| As one of the main crops in China,rice is also an indispensable food source in people ’ s daily life.Accurate monitoring of rice planting area in China is of great significance for China ’ s food security strategy,rice yield evaluation,crop planting mode and rural agricultural economic development.Traditional rice planting monitoring often relies on manual statistics,which is time-consuming,laborious and inefficient.In recent years,with the rapid development of remote sensing technology,compared with the traditional agricultural monitoring,agricultural monitoring based on remote sensing technology has entered the fast lane.Compared with manual statistical methods,agricultural information monitoring based on remote sensing technology has the advantages of wide monitoring area,strong timeliness,high efficiency,low cost and high accuracy.However,there are also some problems in the monitoring of rice planting information using remote sensing technology.Firstly,although optical remote sensing can better realize the monitoring of rice planting information,its imaging is greatly affected by weather conditions such as cloud and rain,so it is difficult to meet the requirements of some specific regions.Therefore,the monitoring of rice planting information based on microwave remote sensing technology has also been further developed.Secondly,the existing rice monitoring method based on remote sensing technology is simple,and the model accuracy is low,which is difficult to meet the needs of modern precision agriculture monitoring.Finally,it is more difficult to monitor rice information in hilly areas with complex terrain and fragmented plots.Therefore,it is of great significance to explore effective and reliable high-precision rice planting information monitoring methods.In view of the advantages and disadvantages of the current rice planting information monitoring based on remote sensing technology,this paper proposes a full convolution network rice extraction model based on time-series microwave remote sensing data.Sentinel-1A is selected as the data source to extract the backscattering coefficient of time series under VV and VH polarization,and the rice monitoring experiments are carried out in the typical plain rice planting area and the typical hilly rice planting area.Firstly,this method overcomes the factors that optical remote sensing images are affected by cloud and rain weather in the selection of remote sensing data sources.Secondly,according to the terrain and surface conditions of different rice planting areas,the rice planting area monitoring in hilly areas is further constructed based on multi-scale(time series,spatial neighborhood,characteristic index)full convolution neural network model.Based on the fully convolutional neural network model,the model uses the normalized water body,the normalized rice,the dynamic time warping(DTW)feature index to construct the feature index data set.After data input,the convolution operation is carried out on the three scales of the time series backscattering coefficient,the feature index data set and the pixel neighborhood range,respectively.The convolution features at different scales are extracted to improve the classification accuracy and reliability.After the extraction of rice,the phenology index was further constructed to analyze the rice growth phenology and realize the monitoring of rice growth phenology in different regions.In order to verify the rice monitoring methods of different rice planting areas,this paper first selects Fujin City,Jiamusi City,Heilongjiang Province,Northeast Plain as a typical plain rice planting area,which has flat terrain and simple crop planting structure.Rice,maize and soybean are the main crops.At the same time,rice samples were extracted from the existing 10-meter crop classification dataset and the sentinel-2A optical image data in Northeast China.The minimum distance method,random forest model and full convolution neural network model were used to extract and analyze the rice planting area and rice growth phenology in the region.The experimental results show that:(1)The classification accuracy of minimum distance method,random forest model and full convolution neural network model are 89.2 %,91.7 % and 96.2 %,respectively.the identification accuracy of rice was93.8 %,95 % and 97.8 %,respectively;kappa coefficients are 0.842,0.878 and 0.943,respectively;(2)The comprehensive comparison and analysis of the minimum distance method and the random forest model can better realize the monitoring of rice planting area.However,compared with the rice monitoring model based on the multi-scale full convolution neural network,there are certain differences in accuracy.In general,the multiscale factor full convolution neural network has the highest accuracy classification identification,and the model is the most reliable,which can well realize the monitoring of rice planting information in the plain area.The phenological period of rice growth in the Northeast Plain was analyzed by constructing the phenological index.It was found that the rice growth in this area had obvious phenological characteristics,and the phenological index was characterized by ’ double peaks and double valleys ’.In order to further realize and verify the monitoring of rice planting information in areas with complex terrain and climatic conditions,Fushun County,Zigong City,Sichuan Province is selected as a typical hilly rice planting area in this paper.The terrain conditions in this area are relatively complex,the climate is cloudy and rainy,and the plots with rich ground types are fragmented,so the extraction of ground types is relatively difficult.In order to obtain the real ground rice sample data,a five-day ground artificial survey was conducted in Zhongshi Town,Fushun County from May 25 to May 30,2021.The main surface features of rice,architecture,water and forest land were investigated.In order to verify the applicability of each rice extraction method model in hilly area,the minimum distance method,random forest model and full convolution neural network model are also selected to extract the rice planting area in this area.The experimental results show that:(1)The classification accuracy of minimum distance method,random forest model and full convolution neural network model is 84 %,87 % and 90 % respectively;the identification accuracy of rice was 86 %,88 % and 92 %,respectively;kappa coefficients are 0.8,0.84 and 0.87 respectively;(2)Through the comparison and analysis,it is found that the accuracy of rice planting extraction in hilly areas under the same method decreased to varying degrees compared with that in plain areas.However,the rice extraction based on the total convolution neural network model is still the most accurate and reliable method.On the basis of relatively high-precision rice identification,it also better retains the edge authenticity of the fragmented planting plots in hilly areas,and can well realize the extraction of rice planting information in hilly areas.In summary,this paper is based on the application of Synthetic Aperture Radar(SAR)in the field of agriculture to explore effective and high-precision rice extraction methods.The feature index data set was constructed,and the deep learning theory was introduced into the model.According to the actual situation of rice planting,the multi-scale fully convolutional neural network model was further constructed to realize the extraction of rice planting information in plain areas and hilly areas of typical rice planting areas.Compared with the traditional rice identification method based on remote sensing technology,the model has higher accuracy and reliability,and has stronger universality in different rice planting areas.At the same time,combined with the analysis of rice phenology,it can better monitor the rice growth in various regions.Finally,it provides some technical support for food security strategy,rice yield evaluation,crop planting mode and rural agricultural economic development. |
PDF Full Text Request