Extraction For Cropping Pattern Based On Assimilating Multi-source Remote Sensing Imagery

Cropping pattern is defined as the spatial distribution and succession of crop types on an agricultural unit during one year,which is a concrete form for cropland use.Investigation for cropping pattern is critical for global food security,efficient cropland management,and environment protection.Remote sensing with charactistics of large extent and near real time has been a main technology used for crop classification and extraction.Specially,the available multi-source multispectral imagery and its free policy enhance the cropping pattern investigation and application.However,some methodology drawbacks and technical difficulties(e.g.,spatial-temporal dataset reconstruction,classification methods,etc)has limited the cropping pattern research with remote sensing approach.At first,the quantity and quality of satellite imagery have impeded the construction for spatial-temporal dataset in a long time.Recent development of multi-source satellite imagery provides a new opportunity for spatial-temporal dataset.By integrating multi-source satellite imagery,there will be a real dataset rather than a simulated dataset,which is essential for remote sensing application.But the spectral difference among multi-source satellite imagery will affect their integration.Thus,study on assimilation of multi-source imagery is the basis of combining these satellite images.Secondly,previous study on cropland use are based on the shape of time series vegetation indices,which ignored many information of actual cropland use status.And then,the definition for cropping pattern is not clear.Most research issues are focusing on local crop types classification therebefore,involving no structure information of crop planting.And the classication methods for crop types are basicly supervised classification based on training samples,which has limitation on different time,region,and dataset.In addition,the lack of automatic classification and extraction methods for crop types also limit the application of agriculture management.Therefore,our objectives are mainly including:(1)spectral assimilation for multi-source imagery,(2)reconstruction of spatial-temporal dataset with high resolution by integrating multi-source imagery,(3)extraction for cropland use by combing time series NDVI and actual planting states,and(4)extraction for cropping pattern under the supporting of cropland use results and automatic crop type classification methods.The main research issues and conclusions include:(1)Cross-comparison the continuity of Sentinel-2 MSI,Landsat-8 OLI,Landsat-7 ETM+,GF-1 WFV,HJ-1 CCD images.We used all images in 2017 with a spatial extent of 10°×10°,located in Hubei and near Hubei Province,China,as datasets for spectral assimilation.The pre-processing for these images included othrocorrection,registration,tiling,atomosphere correction,and nadir BRDF adjust.Finally,we collected large amounts of samples by sampling every ten pixels for building the assimilation model between each two-source imagery.The results shown that all imagery exsiting good linear fitting relationship.And the Landsat-8 OLI is the best target for assimilation of all five source satellite imagery.(2)This study developed an image composite method among multi-imagery based on the results of spectral assimilation.In order to meet the demands of extraction for cropland use and cropping pattern,we reconstructed the time series multispectral images with visual and near-infrared bands by integrating all images.And then,we calculated the time series NDVI using the integrated dataset.The MODIS NDVI product were used to assess the quality of our NDVI datasets.The result shown good performance both on NDVI values and temporal tendency on every land cover types by comparing two NDVI datasets.(3)In this study,a new cropland use extraction framework was poposed based on the actual survey of study area.Moerover,the spectral window method in information processing was used to modify the searching strategy of extraction for phenological features.In this way,the algorithm execution efficiency has been accelerated largely.In our cropland use extraction results,the phenomenon of seasonal abandonment in the study area were actualy extracted,which was not performed in the previous methods.While this is very important for cropland management.(4)A franmework of "cropland use-multiple planting area-thematic crop types extraction—integration results for cropping pattern" was proposed in this study,who is not only about the crop type classification but also crop type succession regular.A new automatic classification method for oilseed rape and winter wheat was peoposed,basing on the spectral and color features during the flowering period.The tests on different time,region,and dataset were also obtained good performance with larger than 90%overall accuracy.Therefor,this method has good robustness and can be used to regional mapping producing.Moveover,the special spectral characteristics of red edge among paddy rice,soybean,and corn has been found.And a dicision tree method based on those features was developed to classify paddy rice and soybean,which achieved over accuracy larger than 90%.