Study On The Extraction Of Plastic Greenhouse Based On Landsat 8 OLI Image

Plasticultural has the functions of heat preservation,moisture retention and soil conservation,which can effectively improve the yield and quality of crops,so it has been widely used.The global plasticultural area(mainly including greenhouses,low and high arches,and plastic-mulch landcover)is rapidly expanding at a rate of 20%every year.In 2017,the global plasticultural area has reached 20.22 million hm~2.However,with the widespread use of plastic agricultural films,many problems such as environmental pollution have also been brought.Therefore,it is necessary to quickly and accurately obtain plasticultural area and its distribution information to monitor it in order to provide decision-making basis for agricultural management,environmental protection and other issues.At present,the research on the extraction of plastic agriculture using remote sensing methods is mainly concentrated in agricultural developing areas with large areas of densely distributed plasticultural area and a single type of surface landcover.There are still huge challenges for the extraction of plasticultural area with scattered distriution in areas with complex landcover.In addition,most of the existing researches on the extraction of plasticultural area used high-resolution images,which have the characteristics of small study area,and the algorithm is not suitable for large area extraction of plasticultural area.Based on this,this paper attempts to extract scattered plastic greenhouses in areas with complex landcover using medium-resolution remote sensing images.In this study,Xuzhou city,Jiangsu Province was selected as the research area,and Landsat 8 OLI remote sensing images were used as data sources to extract the plastic greenhouses in Xuzhou.The object-based method and the deep learning algorithm were used to extract the plastic greenhouses,and the advantages and disadvantages of the two methods were compare,then the method with high precision was selected to extract plastic greenhouses in Xuzhou from 2014 to 2018 and the temporal and spatial changes of plastic greenhouses were analyzed.The main research work is as follows:(1)According to the information of seven band of Landsat 8 image,the spectral curves of plastic greenhouse and other land types were analyzed.Based on the spectral and geometric characteristics of the ground objects,a threshold model is established to extract the plastic greenhouses.Firstly,the images was preprocessed including radiometric correction,FLAASH atmospheric correction and image clipping.Then,the spectral differences between the plastic greenhouses and other landcover were analyzed.Five distinguishing features were established,namely Brightness,Band NIR,PGHI,CSBI and PMLI.Then the Shape Index and Shape Index 1 are established according to the geometric characteristics of the plastic greenhouses.Brightness was used to remove vegetation,dark water and dark buildings;Band NIR was used to remove bright water;PGHI was used to remove bare land and white buildings;PMLI was used to remove plastic-mulched landcover.In order to optimize the classification results,Shape Index and Shape Index 1 were used to remove irregular objects.Based on the above seven features,the threshold model was established,and the plastic greenhouses of Xuzhou in 2014 was extracted by the object-based method,with the accuracy of more than 95%.(2)As a new technology of image recognition,deep learning can automatically learn image features for classification.The U-net model for small sample classification was studied to extract plastic greenhouses.This paper improved the U-net model input layer according to the characteristics of Landsat 8 images.With the help of high-resolution remote sensing images,real land surface labels were made,which were divided into two categories“plastic greenhouses”and“others”.The images in 2014were cut into 32×32 size pictures,and a total of 19800 pictures were obtained.Choosing 13860 pictures as training samples,3960 pictures as verification samples and1960 pictures as test samples.Under the framework of tensorflow,the improved U-net model based on sample training was used to predict the test pictures,and the overall accuracy is up to 86%.(3)Comparing the two methods,object-based method is selected to extract the plastic greenhouses of Xuzhou from 2014 to 2018.It is found that many scattered plastic greenhouses decreased in 2015 and 2016 compared with 2014 because of El Niro incident;In 2017 and 2018,plastic greenhouses increased year on year,which mainly distributed in Peixian County,the northwest of the main urban district,Suining and the east of Xinyi,making the distribution of plastic greenhouses more and more denser.This was mainly affected by the land transfer system,more and more farmers transfer land to professional companies for management,which resulted in that the increase of dense plastic greenhouses and the decrease of scattered plastic greenhouses.