Research On Field Weed Recognition And Semantic Segmentation Based On Remote Sensing Images

Weed identification and segmentation are crucial for accurate weed removal in the field.Crop cultivation in China presents the characteristics of wide area,large variation and seasonal change,and crops are extremely seriously affected by weed damage.The traditional weeding method is by manual weeding,but manual weeding is inefficient and labor-intensive,and is gradually replaced by other weeding methods.Pesticide weeding is the most common method,but it cannot be applied locally to weed areas,which increases production costs and pollutes the environment,and food safety can be affected.In order to achieve green agriculture and precision agriculture in agricultural environment,weed identification and segmentation to locate weed areas in the field has become the primary prerequisite for accurate weed removal in the field.The common method of field weed detection is to collect farmland images by low altitude drones,and then use machine vision technology to identify and segment weeds,however,the images taken by drones are only suitable for small area field weed detection,and machine vision technology is mostly used to extract shallow features from small batch image data.Aiming at the problem of large and extensive farming areas and complex agricultural environment in China,this paper takes remote sensing images of common land cover scenes,abnormal scenes among farmlands and field weeds as the research objects,and combines the features of wide coverage area and fast acquisition speed of aerial and satellite remote sensing images to carry out the research of field weed recognition and semantic segmentation based on remote sensing images.In order to extract deep features of remote sensing images and improve the accuracy of field weed recognition and semantic segmentation,we use convolutional neural network to recognize and semantic segment remote sensing images,and then identify agricultural scenes from common land cover scenes,and further segment the location of field weeds and other abnormal areas in fields from agricultural scenes,and add feature visualization algorithm to enhance the interpretability of convolutional neural network and improve the feature The feature visualization algorithm is added to enhance the interpretability of the convolutional neural network and improve the feature representation capability.In order to demonstrate the effectiveness of the algorithms in this paper,a weed detection system based on remote sensing images is designed and implemented.The main research contents and conclusions are as follows.(1)A remote sensing image scene recognition algorithm based on RCF(Res Net50-CBAM-FCAM)network is proposed to address the problem of similarity of scenes between categories and large differences of scenes within categories in remote sensing images,and the limited feature extraction ability of convolutional neural networks.The algorithm adds the Convolutional Block Attention Module(CBAM)and Full Convolutional-Class Activation Mapping(FCAM)branches to Res Net50,and uses the The attention mechanism is used to fuse the branch features with the extracted channel attention features and spatial attention features,respectively,to improve the extraction capability of Res Net50 for target object features of remote sensing scene images.It is experimentally demonstrated that the algorithm achieves 94.08% and 96.64% overall recognition accuracy with training sets of 20% and 50% of the total samples on dataset AID;90.63% and 93.38% overall recognition accuracy with training sets of 10% and 20% of the total samples on dataset NWPU-RESISC45;90.63% and 93.38% overall recognition accuracy with training sets of 10% and 20% of the total samples on dataset Agriculture-Vision,the training set is 80% of the total samples,and the overall recognition accuracy reaches93.42%.(2)In order to solve the problems that the target objects of agricultural remote sensing images are of different sizes,the distribution areas are many and scattered,and the semantic pairs are not aligned when the features of the semantic segmentation network are fused,a remote sensing image semantic segmentation algorithm based on FAm F(FPN-ASPP-m FAM)network is proposed.The algorithm fuses RGB and NIR remote sensing image data,makes full use of farmland planting feature information,replaces ordinary convolution with empty convolution in the feature extraction stage of FPN(Feature Pyramid Networks,FPN),and introduces the empty convolution feature pyramid module(Atrous Spatial Pyramid Pooling(ASPP)to extract multi-scale high-level semantic feature information and global location information of target objects;add FCAM branches and fuse FCAM branch features with backbone features through the attention mechanism;in the upsampling stage,use the improved multi-scale Flow Alignment Module(muti-FAM)to replace the upsampling operation.In the upsampling stage,the multi-scale Flow Alignment Module(muti-FAM)is used to replace the upsampling operation to solve the problem of semantic misalignment during feature fusion.It is proved that the algorithm can achieve 92.48% and 54.90%of the mean pixel accuracy and mean intersection ratio on the dataset Agriculture-Vision with a training set of 80% of the total samples,and 93.93% and90.50% of the mean pixel accuracy and mean intersection ratio on the dataset WEED with a training set of 80% of the total samples.(3)In order to explain the operation mechanism of RCF network and FAm F network,the feature map visualization algorithm based on FCAM network is proposed.The output layer is constructed based on CAM,the fully connected layer of the recognition network is replaced by a convolutional layer and a global average pooling(GAP)layer,and the codec sampling structure is increased to visualize the feature map while the network is trained,and the feature representation capability of the target object is increased during feature extraction.The experimental results show that the method can better focus on the target region in the image and make a reasonable explanation for the convolutional neural network to complete the recognition and semantic segmentation tasks.(4)In order to visualize the effect of RCF network and FAm F network,a weed detection system based on remote sensing images was designed and implemented.The system implements registration and login,basic image data management,weed recognition and weed segmentation functions.The remote sensing image scene recognition algorithm based on RCF network and the remote sensing image semantic segmentation algorithm based on FAm F network are applied to the system weed recognition and segmentation functions respectively by using Flask framework to support the online completion of weed recognition and segmentation tasks.After system testing,the system can improve the efficiency of weed detection and detect weeds more conveniently.