Research On Semantic Segmentation Algorithm Of Crops Based On UAV Remote Sensing Images

The area measurement of agricultural land will affect the income level of farmers,and at the same time,it also plays a certain role in influencing social and economic development.The traditional method of measuring the ground needs to rely on high labor costs,and cannot respond quickly to changes in the types of crops.Semantic segmentation technology based on deep learning is an effective method for solving area measurements.The use of semantic segmentation technology can achieve end-to-end division of remote sensing image plots,which can reduce labor costs and improve measurement accuracy,so it has an important application in the value of ground surveys.The specific process of using drone remote sensing images for ground measurement is to use drones to take aerial photography to obtain images of the area to be measured,and then use convolutional neural networks to find deep features in the images that represent different crops and obtain segmentation results,and then compare the algorithm Carry out optimization iterations to meet the specific needs of crop segmentation in some special areas.Based on the data set provided by the government of Xingren County,Guangxi,this article reasonably segmented the original data and corrected some of the incorrectly labeled parts,and established a data set suitable for segmentation tasks.Researched various feature extraction networks including ResNet,and finally selected Xception as the feature extraction network for segmentation.The IM-DeepLabv3+model structure is proposed.This structure separates the segmentation process of civil buildings by introducing UNet as an auxiliary segmentation network to improve the segmentation effect of civil buildings.At the same time,the edge segmentation effect is optimized by combining the label smoothing strategy and completed ablation experiment.In the fine-grained segmentation,DeeepLabv3+is selected as the baseline of the segmentation network,and the performance of the model is more stable by optimizing the loss function.Finally,the overall design of the algorithm flow is completed.At the image input end,the pre-processing is done considering the influence of the lighting in the real environment,and the morphological post-processing is done at the output end of the image according to the general conditions of the real environment,and used in the prediction stage.The prediction results were further optimized by the method of expansion prediction and up-down and left-right turning and averaging,made performance superior to other methods.