Remote Sensing-based Winter Wheat Identification Considering Temporal-spectral Intra-class Difference Characteristics Of Vegetation Index

Wheat is the worlds third-largest crop and is the most widely grown crop worldwide,of which the planting area of winter wheat accounting for more than 80%.The accurate and timely information of winter wheat areas is therefore vital for crop yield estimation,growth monitoring,and agricultural policymaking.The development of the remote sensing technology makes it possible to monitor winter wheat areas at fine spectral,temporal,and spatial scales.Specifically,remote sensing images with the wide-area coverage,short and reliable revisit periods,relatively easy data acquisition,and low cost provide a viable means to detect winter wheat fields.However,remote sensing-based crop identification is a complex task,which not only involves the practicality and execution of the technology but also considers the credibility and accuracy.The high-accuracy classification can hardly be achieved by a single image.Time series data,especially vegetation index(VI)time series,has been a hot topic in remote sensing-based crop classification.In recent years,with the development of earth observation technology,multi-source,multi-spatial-temporal resolution satellite sensors have emerged.The amount of remote sensing data at home and abroad has increased rapidly,which provides a wealth of data sources for remote sensing-based winter wheat identification.Both readers and researchers are commonly concerned about how to efficiently select the reasonable dataset to conduct winter wheat identification and improve the accuracy.In order to resolve the existing issues,this paper conducted the remote sensing-based winter wheat identification based on GF-1 WFV(Wide Field of View)images and MODIS(Moderate-Resolution Imaging Spectroradiometer)datasets.Firstly,focused on the North China Plain,we analyzed the temporal,spectral characteristics of winter wheat at different growing seasons using GF-1/? time series.Combing with the differences of time series spectral characteristics of winter wheat and other land features,the remote sensing-based winter wheat recognition method was explored.Based on the vector analysis,the angles and distances of the N-dimensional vector were introduced into the ? time series.Then a vector analysis model for remote sensing-based winter wheat identification was proposed based on the GF-1 WFV images.Meanwhile,considering the accuracy and timeliness,the optimum ? time series for winter wheat identification was further explored based on the proposed model.Secondly,the model was applied to MODIS EVI dataset and was further improved.The results were evaluated using the Landsat images and other statistic datasets.The evaluation showed that the accuracy of winter wheat identification had been improved than before.Finally,the applicability of the improved model was further evaluated using MODIS data in Kansas State,USA.After that,combining with the land fragmentation indices,the factors influencing winter wheat identification accuracy were further discussed in this study.The main conclusions of this study include:(1)Based on the spatial angle and distance of ? time series,the vector analysis model of remote sensing-based winter wheat identification was proposed.GF-1/WFV NDVI time series covered the whole growth period of winter wheat were constructed in this study.Based on the phenological characteristics of winter wheat,the temporal spectrum differences between winter wheat and other land types were compared and analyzed.Combing with Spectral Angle Mapping,taking NDVI time series as N-dimensional vector,the vector analysis model of remote sensing-based winter wheat identification was constructed.The developed model considered the spatial angles and distances characteristics of winter wheat in ? time series.Overall accuracies in the confusion matrix were evaluated by validation samples(94.83%).The accuracy was improved by 8.33%compared with other methods,which indicated our developed model could effectively realize the remote sensing-based winter wheat identification and achieve the higher accuracy.(2)Based on the developed model,the timeliness of remote sensing-based winter wheat identification was further explored.Considering the timeliness,the developed model was implemented to test the optimal time series data for winter wheat identification.The time series sequences covered different phenological periods of winter wheat were tested to select the optimum period with higher winter wheat identification accuracy.The evaluation showed that the accuracy of the ? time series covered sowing date to returning green stage reached more than 90%using this proposed model.(3)A vector analysis model considering temporal-spectral intra-class difference characteristics of EVI temporal spectrum was proposed for remote sensing-based winter wheat identification.Based on MODIS data,the EVI temporal spectrum characteristics of winter wheat was analyzed.We found that different growth conditions and landscape factors(e.g.,irrigation,fertility,climate,topography,fragmentation)at large areas could cause intra-class differences affecting the spectral signatures of winter wheat.Intra-and inter-class confusion due to these factors would degrade the ability of the image classifier to produce accurate maps for remote sensing-based winter wheat identification.In this study,an improved model for mapping winter wheat based on intra-class variability was presented that integrated the angles and distances of multidimensional vectors and adopted multiple subclasses as training samples.The improved model was applied to the North China Plain,and the results were evaluated by Landsat images and ground datasets.The evaluation showed that the overall accuracy of winter wheat reached 85.00%,which increased by 15%compared with the traditional method(maximum likelihood classification,MLC).Meanwhile,using the? time series from the sowing date to the returning-green stage,the accuracy was 70.17%in the North China Plain.(4)Applicability evaluation of the improved model of remote sensing-based winter wheat identification was conducted at different regions.In order to verify the applicability of the improved model,the paper applied the model in Kansas State in the USA,which was located in the similar latitudes as the North China Plain.The accuracy of winter wheat identification results in Kansas State was evaluated at the regional,county,and pixel scales,respectively.The results showed that the improved model could effectively identify the distribution of winter wheat at different regions.The mapping results compared well with those from multiscale ground information.Compared with Landsat data,the accuracy at the pixel scale reached 90.33%.Meanwhile,the ? time series covered from the sowing date to the retuming-green stage was used to validate the improved model further.The results also showed that the accuracy reached 80.67%.The application in Kansas State inhibited that the improved model proposed in this study has two key advantages over traditional methods:the high accuracy and wide application.(5)Based on landscape metrics,factors influencing the winter wheat identification model were clarified.The crop identification accuracy based on remote sensing images with the medium resolution is closely related to the spatial heterogeneity of farmland.Landscape metrics were used to express the landscape's fragmentation,and the impacts of fragmentation on the accuracy of remote sensing-based winter wheat identification were quantitatively analyzed.The results showed a strong positive correlation between landscape fragmentation index(FRG)and the accuracy(r = 0.99).Similarly,higher PLAND corresponded to higher winter wheat identification.When the PLAND value was more than 20%,the accuracy of winter wheat identification was more than 90%.Further analysis indicated that the model performed better in areas with lower landscape fragmentation,which may partly explain the relatively higher accuracy of winter wheat mapping in Kansas.Moreover,due to the specific land allocation policy in China,croplands in the same region might be dominated by different cropping patterns or growing times,which shows a high degree of fragmentation on remote sensing images and makes it more complicated for crop mapping.This means that GF-1 images are more suitable in China for winter wheat mapping in large areas.This also suggested that the MODIS EVI 250 m data has a significant advantage for crop mapping in intensively managed landscapes with a lower fragmentation degree.