| Over recent years,the rapid development of deep learning and its wide application in remote sensing observation,the ability to use this technology to extract crop information has been greatly improved.For crop growth period,which is a basic but extremely important agricultural information,a growing number of experts and scholars,using deep learning techniques,have conducted extensive research,but the results obtained are often difficult to meet the needs of remote sensing for high-precision identification of crop growth period on a large regional scale.To promote the solution of the above problems,this study attempts to establish a highprecision crop information extraction method by satellite remote sensing that fits the intrinsic characteristics of the target crop growth process using various deep learning algorithms,and applies it to the remote sensing analysis of rice types and their growth stages in Jianghan Plain.In this process,the study first analyses the growth characteristics of different rice types in the study area during their respective fertility periods and the differences between them and other features,constructs a remote sensing feature index data set consisting of two indices,NDVI and MNDWI,and then integrates the improved Mean Shift segmentation algorithm and SVM algorithm to perform image segmentation and classification in turn,and obtains Then,through the reconstruction of multispectral images covering the whole growth cycle of rice in the study area,the time-series spectral reconstruction images containing time-dimension information and recording the spectral characteristics of different types of rice and their growth process were obtained.Finally,by combining two different deep learning methods,Temporal Convolutional Network(TCN)and Long short-term memory neural Network(LSTM),the above time-series spectral reconstruction images are classified.Realize the remote sensing diagnosis of rice multigrowth period.The major accomplishments are as follows:(1)Based on sentinel-2 satellite images with long time series,a series data set of remote sensing multi-feature indexes covering the complete growth cycle of major rice types in the study area was constructed(including MNDWI index on February 21 and April 17,respectively).And on June 6,based on July 31,August 30,and October 24 September 14 th NDVI index),which keeps the spectral characteristics of the research in different rice own unique,and can objectively reflect the various types of rice at different developmental stages of growth status differences between laid the groundwork for subsequent rice types of high precision to distinguish.(2)Using the superpixel temporal remote sensing feature index time-series images generated based on the improved Mean Shift segmentation algorithm in this paper as the data source,the SVM algorithm was applied for classification,the accuracy of rice remote sensing classification can be improved.Classification accuracy evaluation was carried out based on the field survey sample plots,and the results showed that the overall classification accuracy(OA)and Kappa coefficient of this method reached 94% and 92%,respectively.Compared with the SVM classification of remote sensing feature index time-series images at the pixel scale and the classification results based on the conventional Mean Shift algorithm combined with the SVM algorithm,the OA and Kappa coefficients of this method are improved.Meanwhile,the conventional Mean Shift algorithm can only segment a single band,which is prone to wrong segmentation,while the improved Mean Shift segmentation algorithm can fuse the time series characteristics for multiband image segmentation.Therefore,from the actual classification results,using the improved Mean Shift segmentation method in this paper,not only fewer missed segmentation cases occur,but also can better retain the parcel The edge information of the parcel is better preserved.(3)Through to the commonly used in crop growth period of satellite remote sensing monitoring,the summary of the commonly used prediction model and algorithm analysis,especially it analyzes all kinds of related methods in crop growth period of remote sensing to identify advantages and limitations of existence,in this paper,TCN and LSTM algorithms were selected to carry out the remote sensing recognition experiment of multi-growth stage(seedling stage,tillering stage,jointing and booting stage,full-ear and flowering stage,grain-filling and milk-ripening stage,full-ripening stage and stuffling stage)of all rice types.The results showed that TCN model was better than LSTM model in predicting rice multiple growth period,and its OA accuracy reached 89%,exceeding the latter by 9%.But compared with the classical LSTM model,the TCN model was prone to misclassify the stubble field stage as the nodulation stage.(4)By comparing and analyzing the existing problems of TCN and LSTM models in satellite remote sensing identification of rice growth period at regional scale,this paper proposed a satellite remote sensing monitoring method of rice growth period based on TCNLSTM comprehensive model,that is,on the basis of the application of TCN model,the rice growth period identified as jointing stage was classified by LSTM model.The results of the experiment show that the classification method based on TCN-LSTM integrated model could combine the advantages of TCN and LSTM,and achieve higher accuracy of rice growth period prediction.Compared with the single TCN model,the OA and Kappa coefficients of the integrated model are improved by 7% and 9%,respectively,and could improve the problem that TCN could easily classify the stubble stage into the wrong one when identifying the jointing stage.Compared with the single LSTM model,OA and Kappa coefficients increased by 16%and 20%,respectively,and the recognition performance of tillering stage,filling stage and milk stage was also significantly improved,making CCA coefficients of these three growth stages directly increased by 45%-47%.(5)The comprehensive deep learning model TCN-LSTM proposed in this paper was combined with the reconstruction of time series multispectral images covering the whole growth cycle of major rice types,and a good remote sensing recognition effect of rice multigrowth period was achieved.Among them,this method has the best recognition performance for seedling stage and jointing and booting stage,and its accuracy CCA coefficient is up to 98%and 96%,respectively.The CCA coefficients were 87%,92% and 84%,respectively.The identification performance of the flowering stage of the aligned panicle was average,and the CCA coefficient was about 76%.The recognition performance to tillering stage was the worst,the CCA coefficient was only 59%,which requires further methodological improvement in the future.(6)In the study area,there was not only obvious spatial differentiation in rice cultivation,but also obvious inconsistency in rice growth period.Early and middle rice in one season was mainly distributed in the hilly area of Shayang County.Most of the first-season early rice was distributed in the southeast of Shayang County,near the Changhu Lake.The traditional planting mode of mid-season rice is widely planted in Jiangling County,Xiantao City,Tianmen City and the north of Qianjiang City.In recent years,shrimp rice,which has developed rapidly,has been cultivated on a large scale in Qianjiang City,Shashi District,Jiangling County(mainly Sanhu farm)and Jianli city,where the main canal of the Four Lakes flows.In addition,because the research area of rice planting time is not unified and so on,result in rice growth period around the process changes exist significant difference,at the end of July,for example,in the north of Jiangling,Xiantao,Qianjiang,etc to a wide range of conventional rice planted in entering the jointing-booting stage,and in places such as Qianjiang south and east of the gong ’an county,most shrimp farms middle-season rice is still in the late tillering. |
PDF Full Text Request