Research On Winter Wheat Distribution Information Extraction Method Based On Improved Deeplabv3+

Winter wheat is one of the important food crops in northern China.Timely and accurate access to winter wheat distribution information is of great significance for formulating relevant agricultural policies,optimizing winter wheat planting and distribution areas,and food security in China.Gaogao-2 is the first civil sub meter high-resolution remote sensing satellite in China,which provides high-quality data support for agricultural remote sensing research in China.At present,traditional remote sensing technology is still the most commonly used means to obtain winter wheat distribution information,but it has the problems of low extraction accuracy and insufficient automation.Therefore,this paper takes Daiyue district,Tai’an City,Shandong Province as the research area,and gaofen-2remote sensing image as the data source,and proposes a winter wheat distribution information extraction method based on improved deeplabv3+.The main research contents and conclusions of this paper are as follows:(1)In view of the small internal gap of winter wheat and the difficulty of boundary pixel classification in remote sensing images,the original Deeplabv3+ model is improved as follows: the attention mechanism is introduced to effectively restore the edge detail features;Mobilenetv2 backbone is used to extract features to reduce over fitting and lightweight the network at the same time;Replace the normalization method to reduce the impact of batch size on normalization effect.The Deeplabv3+ model before and after the improvement is trained on the winter wheat extraction data set.The extraction accuracy on the test set is96.03%,which is 1.78% higher than that before the improvement.(2)Aiming at the improved Deeplabv3+ model in this paper,the data set is extracted by winter wheat to verify the effectiveness of the improvement.The experimental results show that using Mobilenetv2 as the backbone network,the single round iteration time is reduced from 947 s to 462 s,the single sheet prediction time is reduced from 156 ms to 35 ms,and the parameter quantity is reduced from 209 mb to 22.3mb,which can effectively alleviate the over fitting problem,reduce the dependence of model training on hardware conditions,and significantly improve the performance of the model;According to the results of ablation experiment,using only Mobilenetv2 backbone network,although the average intersection to union ratio decreased by 0.43%,it increased by 2.38% and 0.91%after adding attention mechanism;After normalization,the average cross merge ratio of the addition group was increased by 0.62%.The above results proved the effectiveness of the improvement.(3)Two regions with concentrated distribution and large proportion of winter wheat and scattered distribution and small proportion of winter wheat are selected,and the method proposed in this paper is compared with other deep learning and traditional remote sensing methods.The results showed that the accuracy of the method of extracting winter wheat distribution information based on deeplabv3+ proposed in this paper reached 96.55% and97.27% respectively;The accuracy rates of unsupervised classification,supervised classification and UNET were 85.53% and 87.71% respectively;90.51 % 、 92.78 %;93.92%、96.82%。 It can be seen that the accuracy of the method proposed in this paper is the highest,and the accuracy difference between the two images is the smallest,which proves the superiority of the method proposed in this paper.(4)The improved Deeplbav3+ method is used to extract winter wheat in Xiazhang Town,Daiyue district and Culai Town,Cuwen district.Combined with the field verification points,the extraction accuracy is calculated,and the extraction accuracy is 93.6% and92.2% respectively;At the same time,according to the extracted spatial distribution information of Winter Wheat in Xiazhang Town,the planting area of Winter Wheat in Xiazhang town in 2020 is 42200 mu,which is only 2800 mu different from the 45000 mu provided by Daiyue district government,and the relative error is 6.22%,it proves that the method proposed in this paper has practical significance.