| The development of high-resolution remote sensing satellite technology has played a huge role in promoting the level of high-resolution earth observation in China.The launch and application of high-resolution satellite series help China obtain higher precision data sources in resource monitoring,agricultural yield prediction,disaster management and other aspects.One of the most widely used aspects in remote sensing image is crop agricultural management.High resolution remote sensing satellite data can provide high-precision crop information,which play a huge role in promoting crop monitoring,disease management,yield evaluation and so on.At present,there are many algorithms for crop extraction from remote sensing images,but the traditional crop extraction methods not only have limitations,but also become more and more difficult in processing massive remote sensing information.Remote sensing images processed by empirical and deterministic methods can help to predict the current situation of various crops in the whole growing season.It is an important task to accurately identify cotton fields from remote sensing images in precision agriculture.Based on this,this study uses the framework of deep learning to identify and estimate the cotton yield in the Weiku oasis of China using a variety of satellite remote sensing images.Firstly,the optimized pixel level multi-dimensional dense connected convolutional neural network(Dense Net)is used to identify the cotton field in Weigan River Basin Oasis.In addition,it is compared and evaluated with four widely used classical convolutional neural networks(CNN),including Res Net,VGG,Seg Net and Deep Lab v3+.The results show that compared with other models,Dense Net can recognize the characteristics of cotton crops in a short time and converge at a faster speed.Several indexes(P,F1,R and m Io U)generated by confusion matrix were used to test the performance of the model,and then the identified cotton fields were visualized.Compared with the mainstream models widely used before,the recognition effect of the optimized densenet model is significantly improved.F1 is 0.953,m Io U is 0.911.In addition,it can also distinguish cotton fields from clouds,mountain shadows and towns.Then,based on the improved Dense Net algorithm,the cotton field recognition of multi-source and multi temporal remote sensing images is carried out,and the cotton field planting area in the Wei Ku Delta region is extracted to realize the rapid monitoring of cotton field area change.The multi-source remote sensing image mainly considers the data of gaofen-1 and Landsat.Based on the satellite image of GF-1,the change of cotton planting area from 2015 to 2018 was analyzed and compared with the official statistical data.Due to the limited data of GF-1,the change of cotton planting area from 1993 to 2018 was analyzed based on Landsat satellite images,and compared with the official statistical results.It is proved that the limited remote sensing data can still be used to analyze the temporal and spatial changes of cotton field area.Finally,according to the phenological characteristics of cotton,the meteorological data and NDVI data which affect the cotton yield in the growth period were selected,and the data were modeled with the county cotton yield data to estimate the cotton yield.We build the convgru neural network which can fuse the temporal and spatial characteristics and deal with the long time series big data,and compare it with the neural network fitting machine which deals with a small amount of data.The results showed R~2is more than 0.9.Finally,we used Conv GRU model to predict the yield of three counties(Kuche County,Xinhe County and Shaya county)in the Weigan River Basin Oasis. |
PDF Full Text Request