Semantic Segmentation Of Crop Remote Sensing Images Based On Deep Learning

Semantic segmentation technology of remote sensing image plays an important role in crop planning,vegetation detection and agricultural land monitoring.However,there are some problems in the task of crop remote sensing semantic segmentation,such as class imbalance and unclear boundary division,and the high similarity between crop classes and large differences within classes in some samples make the semantic segmentation of crop remote sensing images more challenging.In order to solve these problems,this thesis studies the semantic segmentation of agricultural remote sensing images based on deep learning technology.The main research work of this thesis is as follows:(1)Aiming at the problems of unbalanced categories and unclear boundary division of agricultural remote sensing images,an agricultural remote sensing image segmentation network based on parallel attention(PANet)is proposed.In the data preprocessing stage,data enhancement and expansion are carried out to reduce the gap between the distribution of different types of samples and enrich the diversity of data sets.The network uses the feature pyramid structure to integrate high-order semantic features and low-order spatial information to enhance the processing ability of the network to image details.At the same time,a parallel attention structure is proposed to enhance the output characteristics of pyramid structure.In the parallel attention structure,the coordinated attention module is used to obtain the channel attention features containing location information,so that the network can locate and recognize the target object more accurately.Experiments on the Barley Remote Sensing Dataset released in the Tianchi Agricultural Brain AI Challenge show that the mean intersection over union of PANet is 65.94%,and the segmentation results of crop edge and other features are more accurate than the comparison model.(2)Aiming at the problems of high similarity between crops and large differences within crops and category imbalance in some samples,asymmetric convolution class relation network(A-CRNet)is proposed based on PANet network.The network adopts encoder-decoder architecture.In the encoder,asymmetric convolution is used to enhance the horizontal and vertical features and improve the extraction ability of different dimensional features.Then,a class feature enhancement attention mechanism is proposed,which is composed of channel attention mechanism and spatial attention mechanism to enhance location information.The channel attention mechanism learns different channels,that is,the semantic differences between different crops,and enhances the differentiation of different crops;Strengthen the spatial attention mechanism of location information to obtain the relationship between pixels in each channel,that is,the characteristics of the same crop,so as to reduce the misjudgment of the same crop.Next,the category relationship module using class feature enhancement attention mechanism is introduced to obtain crop category relationships of different scales.In the decoder,the relationship between crop categories is fused to enhance the recognition ability of the network for crop features of different scales and improve the accuracy of crop boundary segmentation.In addition,through data preprocessing,data enhancement and category balance loss function,the problem of category imbalance in crop remote sensing images is further alleviated.Experiments on Barley Remote Sensing Dataset show that the mean intersection over union and overall accuracy of A-CRNet network reach 70.11% and 83.45% respectively.In the complex ground object background of remote sensing image,we can accurately distinguish different similar crops,correctly identify the same crop with large feature differences,and the extracted target boundary is clear.