A phenology-based vegetation index for improving ratoon rice mapping using harmonized Landsat and Sentinel-2 data

Yunping Chen ,Jie Hu ,Zhiwen Cai ,Jingya Yang ,Wei Zhou ,Qiong Hu ,Cong Wang,Liangzhi You,Baodong Xu#

1 Macro Agriculture Research Institute,College of Plant Science and Technology,Huazhong Agricultural University,Wuhan 430070,China

2 College of Resources and Environment,Huazhong Agricultural University,Wuhan 430070,China

3 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,China

4 Key Laboratory for Geographical Process Analysis & Simulation of Hubei Province/College of Urban and Environmental Sciences,Central China Normal University,Wuhan 430079,China

5 International Food Policy Research Institute,NW,Washington,D.C.20005,USA

Abstract Ratoon rice,which refers to a second harvest of rice obtained from the regenerated tillers originating from the stubble of the first harvested crop,plays an important role in both food security and agroecology while requiring minimal agricultural inputs.However,accurately identifying ratoon rice crops is challenging due to the similarity of its spectral features with other rice cropping systems (e.g.,double rice).Moreover,images with a high spatiotemporal resolution are essential since ratoon rice is generally cultivated in fragmented croplands within regions that frequently exhibit cloudy and rainy weather.In this study,taking Qichun County in Hubei Province,China as an example,we developed a new phenology-based ratoon rice vegetation index (PRVI) for the purpose of ratoon rice mapping at a 30 m spatial resolution using a robust time series generated from Harmonized Landsat and Sentinel-2 (HLS) images.The PRVI that incorporated the red,near-infrared,and shortwave infrared 1 bands was developed based on the analysis of spectro-phenological separability and feature selection.Based on actual field samples,the performance of the PRVI for ratoon rice mapping was carefully evaluated by comparing it to several vegetation indices,including normalized difference vegetation index (NDVI),enhanced vegetation index (EVI) and land surface water index (LSWI).The results suggested that the PRVI could sufficiently capture the specific characteristics of ratoon rice,leading to a favorable separability between ratoon rice and other land cover types.Furthermore,the PRVI showed the best performance for identifying ratoon rice in the phenological phases characterized by grain filling and harvesting totillering of the ratoon crop (GHS-TS2),indicating that only several images are required to obtain an accurate ratoon rice map.Finally,the PRVI performed better than NDVI,EVI,LSWI and their combination at the GHS-TS2 stages,with producer’s accuracy and user’s accuracy of 92.22 and 89.30%,respectively.These results demonstrate that the proposed PRVI based on HLS data can effectively identify ratoon rice in fragmented croplands at crucial phenological stages,which is promising for identifying the earliest timing of ratoon rice planting and can provide a fundamental dataset for crop management activities.

Keywords: ratoon rice,phenology-based ratoon rice vegetation index (PRVI),phenological phase,feature selection,Harmonized Landsat Sentinel-2 data

1.lntroduction

Paddy rice provides a staple food for more than half of the world’s population and occupies more than 12% of the global cropland area (Wanget al.2015;Denget al.2019;Singhaet al.2019).The increasing global food demands driven by population growth and dietary changes have drawn the attention of the scientific community.For instance,global rice production needs to increase by 26% by 2035 to meet the requirement of rapid population growth (Yamanoet al.2016;Donget al.2017).Nevertheless,the growth rate of rice yields has slowed significantly in recent years due to the shortage of rural labor,the changing structure of the workforce,and the increasing cost of agriculture (Liuet al.2012;Denget al.2019).Thus,increasing rice production is pivotal for ensuring food security worldwide.

Previous studies suggested three main ways to increase rice production: expanding the cropland area,increasing rice production per hectare and improving the multiple cropping index (Ray and Foley 2013).Due to rapid urbanization and industrialization,as well as the ceiling effect of rice breeding,the former two crop management strategies are difficult to implement (Peng 2014;Denget al.2019;Wanget al.2021).Therefore,increasing the multiple cropping index of arable land,such as through double rice cropping,has become a key initiative for ensuing high and stable rice production.However,taking China as an example,the area of the double rice cropping system has decreased by approximately 40%,primarily due to increasing labor costs (Donget al.2017;Denget al.2019).Moreover,the double rice cropping system exhibited the highest carbon footprint intensity among rice production systems due to its requirement for more agricultural inputs (Linet al.2021;Xuet al.2022).Thus,there is an urgent need to find low-cost and -labor rice cropping systems that are environmentally sustainable to effectively improve the multiple cropping index.

Ratoon rice is the practice of obtaining a second harvest from tillers originating from the stubble of the first harvested crop (the main crop).This process allows one sowing to produce two harvests,which can improve the multiple cropping index (Jones 1993;Peng 2014).The advantage of ratoon rice is that it can improve rice production and enhance resource efficiency with minimal agricultural inputs (Huanget al.2022;Yonget al.2022).Specifically,ratoon rice achieved 72-129% higher annual grain production,net energy production,and net economic return than single rice crops,and yet it reduced energy inputs,production costs and global warming potentials by 32-42% compared to double rice crops (Yuanet al.2019).Additionally,ratoon rice is an effective rice cropping system with largely reduced vulnerability to extreme weather,such as the high temperature in summer in the middle and lower reaches of the Yangtze River or the prolonged rainy period in southern China (Wanget al.2021).The specific characteristics of production improvement,ecological protection and disaster resistance provided by ratoon rice make it an effective response to global food security and climate change issues.To determine the current planting area of ratoon rice as well as its expansion potential,accurately obtaining the spatial distribution of ratoon rice crops is critically important.

Satellite remote sensing is an indispensable earth observation technique that has proven to be an effective and valuable tool in rice mapping due to its wide coverage,real-time capabilities,and low cost (Qiuet al.2015;Donget al.2016;Huanget al.2018;Paulet al.2020).The methods for rice mapping by remote sensing images can generally be divided into two categories: classifiers and thresholds.Classifier-based methods employ mono-or multi-temporal remote sensing data of rice and other crop samples to train the classification model (e.g.,maximum likelihood,support vector machine,dynamic time warping,long short-term memory,deep convolutional neural networks,and others) (Oguroet al.2001;Liet al.2014;Castro Filhoet al.2020;Weiet al.2021).Then,the trained model is used to identify rice and other land cover types from satellite images.Nevertheless,this method is limited in this application by low computational efficiency and redundant input data (Donget al.2016).In contrast,threshold-based methods,which develop a specific vegetation index with thresholds set at key phenological periods of rice growth,can effectively address these limitations to better identify rice.For instance,the criterion of the land surface water index (LSWI)+0.05 greater than the normalized difference vegetation index (NDVI) or enhanced vegetation index (EVI) was proposed to identify the flooded/transplanting period of rice (Xiaoet al.2005).Zhanet al.(2021) used the shape feature of the Sentinel-1A vertical-horizontal (VH) backscatter time series in the flooding phase to set the critical threshold for rice identification.However,current studies mainly focus on distinguishing rice from other non-rice crops rather than different rice cropping patterns.The difficulty in extracting ratoon rice is that it has similar phenological characteristics to other rice cropping systems (e.g.,double rice),as denoted in the time series of commonly used vegetation indices (Liuet al.2020).Therefore,these limitations highlight the necessity of developing a new vegetation index based on specific growth stages for capturing the distinguishing characteristics of ratoon rice.

Furthermore,due to the high heat demand for the growth of ratoon rice,it is generally cultivated in lowlatitude regions characterized by cloudy weather and fragmented croplands.Thus,images with high spatiotemporal resolutions are essential for mapping ratoon rice with sufficient accuracy.In general,the short revisit cycle of satellite sensors has coarse spatial resolutions (e.g.,Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS)),while the high spatial resolution sensors are accompanied by low temporal resolution (e.g.,Landsat) (Griffithset al.2019;Nguyenet al.2019;Shaoet al.2019).Considering the planting characteristics of ratoon rice,the integration of multisource high spatial resolution images is an effective way to obtain high resolution spatiotemporal observations and consequently improve the accuracy of ratoon rice mapping (Yinet al.2019).In particular,the Harmonized Landsat Sentinel-2 (HLS) products,which combine Landsat Operational Land Imager (OLI) and Sentinel-2 Multispectral Imager (MSI) sensors,provide surface reflectance data with spatial and temporal resolutions of 30 m and less than 4 days,respectively (Claverieet al.2018).Therefore,the high spatiotemporal HLS data have great potential for assessing the spatial distribution of ratoon rice at a large scale.

In this study,we developed a phenology-based ratoon rice vegetation index (PRVI) to capture the individual features of ratoon rice at specific phenological stages through the detailed analysis of its spectral differences with other land cover types,such as single rice and double rice cropping systems,throughout the whole growth period.Taking Qichun County in Hubei Province,China as an example,the spatial distribution of ratoon rice was identified based on the proposed PRVI derived from HLS products.In addition,the performance of the PRVI was compared to three widely used vegetation indices (i.e.,NDVI,EVI and LSWI) to better understand the potential of PRVI in terms of ratoon rice mapping.

2.Materials and methods

2.1.Study area

The study area was Qichun County in eastern Hubei Province in the northern part of the middle reaches of the Yangtze River in China (29°59′-30°45′N,115°12′-15°56′E).The terrain of Qichun is complex,with mountainous forest in the north and plain cropland in the south,and the elevation ranges from 0 to 1,200 m.This region falls within the subtropical continental monsoon climate and experiences an average of 249 frost-free days per year,an annual precipitation of 1,341 mm,2,025 h of sunshine each year,and annual temperature of 16.8°C.The area covered by this county is approximately 2,400 km2,28% of which is cropland characterized by fragmented landscapes.Although paddy rice is the main crop type in this region,the cropping patterns of paddy rice vary,i.e.,single rice,double rice,and ratoon rice,which is what makes this region suitable for exploring the specific characteristics of ratoon rice.

The crop calendars for the three rice cropping systems in Qichun County,according to field surveys and farmerreported data,are shown in Fig.1-A.Ratoon rice undergoes two rice cycles between late March and the end of October that consists of a single sowing and two harvests.Note that the overlapping growth periods between ratoon rice and double rice may introduce more uncertainties when trying to distinguish them,which highlights the importance of developing a specific method for identifying ratoon rice.Furthermore,the whole growth period of ratoon rice can be divided into seven phenological stages (Faruqet al.2014;Donget al.2017),i.e.,seeding (SS),tillering (TS),booting (BS),grain filling and harvesting (GHS),tillering of the ratoon crop (TS2),booting of the ratoon crop (BS2),and grain filling and harvesting of the ratoon crop (GHS2),as shown in Fig.1-B.Ratoon rice exhibits different characteristics at these seven phenological stages due to biophysical and biochemical changes (e.g.,leaf pigments,leaf water content,and canopy structure).These typical phenological features can be sufficiently captured by the spectral reflectance,which is critical for ratoon rice identification.

Fig.1 The main crop calendars in the study area and photos of ratoon rice at the seven phenological stages.A,crop calendars for ratoon rice,single rice,and double rice cropping systems in Qichun County,Hubei Province,China.B,field photos of ratoon rice at each phenological stage.

2.2.Data

HLSThe HLS data developed by NASA (https://hls.gsfc.nasa.gov/) provide surface reflectance products derived from Landsat-8 OLI and the Sentinel-2 MSI images (Claverieet al.2018).Specifically,the HLS data include four types: S10 (Sentinel-2 MSI surface reflectance data at spatial resolutions of 10,20 or 60 m),S30 (Sentinel-2 MSI harmonized surface reflectance data resampled at a 30 m spatial resolution),L30 (Landsat-8 OLI harmonized surface reflectance data resampled at a 30 m spatial resolution) and M30 (5-day Landsat-8 OLI or Sentinel-2 MSI harmonized surface reflectance resampled at a 30 m spatial resolution).Based on data accessibility,S30 and L30 were selected to maximize the numbers of observations from the different satellites.Both S30 and L30 were processed for atmospheric correction,cloud mask,view and illumination angle (BRDF) adjustment,and bandpass adjustment to ensure consistent surface reflectance between both sets of images.The HLS product includes a quality assurance (QA) flag denoting the state of cloud shadows,clouds,adjacent clouds,and cirrus clouds for each pixel.In this study,the images with cloud coverage less than 15% were selected.Moreover,the images that covered the study area were carefully assessed through visual inspection to select only highquality observations.We obtained a total of 19 highquality HLS images,consisting of five L30 images and 14 S30 images acquired from March to October 2019.The number of high-quality HLS images available for each phenological stage of ratoon rice ranged from 1-4,with more than three images available for both SS and GHS and only one image available for TS2 (Fig.2).In this study,six spectral bands from both the HLS L30 and S30 data,i.e.,Blue,Green,Red,NIR,SWIR 1 and SWIR 2,were employed to develop the vegetation index for identifying ratoon rice (Table 1).

Table 1 The band names and wavelengths of the Harmonized Landsat and Sentinel-2 (HLS) L30 and S30 data1)

Fig.2 The observation dates of the Harmonized Landsat and Sentinel-2 (HLS) L30 and S30 products at the different ratoon rice phenological stages over the study area in 2019.

Field samplesTo collect the training and validation samples of ratoon rice and other land cover types,we conducted field surveys on two dates in 2019: May 24 (booting stage of the main season) and September 5(booting stage of the ratoon season).These two dates were selected since the biomass of ratoon rice accumulated rapidly in the field at these times,which can be easily identified among the various rice cropping systems.For each sample,we took field photos and recorded the associated geographical locations by GPS.To avoid the spatial errors of these ground samples,Google Earth imagery was used to check whether each sample was located within the field rather than on the road.Finally,a total of 1,476 crop samples were collected,including 258 ratoon rice,99 double rice,168 single rice,356 other crops (e.g.,cotton,corn,soybeans),and 595 non-cropland samples (e.g.,forest,water bodies,buildings).In this study,the field samples of each land cover type were divided into portions of 70 and 30% for developing the ratoon rice vegetation index and validating the performance of the vegetation index,respectively.

2.3.Methodology

Development of a PRVlThe framework for developing the PRVI based on HLS data is shown in Fig.3.It consists of three major steps: (1) analysis of the spectrophenological separability between ratoon rice and other land cover types,(2) selection of the optimal spectro-phenological features for ratoon rice,and (3) establishment of the PRVI formula.A detailed description of each step can be found as follows:

Fig.3 Developmental framework of the phenology-based ratoon rice vegetation index (PRVI).

(1) Analysis of spectro-phenological separability between ratoon rice and other crop types.Mapping ratoon rice is impacted not only by non-rice crops but also by the similar features of other rice cropping systems,especially double rice (Fig.1).The separability index (SI) proposed by Somers and Asner (2013),which can separate different land cover types by assessing their intraclass and interclass variability,was adopted in this study to select the key phenological phases and spectral characteristics of ratoon rice.TheSIcan be calculated according to eq.(1):

m={Blue,Green,Red,NIR,SWIR1and SWIR2},t={70,78,105,142,...,270,290,302}

where c andjdenote ratoon rice and other classes (i.e.,double rice,single rice,other crops and non-cropland),respectively;bis the total number ofjclasses (four in this case),mrefers to the six spectral bands in the HLS data,andtrefers to the 19 observation dates from day of year (DOY) 70-302;are the mean values of spectral featuremon DOYnfor ratoon rice and other land cover types,whileσcandσjare the standard deviations of the specific feature for ratoon rice and other land cover types,respectively;anddenote the interclass heterogeneity and intraclass heterogeneity,respectively.A largeSIcjresults from high interclass heterogeneity and low intraclass heterogeneity,indicating the higher separability between ratoon rice and the other land cover types.To illustrate the overall separability of a feature to identify ratoon rice,we calculated a global separability indexSIglobalby averaging allSIcjvalues.SIglobalis a global indicator that denotes the ability of each spectro-temporal feature to distinguish the target class from other classes (Huet al.2019).

p={SS,TS,BS,GHS,TS2,BS2,GHS2}

whereiornrepresents the start or end date of each phenological stage,andpis the specific phenological stage of ratoon rice (Fig.1).Similarly,a highvalue indicates that ratoon rice can be easily identified at phenological stagep.

The average global separability indexof ratoon rice,where the horizontal and vertical axes represent the different phenological phases (seven stages in total) and spectral features (six bands),respectively,is shown in Fig.4.Yellow grid cells represent the greater separability of the corresponding featureF(m,p),indicating the stronger ability of this feature to differentiate ratoon rice from other land cover types in the training dataset.Specifically,the four spectral bands of Red,NIR,SWIR 1 and SWIR 2 were the most sensitive to ratoon rice.The phenological phase of TS2 was the most important due to its highest separability over the spectral span.Although these spectro-phenological features with highmay be the optimal feature set from the feature separability perspective,the feature correlation needs to be further evaluated.

Fig.4 The average global separability index of ratoon rice.SS,seedling;TS,tillering;BS,booting;GHS,grain filling and harvested;TS2,tillering of ratoon crop;BS2,booting of ratoon crop;GHS2,grain filling and harvested of ratoon crop.

(2) Selection of the optimal spectro-phenological features.To assess the correlations of the spectrophenological features,the phenology-based spectral and temporal feature selection (PSTFS) method proposed by Huet al.(2019) was adopted in this study.The PSTFS method automatically determines the final features by balancing the spectral separability with information redundancy as measured by the correlation across different features (Huet al.2019).The coefficient of determination (R2) was used to measure the correlation with the average global separability indexof ratoon rice,as shown in eq.(3):

m1,m2={Blue,Red,Green,NIR,SWIR1and SWIR2},p1,p2={SS,TS,BS,GHS,TS2,BS2,GHS2}

whereF(m1,p1) andF(m2,p2) represent one of the 42 time seriesfeatures of ratoon rice,Var(F(m1,p1)) andVar(F(m2,p2)) are the corresponding variances,andCov(F(m1,p1),F(m2,p2)) is the covariance betweenF(m1,p1) andF(m2,p2).The feature with the maximum value of (F(mmax,pmax)) inwas selected,andR2(eq.(3)) was calculated betweenF(mmax,pmax) and every other featureF(mi,pi).As shown in Fig.3,the threshold of (1-0.02k) decreases with increasing iterations (k) and influences the composition and dimensions of the optimal feature set.Along withR2and the threshold was adopted to evaluate the correlation between two compared features based on a previous study (Huet al.2019).This criterion was critical to remove the similar feature for each iteration,and then to derive the optimal set of spectro-phenological features for identifying ratoon rice.Specifically,if theR2value between the two compared features was less than the correlation threshold (1-0.02k),F(mi,pi) was retained from thematrix,and otherwise it was removed.Then,this process was repeated fork+1 iterations until all features inwere checked,and the optimal feature setF(Fig.3-(2)) was obtained.To determine the separability between each optimal feature,we calculated the average value ofbetween the horizontal and vertical axes in the optimal feature set.

The crucial bands and phenological phases of ratoon rice with both high separability and low information redundancy are shown in Fig.5.The optimal spectrophenological features,i.e.,F(Red,TS2),F(NIR,BS),F(NIR,GHS),F(NIR,TS2),andF(SWIR 1,GHS2),were found to be the most sensitive features for mapping ratoon rice (Fig.5-A).In particular,Fig.5-B and C show that the Red,NIR and SWIR 1 bands,as well as the GHS and TS2 phenological phases,were the most important features for separating ratoon rice from other land cover types.These selected optimal spectro-phenological features provide fundamental information for developing the PRVI to effectively identify ratoon rice.

Fig.5 Crucial bands and phenological phases of ratoon rice.A,final set of optimal spectro-phenological feature charts for identifying ratoon rice.B,the mean separability between ratoon rice and other land cover types in each crucial band was calculated as the mean value of the horizontal axis in A.B,blue;G,green;R,red;NIR,near infrared;SWIR 1,short-wave infrared 1;SWIR 2,short-wave infrared 2.C,the mean separability between ratoon rice and other land cover types in each crucial phenological phase was calculated as the mean value of the vertical axis in A.SS,seedling;TS,tillering;BS,booting;GHS,grain filling and harvested;TS2,tillering of ratoon crop;BS2,booting of ratoon crop;GHS2,grain filling and harvested of ratoon crop.

(3) Establishing the PRVI formula.Based on the sensitivity and separability analysis of ratoon rice properties (Fig.5),three bands of HLS data (i.e.,Red,NIR and SWIR 1)were selected for the crucial phenological phases (i.e.,GHS and TS2) to establish the PRVI.These three bands are also effective in characterizing the greenness of the vegetation,leaf structure,and water absorption (Choudharyet al.2021;Yanget al.2022).The mean reflectance values of the three bands over all training samples in the whole growth period,which can provide additional information for PRVI formula development,are shown in Fig.6.Note that ratoon rice presented a relatively high reflectance over all three bands in the critical phenological phases (DOY 210-250),but the magnitudes of the reflectance in these three bands differed greatly.Specifically,the reflectance variation of ratoon rice in the Red band was larger than that of non-ratoon rice crops on DOY 210-250 as ratoon rice experienced harvest to sprouting,which could indicate that the leaves are changing from green to yellow to light green (Fig.6-A).The reflectance values of ratoon rice in the NIR band varied sharply in the crucial phenological phases (DOY 210-250) due to ratoon rice regrowth after harvest,while the reflectance of non-ratoon rice crops varied only slightly (Fig.6-B).The reflectance change of ratoon rice in the SWIR 1 band was more prominent than those of other land cover types due to the water management in DOY 210-250 (Fig.6-C).Moreover,the reflectance of ratoon rice was the highest in the NIR band,followed by the SWIR 1 and Red bands among the three.Thus,to maximize the difference in the leaf chlorophyll contents and water signatures between ratoon rice and the other land cover types in the crucial phenological phases,we introduced both the SWIR 1 and NIR bands.Considering the divergence between ratoon rice and other land cover types regarding the phenological phases and separability,the PRVI was developed by combining the Red,NIR,and SWIR 1 bands.The PRVI formula can be expressed by eq.(4):

Fig.6 The temporal variability of the Red (A),NIR (B) and SWIR 1 (C) reflectance in 2019 for the different land cover types.DOY,days of year.

whereρRed,ρNIR,andρSWIR1are the reflectance values of the Red,NIR,and SWIR 1 bands in the HLS data,respectively.

Random forest classification based on the PRVlWe used the random forest (RF) classifier,which is a treebased ensemble learning method (Breiman 2001),to map the ratoon rice based on the PRVI.Compared to other supervised classification methods (e.g.,decision tree classifiers,maximum likelihood classification),the random forest is advantageous in overcoming the impact of overfitting and reducing the sensitivity to outliers in the training samples (Belgiu and Dr?gu? 2016;Lucianoet al.2019).Moreover,it can provide the vital features of target crops based on the samples and modeled complex relationships,making it popular in assisting feature selection for land cover mapping (Xiaet al.2022).mtryandntree,which refer to the number of variables randomly selected by the decision tree and the total number of trees generated in the model,respectively,are two important parameters in random forest models.In this study,we set themtryvalue asand ntree as 500 (Pal 2005),wherepis the total number of features in the training set.When training the random forest model,two-thirds of the total samples were used to construct each decision tree,and the remaining one-third of the samples were used to verify the classification results of each decision tree.

PRVl performance evaluationTo evaluate the performance of the PRVI for distinguishing ratoon rice from other land cover types,three other widely used vegetation indices,i.e.,NDVI,EVI,and LSWI,were implemented for the comparative analysis.EVI is an effective indicator for tracking phenological events of crop growth as well as assessing and monitoring seasonal variations of crops and evergreen vegetation (Gurunget al.2009;Sonet al.2014).NDVI not only monitors the spatiotemporal changes in the rice canopy but also detects changes in soil conditions (Xiaoet al.2005;Paulet al.2020).LSWI is sensitive to changes in soil and vegetation water contents and can sufficiently capture the unique flooding characteristics of rice (Delbartet al.2005;Boschettiet al.2014).The NDVI,EVI,and LSWI were expressed according to eqs.(5)-(7),respectively:

whereρBlue,ρRed,ρNIR,andρSWIRare the reflectance values of the Blue,Red,NIR,and SWIR bands in the HLS data,respectively.The performances of these indices were compared to that of the PRVI from three aspects,including the vegetation index temporal profile analysis,SIcalculation and ratoon rice mapping accuracy.Four scenarios of monotemporal in GHS,monotemporal in TS2,multitemporal in GHS-TS2 and multitemporal in SS-GHS2 were selected to evaluate the performances of the PRVI,NDVI,EVI,and LSWI for identifying ratoon rice across the different phenological phases.Additionally,since the PRVI was established with the Red,NIR,and SWIR 1 bands,we also added the combination of LSWI and NDVI (LSWI+NDVI) for performance evaluation.Based on the independent validation samples (30% of the total field samples),we adopted the user’s accuracy (UA) and producer’s accuracy (PA) indicators derived from the confusion matrix to evaluate the ratoon rice mapping accuracy.

3.Results

3.1.Temporal variability of the PRVl,NDVl,EVl,and LSWl for ratoon rice and other land cover types

The temporal variability of the PRVI,NDVI,EVI,and LSWI derived from HLS time series images over all land cover types in the study area is shown in Fig.7.Note that all four vegetation indices exhibited the largest overall separability between ratoon rice and other land cover types in the GHS-TS2 stages,indicating the reliability of these phenological phases.The PRVI performed better than the NDVI and EVI at the GHS-TS2 stages when distinguishing ratoon rice from other rice types (i.e.,single rice and double rice) since the PRVI increased the difference of canopy and water background reflectance in the harvest stage of the first season of ratoon rice (Fig.7-A,B and D).Besides,the PRVI,NDVI,and EVI were able to distinguish rice from non-rice land cover types.The LSWI (Fig.7-C) showed greater variability than PRVI,NDVI,and EVI in the GHS-TS2 phases because the ratoon rice fields need drought conditions for mechanical harvesting and irrigation for tillering and growth during the ratoon season.Nevertheless,the overlapping LSWI temporal trajectories among ratoon rice,double rice and non-cropland at the TS2 stage introduced serious uncertainties for the identification of ratoon rice.Thus,the large differences observed in the PRVI results between ratoon rice and the other land cover types hold great potential for extracting ratoon rice in practice.

Fig.7 The temporal trajectories of phenology-based ratoon rice vegetation index (PRVI) (A),normalized difference vegetation index (NDVI) (B),land surface water index (LSWI) (C),and enhanced vegetation index (EVI) (D) based on Harmonized Landsat and Sentinel-2 (HLS) time series images in 2019 over all land cover types.The pink shading represents the standard deviations of the vegetation indices for ratoon rice.SS,seedling;TS,tillering;BS,booting;GHS,grain filling and harvested;TS2,tillering of ratoon crop;BS2,booting of ratoon crop;GHS2,grain filling and harvested of ratoon crop.Doy,days of year.

3.2.Comparison of spectral separability between PRVl and the other indices

The pairwise separability and average global separability between ratoon rice and the other four land cover classes (i.e.,double rice,single rice,other crops,and non-cropland) are presented in Fig.8.The horizontal axis and vertical axis represent seven phenological phases and four vegetation indices,respectively.As shown in Fig.8-A,ratoon rice and double rice could be easily identified using the PRVI and NDVI at GHS2 due to their large SI in the harvest stage.This is because the ratoon rice had a short growth period in the ratoon season and was harvested in mid-October,while the late double rice grew from July to November.Moreover,Fig.8-B demonstrates that the PRVI,NDVI,EVI,and LSWI were able to identify ratoon rice and single rice at TS2,which can be primarily explained by the differences in growth features between them.At this stage,the single rice flourished and the leaves appeared green,whereas the ratoon rice had begun to sprout and its stubble appeared yellow-green,and the fields were filled with water.Furthermore,the PRVI was better than the NDVI,EVI,and LSWI at separating ratoon rice and single rice at GHS.Additionally,BS and GHS were important phenological phases for distinguishing ratoon rice from other crops and non-cropland with the PRVI,NDVI,and EVI (Fig.8-C and D).During the BS and GHS stages,the ratoon rice showed a flooding signal and then reached its peak canopy growth successively,whereas the other crops and noncropland were sown and growing during these stages.The average global separability of the PRVI,NDVI,EVI,and LSWI in terms of distinguishing ratoon rice from other land cover types is shown in Fig.8-E.Among these vegetation indices,the LSWI performed the worst in identifying ratoon rice because it could not distinguish ratoon rice from double rice,other crops,and non-cropland.In contrast,the PRVI was extremely effective in identifying the physiological characteristics of ratoon rice and attained the highest separability over to the NDVI,EVI,and LSWI,especially at the crucial phenological phases,i.e.,GHS and TS2.

Fig.8 The spectro-temporal SIcj charts and charts based on the phenology-based ratoon rice vegetation index (PRVI),normalized difference vegetation index (NDVI),enhanced vegetation index (EVI),and land surface water index (LSWI) at seven phenological stages.A-D,interclass separability between ratoon rice and double rice,single rice,other crops and non-cropland,respectively,while E indicates the average global separability for ratoon rice.SS,seedling;TS,tillering;BS,booting;GHS,grain filling and harvested;TS2,tillering of ratoon crop;BS2,booting of ratoon crop;GHS2,grain filling and harvested of ratoon crop.

3.3.Ratoon rice mapping and accuracy assessment

As described above,both GHS and TS2 are crucial phenological features for separating ratoon rice and other land cover types.To further understand the potentials of the GHS and TS2 stages,we also selected the whole growth period (SS-GHS2) and designed several phenological-vegetation index scenarios (GHS,TS2 and GHS-TS2) for ratoon rice mapping.Based on validation samples over different vegetation indices and phenological scenarios,the PA and UA of ratoon rice are shown in Fig.9.The accuracy assessment results demonstrated that the multitemporal scenario had better classification accuracy than the monotemporal scenario,as expected.Specifically,the GHS-TS2 scenario combined with the PRVI provided the best mapping performance,with PA and UA of 92.22 and 89.30%,respectively.It is worthwhile noting that the ratoon rice could be better identified with the PRVI in the crucial phenological phases (GHS-TS2) than across all phenological phases (SSGHS2),which was shown by the PA increasing from 85.78 to 92.22% and the UA increasing from 84.44 to 89.30%.This result indicated that the inclusion of more features in the classification not only limited the computational efficiency but also reduced the identification accuracy due to feature redundancy (Huet al.2019,2021).Based on the evaluations of different vegetation indices,the PRVI showed better performance for mapping ratoon rice than NDVI,EVI,and LSWI,and the maximum accuracy achieved was over 92%.Moreover,despite the dramatic changes in the canopy leave structure and water content,the accuracy of the PRVI for ratoon rice identification was higher than that of the combination of NDVI and LSWI (PA=83.33%,UA=88.20%) in the crucial phenological phases,indicating that the vegetation index formula is also critical for crop mapping.

Fig.9 The producer’s accuracy and user’s accuracy of ratoon rice identification over different combinations of vegetation indices and phenological stages.LSWI,land surface water index;EVI,enhanced vegetation index;NDVI,normalized difference vegetation index;PRVI,phenology-based ratoon rice vegetation index.GHS,grain filling and harvested;TS2,tillering of ratoon crop;SS,seedling;GHS2,grain filling and harvested of ratoon crop.

According to the evaluation results of the PRVI at crucial phenological stages,the proposed PRVI was then used to generate the spatial distribution of ratoon rice in Qichun County.The ratoon rice was mainly located in the central,hilly and lakefront parts of the study area from southwest to northeast.Additionally,the slope of the ratoon rice planting region was less than 7° for the purpose of harvesting by machines,as demonstrated in Liuet al.(2020).Thus,ratoon rice is rarely found in the northern part that is dominated by the mountainous areas.The original HLS pseudocolor image (RGB: NIR,red and green bands) on DOY 210 as well as the classification results based on four vegetation indices in the typical area with the distribution of different rice cropping systems is shown in Fig.10.At this observation date,single rice achieved peak growth (dark red pixels in Fig.10-A),whereas ratoon rice and double rice had reached the end of main crop growth (light red pixels in Fig.10-A) and the mature stage of early rice,respectively.The classification results derived by the PRVI exhibited better performance than those of the NDVI,EVI and LSWI,indicating that the PRVI is promising for distinguishing ratoon rice from the other land cover types.

Fig.10 The spatial distribution of ratoon rice and other land cover types in the typical sub-area of Qichun County,Hubei Province,China.A,pseudocolor (RGB: NIR,Red and Green bands) HLS image on days of year (DOY) 210;B-E,classification results derived from the phenology-based ratoon rice vegetation index (PRVI),normalized difference vegetation index (NDVI),enhanced vegetation index (EVI),and land surface water index (LSWI),respectively.

4.Discussion

4.1.Specific advantages of the proposed PRVl

Ratoon rice is a special rice cropping system that involves reaping two harvests from a single sowing,leading to a similar temporal variability between ratoon rice and double rice.Thus,the remote sensing mapping of ratoon rice is more challenging than the mapping of traditional rice cropping systems,such as single rice and double rice cropping systems (Liuet al.2020).Due to the convenience of spectral indices in highlighting a particular flooding type and the chlorophyll in leaves,several vegetation indices,e.g.,NDVI and LSWI,have been developed in recent decades (Zhaoet al.2021).Nevertheless,few indices have been developed to identify a specific rice cropping pattern.The primary reason is the complexity of different rice cropping pattern properties and the similarity of their spectra,which make it difficult to separate ratoon rice from various land cover types.

In this study,the PRVI was developed to address these limitations by capturing the unique spectral characteristics of ratoon rice at specific phenological stages.First,the optimal bands and phenological phases for identifying ratoon rice were derived by feature selection.Then,the PRVI was established based on the combinations of the selected Red,NIR and SWIR 1 bands,which could effectively reduce the redundant spectral effects.The chlorophyll pigments present in green leaves are known to strongly absorb solar radiation in the Red band,and features related to leaf structural properties and biomass show high reflectance values in the NIR band (Choudharyet al.2021).The SWIR band has good absorption of radiation from land covered by water (Choudharyet al.2021),which can sufficiently represent the typical growth features of ratoon rice.The PRVI that integrated the Red,NIR,and SWIR 1 bands could identify the different leaf chlorophyll contents and water signatures between different rice types,and capture the typical growth characteristics of ratoon rice in the GHS-TS2 phase.Specifically,the single rice gradually grew vigorously and the leaves changed from light green to green in the GHS-TS2 phase.In terms of double rice,the early rice was harvested in July (early in the GHS phase),and the late rice was transplanted and subsequently grown,with light green leaves observed.The ratoon rice progressed from harvest to sprouting,and then the stubble appeared yellow-green,and the fields were filled with water during the GHS-TS2 phase.In general,these three rice types were in different growing stages in the GHS-TS2 phase and exhibited very different chlorophyll pigment contents and biomasses,leading to the specific spectral signatures for each rice type.Although the PRVI was established with the Red,NIR,and SWIR 1 bands,its accuracy for identifying ratoon rice was higher than those of the three individuals bands or all bands (Appendix A),indicating that the PRVI could effectively capture crucial spectrophenological features of ratoon rice and thus performed well in ratoon rice identification.

Additionally,compared with the NDVI,EVI,LSWI,and the combination of NDVI and LSWI,the PRVI showed good performance in discriminating the specific signals of different rice types and played an important role in identifying ratoon rice in our study region.Furthermore,it is noteworthy that the PRVI offers great potential for identifying ratoon rice in areas where cloudy or rainy conditions tend to limit the temporal coverage of satellite observations.For instance,the PRVI performed the best in identifying ratoon rice during the crucial phenological stages (GHS-TS2),which indicates that the PRVI can be used to determine the earliest planting timing of ratoon rice.

4.2.Potential of HLS products for ratoon rice mapping

In this study,we explored the potential of HLS products with 30 m spatial resolution for ratoon rice identification.The HLS product that combines Landsat-8 and Sentinel-2 data to directly generate images with high spatiotemporal resolution can overcome the drawbacks of the currently available 30 m resolution crop intensity maps generated by the fusion of MODIS and Landsat data (Haoet al.2019).The spectral diversity and temporal variability between different crop types must be considered in crop mapping,and the integration of such spectrotemporal information can largely improve the accuracy of the classification results (Xiaet al.2022).The dense time series provided by HLS data can capture crucial phenological information and record the spectral features needed for ratoon rice identification over fragmented croplands in the cloudy and rainy southern areas.

To better understand the advantages of the HLS products for identifying ratoon rice,we compared the annual maps of ratoon rice derived from the L30,S30,and HLS images during crucial phenological phases (Table 2).The ratoon rice map generated by the L30 data showed an unacceptably low accuracy,with PA and UA of 36.66 and 47.83%,respectively.Obviously,most ratoon rice pixels could not be identified based on the L30 image alone due to the lack of sufficient quality observations that are essential to capture crucial spectro-phenological features.Although the HLS data had only one more image than S30 in the GHS-TS2 stages,the UA and OA of the ratoon rice map derived from HLS were greater (by 4 and 6%,respectively) than those of the map derived by S30 data.This improved accuracy indicated that more high-quality images are necessary to reflect the specific temporal features of ratoon rice.Specifically,due to the different ratoon rice management practices,the important GHS and TS2 stages cover a range of changes in morphological characteristics and water contents,which were also captured by the high-quality HLS time series.

Table 2 Accuracy assessment of the derived ratoon rice maps using L30 (30 m spatial resolution is derived from Landsat imagery),S30 (30 m spatial resolution is derived from Sentinel-2 imagery),and HLS (Harmonized Landsat Sentinel-2) data with the PRVI during crucial phenological phases1)

4.3.Limitations and future prospects

In this study,the proposed PRVI exhibited a favorable performance for identifying ratoon rice,with an accuracy of around 90%.Nevertheless,several limitations need to be considered in future studies.First,the PRVI was developed for identifying only ratoon rice,so it may not effectively distinguish ratoon rice and other rice cropping patterns simultaneously.Specifically,rice cropping systems are usually accompanied by complex crop rotations,e.g.,rapeseed-rice and wheatrice rotations,which can also affect the growth period of rice.Additionally,some early rice (single rice) is unmanaged after harvest,and tillers will regrow from the stubble.This makes it difficult to distinguish from ratoon rice,as they share similar temporal patterns within the vegetation indices.Thus,further research is needed to concurrently improve the identificationaccuracy of ratoon rice and other rice cropping systems.Second,this study demonstrated that the PRVI could identify ratoon rice effectively and even more accurately than the combinations of several spectral indices (e.g.,PRVI+NDVI,PRVI+EVI,PRVI+NDVI+LSWI,PRVI+EVI+LSWI,and PRVI+NDVI+EVI+LSWI) during the crucial phenological phases (Appendix B),which can be attributed to the feature redundancy by involving unnecessary features in the classification (Huet al.2019).However,only several scenarios were explored in this study,whether the combination of PRVI and other indices can improve the ratoon rice mapping are needed to be considered in future work.Third,the potential for identifying ratoon rice by the PRVI was only evaluated for Qichun County in 2019.In other regions or years,especially before 2015 when HLS data are not available,the robustness and portability of the PRVI in terms of the combinations of different remote sensing data sources (e.g.,GaoFen-1/6,MODIS,SPOT,etc.) need to be further explored.Finally,due to the spectral heterogeneity for the same crop type and the spectral similarity among different crop types,the same crop within a cropland parcel may be misclassified into different types,introducing further uncertainty into the classification results.The cropland parcel is the basic unit for agricultural production management,which can provide an important reference for crop type identification.Therefore,identifying crop types at the cropland parcel scale is an important future prospect for improving the accuracy of crop mapping.

5.Conclusion

The spatial distribution of ratoon rice provides the fundamental dataset for agricultural monitoring and cropping pattern adjustment.In this study,a new vegetation index named phenology-based ratoon rice vegetation index (PRVI) was developed to map ratoon rice at a 30 m spatial resolution using HLS time series images.According to the analysis of spectro-phenological separability between ratoon rice and other crop types and feature selection,the PRVI was established based on the Red,NIR,and SWIR 1 bands in crucial phenological phases of ratoon rice.Subsequently,the PRVI was used to map the spatial distribution of ratoon rice in Qichun County and its performance was comprehensively evaluated and compared to those of the NDVI,EVI,and LSWI.The results suggested that the GHS-TS2 stage is the best phenological period for distinguishing ratoon rice from other land cover types due to the dramatic changes in the canopy leaf structure and water contents.Based on field samples,the PRVI exhibited the best performance for ratoon rice mapping among all vegetation indices,i.e.,the NDVI,EVI,LSWI and the combination of NDVI and LSWI at the GHSTS2 stage,with PA and UA values of 92.22 and 89.30%,respectively.The results of this study indicate the great potential of the PRVI for ratoon rice identification in areas with fragmented agricultural landscapes and complex rice cropping systems,which is promising for decision-making in the promotion of ratoon rice to improve agricultural production with high yields and low labor costs.

Acknowledgements

This work was supported by the National Natural Science Foundation of China (42271360 and 42271399),the Young Elite Scientists Sponsorship Program by China Association for Science and Technology (CAST) (2020QNRC001),and the Fundamental Research Funds for the Central Universities,China (2662021JC013,CCNU22QN018).

Declaration of competing interest

The authors declare that they have no conflict of interest.

Appendicesassociated with this paper are available on https://doi.org/10.1016/j.jia.2023.05.035