Research On Methods Of Dynamic Crop Recognition Using Incomplete Time Series Data

Timely and accurate spatial distribution information of crops can provide basic data for crop yield estimation,land suitability evaluation,land reclamation and other applications.The crop recognition methods based on time series data can make full use of the specific growing pattern and phenological characteristic of various crops to improve the accuracy of crop classification and have been widely adopted as the solution for crop information extraction and crop recognition.However,these methods generally rely on complete time series data,which delays the acquisition of crop mapping and reduces its application value.Therefore,it is of great significance to study timely and dynamically updatable method for crop classification.The main difficulties in implementing such methods are:(1)the medium-resolution time series data is sparse,time-sampling irregular,and part of the data is missing,which makes the dates of data acquired in different years are inconsistent,and reduces the comparability between feature vectors in different years.(2)there is only partial time series in the early stage of the previous period which cannot be fitted when crop maps are updated,making it difficult for traditional time series classification methods to be used directly.Based on the above analysis,this paper develops a dynamic crop recognition method using incomplete time series data.Main conclusions were drawn from this paper as follows:(1)A method of dynamic identification and classification of crops is proposed.Recurrent Neural Networks such as LSTM need to obtain the hidden layer status of all nodes for comprehensive judgment to determine the category of samples,and are difficult to meet the needs of early identification and dynamic updating of crop map.In this paper,we first develop an improved strategy to extract LSTM hidden layer state directly from the beginning to the current time to form state vector.After mean pooling layer and fully connected layer processing,the type currently recognized by the network is obtained,thereby realizing the dynamic training of classifiers by using the time series data of the reference year,and then the crop classification map can be updated continuously according to the input of the interesting year.To the best of our knowledge,this is the first experiment in which LSTM network is applied to dynamic crop identification with good consistent classification results.(2)A time series data reconstruction method with partial dimension missing is developed.Aiming at the problems of sparse time series data,irregular sampling time and partial key data loss due to factors such as sensor selective observation,cloud and cloud shadow,etc.,this paper firstly designs a noise detection algorithm based on discriminant rules to detect noise points in the original data.The denoised time series data is fitted to the curve through the optimized double logistic model,and the missing values are replaced by the fitted values,thereby obtaining seamless time series observation values to meet the crop identification requirements.In this study,Shouxian,Anhui Province,was selected as the study area where the main crop types were the targeted land cover classes in this study.In addition,Sentinel ? time series images from May 2017 to February 2019 and agricultural census data were employed as the main study datasets.The samples collected in the study area were classified using the dynamic identification method this paper proved.Experimental results show that the overall accuracy can reach 95.1%,while the classification accuracy of autumn crops is 96.4% and that of summer crops is 93.8%.All kinds of classification accuracies are significantly higher than that obtained by Random Forest.At the same time,the validity of proved method is verified by expert interpretation and field investigation.Besides,we evaluate the support effect of the proved method on the application of agricultural insurance in the study area.In the future work,we will further study the extension of the proved method on multi-source image data sets and several study areas,and enhance the it's robustness by combining the time series of different remote sensing indicators.