and 2018 training samples to identify crop types at early season.
Reference time series

Fig.5 Procedure of incomplete time series classification.
The classification accuracy of both 30-m RBM and training samples-derived results were evaluated using producer's accuracy (PA),user's accuracy (UA),overall accuracy (OA)and Kappa coefficient (Congalton 1991).In addition,we applied McNemar's test to evaluate the pair-wise statistical significance of the difference between the early-season crop type maps generated by reference time series and 2018 training samples (De Wit and Clevers 2004).McNemar's test is a non-parametric test based on the standardized normal test statistics,calculated as eq.(4):

wheref12denotes the number of samples that are correctly classified by classifier 1 and incorrectly classified by classifier 2;f21denotes the number of samples that are correctly classified by classifier 2 and incorrectly classified by classifier 1.We defined three cases of differences in accuracy between classifiers 1 and 2 according to McNemar's test:
(1) No significance between classifiers 1 and 2 (n):-1.96<Z<1.96
(2) Positive significance (classifier 1 has higher accuracy than classifier 2) (S+):Z>1.96
(3) Negative significance (classifier 1 has lower accuracy than classifier 2) (S-):Z<-1.96
3.Results and discussion
3.1.The reference NDVI and EVI time series of each crop
For each crop,both NDVI and EVI reference time series were generated using the data in the year 2017 (Fig.6).Winter wheat is the major winter crop in this study region,it turns green around day of year (DOY) 90-105,which is similar to the time phase when tree turns green.After the harvest of winter wheat,the fields are mainly used to plant summer maize,which is highly separable from the other crops as the growing season of summer maize are significantly later than the other crops.Cotton and spring maize are the major summer crops in the study region,and these two crops are confused as they have similar growing season.During their greening time phase,NDVI of cotton turns green faster than spring maize;the peak greenness time phases of the two crops are around DOY 210-240,and after that,NDVI of cotton decrease slower than that of spring maize as the harvest of cotton last more than one month (late-September).Therefore,the greening-up and senescence stages contribute more to separating the two crops.Comparing with the reference NDVI time series,the EVI time series at the early stage are less separable,this is mainly because EVI is less sensitive at the low to intermediate green biomass level (Wardlowet al.2007).We then used Jeffries Matusita (JM) distance to estimate the pair-wise separability of the reference NDVI and EVI time series (Bruzzoneet al.1995).Generally,crop separability increased with the time series length (Fig.7).When time series length was longer than 60 days,cotton and spring maize were separable from tree and winter wheat for both NDVI and EVI time series.However,cotton and spring maize had low separability,indicating that these two crops were more confused,which was consistent with Haoet al.(2018b).
3.2.Accuracy assessment of the early-season crop type maps
Image time series used for the early-season crop type classification were generally incomplete time series,so that we firstly estimated the potential of the 30-m RBM for the classification of incomplete time series.We used the reference NDVI time series and the EVI time series of the entire growing season of 2018 to identify crop types,and then used the validation samples with missing value to verify the classification result.Results showed that the OA of the validation samples with missing value was 97.15%,and the Kappa coefficient was 0.9501 (Table 2),which indicated that the most samples with missing value were correctly identified by the 30-m RBM.

Fig.6 Reference NDVI and EVI time series generated from the data in 2017.DOY,day of year.In this figure,the profiles have different colors and each profile denotes a reference NDVI or EVI time series,respectively.
Then,we tried to use the 30-m RBM to identify crop types at early season.The results showed that reference NDVI and EVI time series generated in 2017 could identify the winter wheat in April 15th (Fig.8),and other summer crops (cotton,spring maize and summer maize) of the study regions before June 30th (the time series length was four months) as both PA and UA of the major crops were higher than 85% and OA was higher than 90%,longer image time series could not further improve the classification accuracies,this was consist with our previous conclusion that the major crops could be correctly identified at early season using Sentinel-2 data (Haoet al.2018b).Compared with the early-season crop maps generated by the 30-m RBM,crop maps generated using ground reference data had higher classification accuracy when the time series lengths were short (Fig.9).For example,when using NDVI time series between March 1st and May30th for crop classification,PAs of cotton generated by 30-m RBM and ground-surveyed samples were 75.96 and 62.50%,respectively.When using NDVI time series longer than four months (such as between March 1st and September 30th),PAs of cotton generated by 30-m RBM and ground-surveyed samples were 91.18 and 93.27%,respectively.Finally,when the entire NDVI/EVI time series were used for crop classification,the OA of 30-m RBM result was 97.97 and 96.04%,respectively,which were similar to the accuracies of the crop maps generated using the ground-surveyed samples.

Fig.7 Pair-wise Jeffries Matusita (JM) distance of the reference NDVI (A) and EVI (B) time series with different time series lengths.The time series length starts on March 1st (DOY 61).DOY,day of year.

Table 2 Confusion matrix and classification accuracies of the validation samples with missing value
NDVI and EVI had similar trend when evaluating the effect of time series length on classification accuracies.While,PA of cotton derived from NDVI time series was higher than that from EVI time series.For instance,when the time series length was March 1-August 31,PAs of cotton derived from reference NDVI and EVI time series were 88.24 and 69.61%,respectively.The low classification accuracy of EVI time series was due mainly to the EVI time series confusion between cotton and spring maize.While,NDVI time series showed better performance to separate cotton and spring maize,this is consistent with the crop separability estimation result (Section 3.1).The higher separability of NDVI time series between cotton and spring maize was because NDVI was more sensitive to the vegetation variations at low and intermediate biomass stages,such as green-up and senescence phases (Wardlowet al.2007;Wardlow and Egbert 2008).EVI was more sensitive to the vegetation variations at high biomass stages,but both cotton and spring maize had similar high EVI at high biomass stage.We defined the time phases that both PA and UA of a crop were higher than 85% as reliable time phase,and then evaluated how early the reliable crop maps could be acquiried.Table 3 showed that ground-surveyed training samples could generate reliable winter wheat distribution maps 4-6 weeks before harvest,but 30-m RBM could generate reliable result just around the harvest time phase,maybe this was because the winter wheat NDVI/EVI time series of 2017 and 2018 were mismatched.For the major summer crops,30-m RBM and ground-surveyed samples generated reliable result at similar time phases,the accurate distribution maps of cotton,spring maize and summer maize could be generated 8,6 and 8 weeks before harvest,respectively.
The classification results using NDVI and EVI time series between March 1st and June 30th of two sub-regions showed that although there were some data missing problems in the images,the missing value did not affect the classification result,as the stripe noise in the NDVI image time series (Fig.10-A and E) did not appear in the crop classification maps (Fig.10-C,D,G and H).In addition,NDVI time series showed better potential to separate cotton and spring maize as more cotton pixels were correctly identified using NDVI time series.
3.3.Statistical analysis

Fig.8 Impact of time series length on crop classification accuracy using reference NDVI (A-C) and EVI (D-F) time series in 2018.For the time series length,number denotes month,A denotes the time series end at the 15th of each month,B denotes the 30th/31st of each month,e.g.,3A denotes time series between March 1st and March 15th.
Fig.11 showed that when the time series length was shorter than two months,crop type maps generated from groundsurveyed samples significantly outperformed 30-m RBM crop type maps.For example,when the time series length was between March 1st and April 15th,McNemar's test results between training samples with 2018 NDV/EVI time series and 30-m RBM with 2018 NDVI/EVI time series were generally higher than 10.While,when the time series length was longer than 4 months,classification results of the six scenarios did not have significant difference.For instance,when the time series length was between March 1st and September 30th,the McNemar's test results were lower than 1.96,which indicated that these results did not have significant difference.This result was consistent with the accuracy assessment,as the classification accuracies did not have significant difference when the time series length was longer than four months.

Fig.9 Impact of time series length on accuracy of crop classification using NDVI (A-C) and EVI (D-F) time series with ground reference data in 2018.For the time series length,number denotes month,A denotes the time series end at the 15th of each month,B denotes the 30th/31st of each month,e.g.,3A denotes time series between March 1st and March 15th.

Table 3 Time phase of acquired reliable classification results and crop harvest

Fig.10 Crop classification results at two sub-regions using NDVI and EVI time series images between March 1st and June 30th.A and E,NDVI color composite images of the two sub-regions.These images were composed by NDVI image of September (blue),July (green) and May (red).B and F,the surveyed fields in the two sub-regions.C and G,classification results of the first subregion generated by June 30th using NDVI and EVI time series.D and H,classification results of the 2nd sub-region generated by June 30th using NDVI and EVI time series.The classification results in this figure were generated using the NDVI/EVI time series between March 1st and June 30th.The crop legends of the figure are only for subplots B,C,D,F,G and H.
3.4.Advantages and limitations
In this study,we generated reference NDVI and EVI time series using ground-surveyed and remote sensing data in 2017,and then used the reference time series to identify crop types at early season in 2018.It is notable that the 30-m RBM does not use ground-surveyed samples collected in 2018.Comparing with the previous version which generates reference time series using MODIS data (Haoet al.2016a),the new version generates reference time series at 30-m resolution directly.The limitation of MODIS data derived reference NDVI/EVI time series are:(1) the spatial resolution of MODIS NDVI/EVI time series data is 250 m,so that the reference time series cannot be generated in study regions with small field size;(2) the reference NDVI/EVI time series have to be transferred to Landsat NDVI/EVI when trying to use the reference time series to identify crop types at 30-m resolution,which may induce misclassification.The combination of Landsat-7/8 and Sentinel-2 data improves the temporal resolution of image time series at 10-30 m resolution,which makes it possible to generate reference time series.So that the samples of the small fields could be used to generate the reference time series,and there is no need to conduct the NDVI/EVI transformation.
Another advantage of the 30-m RBM is that the method processes image time series with missing value,if the time series used for classification is irregular,the classifier could automatically generate subset reference time series based on the irregular time series (Fig.5).In addition,as for the early-season crop type mapping,the image time series used for crop mapping is generally the former part in the entire growing season,for example,the time series between March and June,which is exactly the “irregular time series only containing the March-June part.This irregular NDVI/EVI time series could be identified by the corresponding subset of the reference time series.

Fig.11 Impact of time series length on McNemar's test result.NDVI_2018,crop maps generated using NDVI and 2018 groundsurveyed samples;EVI_2018,crop maps generated using EVI and 2018 ground-surveyed samples;NDVI_ref,crops maps generated using NDVI and reference NDVI time series;EVI_ref,crops map generated using EVI and reference EVI time series.For the time series length,number denotes month,A denotes the time series end at 15th of each month,B denotes 30th/31st of each month,e.g.,3A denotes time series between March 1st and March 15th.
However,there are also some limitations for the 30-m RBM.One limitation is that we directly combine the TOA reflectance of the Landsat and Sentinel-2 data to improve the temporal resolution of the 30-m image time series.Although we used similar bands of the two datasets and the use of NDVI and EVI data had good consistency (Harald and Freek 2016;Korhonenet al.2017;Li and Roy 2017;Sotheet al.2017),there are still differences between Landsat and Sentinel-2 data (Fig.2),which may affect uncertainty for crop classification.In the future,the better availability of the harmonized Landsat and Sentinel-2 dataset,such as HLS data,may provide more reliable data for 30-m crop classification (NASA 2017).The other limitation of the 30-m RBM is the generation of reference time series.Similar to the previous version,the reference data set should contain as many crop conditions as possible to identify crop types (Haoet al.2016b),but in this study,the reference NDVI/EVI time series are generated only using the data of 2017.It is obviously that if the crop condition in 2018 is not contained in the reference dataset,the RBM cannot correctly identify crop types.Therefore,the data should be collected continuously to enhance the reference dataset.Finally,although the 30-m RBM could process the irregular time series,the misclassification still occurs when the data of optimal temporal phases are missing(Zhouet al.2013;Haoet al.2014;Schmidtet al.2014),so that the application of 30-m RBM in cloud-prone regions is still a challenge.
4.Conclusion
Generating early-season crop type distributions at 30-m resolution is challenging if training samples are not acquired in the classification year.In this study,we improved the RBM by using 30-m image time series (Landsat-7/8 and Sentinel-2 data) to generate the reference NDVI/EVI time series,and then used this reference time series to identify major crop types.The method was evaluated in Hengshui City of North China Plain,and the main conclusions are as follows:
(1) The reference NDVI/EVI time series generated using the data in 2017 have good potential to replace the training samples in 2018 for crop type identification.When using NDVI and EVI time series data of the entire growing season in 2018 to identify crop types,the overall classification accuracies are 97.97 and 96.04%,respectively,which are similar to the crop mapping results using ground-surveyed samples.
(2) The major summer crops in the study regions could be accurately identified at early season using 30-m RBM,accurate distribution maps of cotton,spring maize and summer maize could be generated 8,6 and 8 weeks before their harvest,respectively;but winter wheat can only be accurately identified around the harvest time phase.In addition,NDVI has higher accuracy than the EVI time series for crop type classification as NDVI has better separability for distinguishing cotton and spring maize at the green-up time phase.
Future work should focus on collecting training samples in multiple years,and multi-sensor harmonized image time series should be further used to improve the temporal resolution of the 30-m resolution image time series.
Acknowledgements
The study was supported by the China National Key S&T Project of High Resolution Earth Observation System (30-Y20A07-9003-17/18) and the National Natural Science Foundation of China (41801359).
杂志排行
Journal of Integrative Agriculture的其它文章
- First report of the South American tomato leafminer,Tuta absoluta(Meyrick),in China
- Mapping the fallowed area of paddy fields on Sanjiang Plain of Northeast China to assist water security assessments
- Variation in spectral characteristics of dissolved organic matter derived from rape straw of plants grown in Se-amended soil
- The circulation of unique reassortment strains of infectious bursal disease virus in Pakistan
- A protective role of resveratrol against the effects of immobilization stress in corpora lutea of mice in early pregnancy
- Effect of dietary supplementation of pyrroloquinoline quinone disodium on growth performance,meat quality and antioxidative ability of broilers
