Research On Extraction Methods Of Agricultural Land Based On High-resolution Remote Sensing Images

Agricultural land resources are national strategic resources and the material basis for the survival and development of the people.Accurate perception of agricultural land information i s a key of realizing effective protection of agricultural land resources.High-resolution data co ntains richer feature information,clearer neighborhood spatial relationship information,and m ore complex spectral feature information,providing development opportunities for automatic classification of agricultural land.In the practical application of agricultural land classificatio n,traditional visual interpretation methods,classification methods based on statistical analysis,and machine learning classification methods have been widely used,and are gradually develo ping towards a higher level of intelligence.This paper focuses on the classification of agricult ural land,combining the characteristics of high-resolution satellite imagery with the requireme nts of agricultural land classification,and studies the semantic segmentation method,feature e xtraction method and classification method of CNN.Based on the disentangled self-attention model,the key technology of high-resolution satellite image agricultural land classification is optimized.Combined with the OHEM method for uneven samples in the Object-detection fiel d,an improved DNL semantic segmentation model with excellent deep learning performance i s constructed.The R-DNL network finally realizes the extraction and fusion of agricultural lan d classification features in complex scenarios,and improves the accuracy and efficiency of ag ricultural land classification.Following is the conclusions can be drawn:(1)Experiments verify that the classic U-Net network and DNL network in the field of d eep learning can be successfully transferred to this research,which is the semantic segmentati on research of agricultural land cover classification based on high-resolution remote sensing i mages.Among them,the model obtained by DNL network training is better than the traditiona l classic semantic segmentation U-Net network in the classification of each category.It is veri fied that the decoupled self-attention model proposed by the DNL network can make it have h igher feature extraction efficiency and feature extraction effect than traditional semantic segm entation networks,achieves 2.46% and 3.11% accuracy improvements in overall accuracy(O A)and average intersection ratio(mIoU),respectively.(2)The main idea of the R-DNL network proposed in this paper is to improve the feature extraction effect in the DNL network to realize more in-depth mining of feature map informat ion,so that the final DNL feature enhancement module of the model can obtain better agricult ural land coverage classification semantic segmentation results.After model training and mod el comparison and analysis,the classification effect of the R-DNL deep self-attention mechani sm model proposed in this paper is significantly better than DNL and the benchmark U-Net ne twork.Specifically,the overall accuracy(OA) and average intersection ratio(mIoU)of the mo del are improved by 5.66% and 8.25%,respectively,compared with the U-Net network.Comp ared with the DNL network,the accuracy has been improved by 3.2% and 5.14% respectively.(3)This paper proposes a pixel-based OHEM sample extraction training method,and app lies the pixel-based OHEM method to the DNL network and the R-DNL network to verify the effectiveness of the method.The experimental results show that the pixel-based OHEM meth od proposed in this paper can effectively improve the training effect of the unbalanced sample categories in the model.In the R-DNL network experiment,each category F1-Score achieved0.79%,5.81%,0.80% and respectively.2.29% accuracy improvement;the overall accuracy(OA)and the average intersection ratio(mIoU) increased by 1.31% and 3.13% respectively.