Information Acquisition Of Sunflower Farmland Based On UAV Multispectral Remote Sensing Images

Sunflower is an important oil-bearing economic crop,which can be planted in high salinity and alkalization zone.Accurate management of sunflower is the basis of large-scale accurate identification of sunflower agricultural conditions,which has a huge impact on the sunflower industry.Rapid mapping of sunflower planting area can provide planting distribution information for yield estimation and automatic operation.The recognition of sunflower growth period can help farmers to fertilize and irrigate more accurately and timely.In addition,under severe weather conditions such as strong wind,sunflower is prone to lodging,which seriously affects its yield and quality.The rapid identification of sunflower lodging can provide help for subsequent management and claim settlement.However,the traditional method of obtaining sunflower field information mainly relies on artificial field investigation,which is time-consuming and inefficient.In recent years,with the development of deep learning and the application of UAV Remote Sensing Technology in the field of agriculture,it provides a platform and solution for sunflower automatic and rapid agricultural information acquisition.Taking the UAV multi spectral remote sensing image of Shahao canal irrigation area in Hetao Irrigation District of Inner Mongolia as the research object from 2018 to 2019,combined with deep learning algorithm and UAV remote sensing technology,this study solved the research of three key aspects in the field of sunflower agricultural acquisition,mapping sunflower,sunflower period recognition and sunflower lodging monitoring.The main research contents and conclusions are as follow.(1)Based on DeepLab V3 + deep semantic segmentation network,this paper proposes a farmland crop classification method for UAV multispectral remote sensing image.By modifying the input layer structure,fusing multispectral information and vegetation index feature map,and using Swish activation function is used to replace ReLU activation function to optimize the model.We build and train the model on the remote sensing image of 2018,and test its generalization performance of on the image of 2019.The results show that the mean pixel accuracy and the mean intersection over union of the improved DeepLab V3+ model are 93.06% and 87.12%,respectively,Compared with SVM model,it can be improved by 17.75% and 20.8%,respectively.And it can be improved by 2.56% and2.85% compared to DeepLab V3+,respectively.This method obtains the best effect and has fast speed,which provides a simple and effective scheme for using UAV multispectral remote sensing image to interpret the farmland planting.(2)Based on UAV multispectral remote sensing image,a plot level Sunflower Growth Period recognition scheme was proposed.By calculating the ROI region of pixel level depth semantic segmentation model,the field parcel level recognition results that meet the actual needs are obtained.To solve the error caused by the difficulty of labeling the growth period of sunflower at pixel level,this study proposed a weight loss function for the difference between species,which gives different weights to different types of classification error,and effectively solves the problem of low recognition accuracy of sunflower growth period.The experimental results show that the PSPNet model based on deep semantic segmentation combined with the loss function proposed in this paper achieves the best recognition effect.The accuracy of Sunflower Growth Period recognition in the test set can reach 89.02%,which shows that the proposed scheme has good recognition accuracy,and provides an effective solution for sunflower growth period recognition based on remote sensing image.(3)Based on UAV remote sensing image,a sunflower lodging recognition method based on image fusion and depth semantic segmentation is proposed.Firstly,the multispectral image with low resolution is transformed into multispectral remote sensing image with high resolution by matching the multispectral image and the visible image.Then,the visible light image and the transformed multispectral image are fused to obtain high-quality multispectral image with rich spectral information and high spatial resolution.Thirdly,the Segnet depth semantic segmentation model is improved,and the sunflower lodging recognition model is effectively improved by using the methods of jump layer connection,conditional random field and depth separable convolution.Experiments show that the recognition accuracy of sunflower lodging based on image fusion is higher than that of non-image fusion in support vector machine,full convolution neural network(FCN),Segnet and improved Segnet.At the same time,the performance of deep semantic segmentation method is always better than that of SVM method.The recognition accuracy of the improved Segnet is 84.4% and 89.8% respectively for the image fusion group,and76.6% and 83.3% respectively for the non image fusion test group and the image fusion test group of No.2 test field generalization performance test set.This research method combines the advantages of image fusion and deep semantic segmentation,and provides a useful reference for sunflower lodging monitoring.