Research On Paddy Field Segmentation Method Of UAV Low Altitude Remote Sensing Image Based On Deep Learning

Obtaining accurate farmland boundary information is an important foundation for precision agriculture practice,and is important for constructing high-precision farmland maps,guiding intelligent farming machines to implement autonomous field operations and accurately measuring farmland areas.This paper addresses the problem of complex farming environment in southern hilly areas and the low accuracy of extracting paddy field plots from UAV low-altitude remote sensing images,and uses DeepLabV3+full convolutional neural network to achieve accurate segmentation and extraction of paddy field plots using UAV aerial images of paddy fields.The main research contents and conclusions are as follows.1)To address the lack of a large number of paddy field samples in the existing low-altitude remote sensing dataset,the DJI Phantom 4-RTK small UAV was used to obtain orthophotos of the paddy field.The paddy field image data were enhanced by using three image geometric transformation methods and adjusting the brightness and contrast to increase the number of paddy field image samples to five times the original sample size in order to avoid over-fitting the model.Sample labels were created by referring to the POSCAL VOC 2012 dataset format,and pixel-level labelling of the paddy plots was achieved using Labelme annotation software,and semantic label maps were generated.Finally,the data format and labels were adjusted to construct a semantic segmentation dataset covering four types of paddy field scenes.2)A semantic segmentation model for low-altitude paddy fields based on improved DeepLabV3+.By analysing the shortcomings of DeepLabV3+ in the segmentation results of the paddy field dataset in this paper,corresponding optimisations and improvements are proposed to the network structure.The number of branches of the ASPP module parallel atrous convolution is increased in the encoder,and a combination of smaller-scale dilation coefficients is set for it to improve the extraction capability of small-scale features;the shallow feature maps with a scale of 1/8 of the input image are added to the decoder for combined upsampling,and depthwise separable convolution is used to decouple the image depth and spatial information.Comparative experiments were conducted using the improved model(D2-DeepLabV3+)and the original model,UNet,Seg Net and PSPNet models.The results showed that the mean MIo U and F1 score of D2-DeepLabV3+ in four types of paddy field scenes were 82.96% and 90.89% respectively,which were improved by the D2-DeepLabV3+ model generally outperformed the other segmentation models and achieved better segmentation results.3)A semantic segmentation model for low-altitude paddy fields combining attention mechanism and adaptive spatial feature fusion.In response to the problems that the D2-DeepLabV3+ model has low accuracy in segmenting thin and narrow field monopolies and mis-segmentation in some regions,a second generation improved DeepLabV3+ paddy field segmentation model is proposed.The attention mechanism sc SE module is incorporated into the backbone network of the encoder to enhance the representation of paddy field features by remapping the feature map;the adaptive weight coefficients are obtained by using the ASFF algorithm in the decoder to solve the problem of inadequate multi-scale feature fusion in the upsampling process.The optimal combination of dilation rates is determined by comparing the segmentation effect of the model with different ASPP dilation rates to enhance the model’s ability to extract fine features of paddy fields.The segmentation results of this model(SA-DeepLabV3+)were compared with D2-DeepLabV3+ and UNet,Seg Net and PSPNet models,and the results showed that the index parameters and segmentation effects of the SA-DeepLabV3+ model were effectively improved in four types of paddy field scenes,in which the mean MIo U and F1 score were improved to 85.56% and 92.22% respectively,with better segmentation integrity for paddy plots and thin and narrow fields,and the model performed optimally.In summary,this paper takes UAV low-altitude remote sensing images of paddy fields as the research object,and constructs a semantic segmentation paddy field dataset by means of image enhancement techniques and manual labeling.Based on the DeepLabV3+network structure for improvement and optimisation,two generations of paddy field segmentation models,D2-DeepLabV3+ and SA-DeepLabV3+,were successively proposed.The comparative experimental results show that both models can segment and extract paddy plots better,among which the SA-DeepLabV3+ paddy field segmentation model with the introduction of sc SE module and ASFF algorithm has better robustness and stronger adaptability.The research results can provide an important basis for further acquiring high-precision paddy field boundary positioning information and constructing high-precision maps of several paddy fields in larger areas,which can play a positive role in promoting efficient and accurate information management of paddy field cultivation.