Research On New Building Identification Model Based On Remote Sensing Image

One of the most important tasks in the land monitoring business is to supervise the construction,demolition,alteration and expansion of buildings on the ground.When a plot exists in a building without approval,the government department needs to send someone to investigate whether there is illegal land occupation.For large cities and their suburbs,it is difficult for staff to patrol and supervise the city every day.Today’s technology can rely on high-resolution satellite remote sensing images and deep learning algorithms to innovate existing workflows.The identification of new buildings is different from the image difference detection.Many of the areas in the image change are not new building areas,so it is difficult to solve only relying on traditional image processing algorithms.The emergence of deep learning methods based on full convolutional neural networks Image-level segmentation of images is achieved to solve this semantic level of image recognition segmentation.This paper first constructs a two-phase difference map and adds new building annotations in the difference map as training data,and enters the Deep Lab network training to obtain a deep learning model for identifying new buildings.The author of the sampling process in the second half of the network is used.Both the bilinear difference method and the convolutional layer are connected to the BN layer as an improved optimization.Experiments show that the trained model has the ability to recognize new buildings but is affected by the difference map blur effect.The author also applies the remote sensing image.The building directly marks and trains the building identification model.The test data for 2015 and 2017 are used to identify the building area using the building identification model,and then the difference between the two-year identification results is obtained to obtain the newly added building area.In the training aspect,the author made a structural improvement based on U-net,and introduced a multi-layer cavity convolution layer after convolution sampling.The model obtained after training has a good recognition effect on the building area and has more on the newly added building area.Good recognition effect,its model recognition ability is obtained through the evaluation of f1 value.Finally,the author reconstructs the far-infrared channel data into the new three-channel data,and uses two network model hybrids to train and optimize the training deep learning model to realize the identification of new buildings.On the loss function,the author uses the cross-entropy function plus the L2 regular combination loss function,and the author has done a lot of data enhancement processing to provide more rich features for the model training data.Experiments show that more complex new buildings can be identified more accurately through a hybrid learning approach.This paper realizes the identification of new buildings by using deep learning methods combined with the characteristics of remote sensing images.Several research improvements have been made to the structure of the neural network in the deep learning method,and the parameter training optimization has been adjusted,and finally a model that can effectively identify the newly added building is obtained.