| Due to the sharp decrease of agricultural labor population in China,it is an inevitable choice for China’s future grain production to take the road of large-scale utilization of cropland field.The traditional research on cropland field scale is usually based on statistical data and cropland field patches,but the statistical data analysis of cropland scale lacks spatial distribution information,and the cropland field patches lacks the specific morphological characteristics of cropland.Cropland field is the basic unit of agricultural production,which can be used as an important indicator to measure and characterize the scale of cropland field land use.However,there is a lack of cropland field extraction from high-resolution remote sensing images and research on the scale of cropland field large-scale utilization.Based on this,this study focuses on the extraction of cropland field and the characterization of cropland large-scale utilization based on cropland field.As a widely used deep learning model,Convolutional Neural Network(CNN)can form more abstract high-level features by combining low-level features,which is feasible to extract cropland from high-resolution remote sensing images.However,there are many CNN models,and each has its own advantages and disadvantages.Considering the difficulty of CNN network selection,the limitation of training samples and the design of network parameters,this thesis uses the Ensemble Deep Learning(EDL)model to integrate the advantages of each CNN network(FCN、Ps PNet、Seg Net、Unet)to achieve land extraction in high-resolution remote sensing images and improve the extraction accuracy of cropland field.EDL firstly obtains different training sets by bagging sampling method,then applies the training sets to the several CNN,calculates the corresponding probability of cropland boundary pixel by pixel,and finally integrates the probability map according to the average value to obtain the boundary of cropland,so as to realize the extraction of cropland.Landscape index of cropland is the main evaluation method for the large-scale utilization of cropland field,but there are many landscape indexes,and each factor has redundant information,so it can not be directly used to represent the scale of cropland.In this study,through the principal component analysis of landscape index,we can get the CS(crop size)value to describe the scale of cropland,describe the fragmentation of cultivated land,and then evaluate the scale of cultivated land use.Based on this,this thesis selects Fujin city of Heilongjiang Province,Qianjiang City of Hubei Province,and Wuming District of Nanning city of Guangxi Zhuang Autonomous Region as the research areas.The natural conditions,economic conditions and agricultural mechanization level of these three research areas are different.Then,the EDL model is used to extract cropland field,and six landscape indexes(Mean Patch Size,Patch Density,Edge Density,Area-Weighted Mean Shape Index,Area-Weighted Patchfractal Dimension,Patch Stability)are extracted based on cropland field,and the characteristics of cropland use scale in these areas are analyzed.After data standardization processing and factor analysis applicability test,the cropland land use scale in three study areas is analyzed.Finally,the comparative study of cropland scale was carried out.The results show that: 1.The overall accuracy of EDL method is 96%,which is 1% higher than FCN,Seg Net and Unet,and 2% higher than Ps PNet.Compared with a single classifier,EDL can reduce the bias and improve the accuracy of parcel extraction.EDL model also can integrate the advantages of multiple convolutional neural networks and improve the classification accuracy.2.Through principal component analysis,the CS value of Fujin city is 1.6.The CS value of Qianjiang City is 0.82,and that of Wuming district is-2.42.That is to say,the land fragmentation of Fujin city is the lowest and the scale of cropland is the largest.The land fragmentation of Qianjiang City is low and the scale of cropland is large.Wuming district has the highest land fragmentation and the smallest cropland scale.The EDL model proposed in this study can better extract cropland field,which provides a new method for the extraction of cropland field;the use of principal component analysis method can quantitatively describe the scale of cropland field,further grasp the internal law of the distribution of cropland scale in China,and provide decision support for the formulation of China’s land policy. |
