Research And Application Of Land Type Automatic Recognition Based On Remote Sensing Image

Remote sensing image is an important modern technology in the field of geographical research.It can be used to classify and identify land use types,and the results can be used to grasp the change law of soil and water loss,measure the effect of soil and water conservation,and provide a decision basis for subsequent soil and water conservation and ecological environment protection.With the development of artificial intelligence,the automatic recognition method based on deep learning has been widely used in the field of remote sensing image feature classification.Through the Fully Convolutional Network,a variety of land types can be automatically identified and classified.At present,there are still some difficulties to achieve accurate classification of various land types using deep learning technology,among which there are two kinds of current situation that affect the accuracy of recognition.The first one is the unbalanced distribution of categories,and the sample data of different categories have a big gap;the other one is that the recognition of adjacent boundary is wrong,and the semantic information of the boundary between categories is not clear.In order to make full use of the high-resolution remote sensing image and further improve the recognition accuracy of various land types,in view of the above two phenomena,this paper carried out the related research of land type automatic recognition technology based on remote sensing image.First of all,an online hard sample mining algorithm is applied to solve the problem of unbalanced category distribution in pixel level semantic segmentation.In this paper,the OHEM algorithm is added to three classical semantic segmentation models,and the optimal combination of the model parameters is selected.Experiments prove the effectiveness of OHEM algorithm,which can improve the IoU of category with a small sample size,and the overall MIoU and recognition accuracy of the model are improved.Then,a bi-directional affinity field loss function is proposed to solve the problem of unknown semantic information of adjacent pixels in segmentation and recognition.In this paper,the combined calculation method of the affinity field loss function is improved to matrix vector addition,and the loss function is optimized to bi-directional calculation.Experiments show that Bi-AFF loss function can greatly improve boundary recognition by considering the correlation between pixels fairly and keeping the effective information of region-level supervision to the maximum extent.In addition,the ablation experiment proves that the model is more effective when the OHEM algorithm and the Bi-AFF loss function are combined.Finally,this paper applies the theoretical results of the study to the actual scene,and designs an automatic identification prototype system of various land types.The application of this system tool proves its practicability in the application scene of geographical annotation identification,providing a certain reference value for relevant researches on water and soil conservation.