Research on the technology of terrain classification and change detection based on deep learning

With the rapid development of urbanization in China, the change of land property is very rapid. The detection of land change is helpful to the government’s supervision and the development and protection of land resources. The traditional methods of land feature classification and change detection of remote sensing image generally adopt the methods of manual annotation and recognition. Due to the performance defects, they can not meet the requirements of remote sensing image business application. In this paper, the technology of satellite remote sensing image feature classification and change detection based on deep learning is proposed, and several demonstration cases of automatic classification and change detection system of buildings, roads, water bodies, forest land and other features based on high-resolution remote sensing image are formed. Relevant research results show that the technology can meet the requirements of remote sensing image classification and change detection, and the accuracy of the review is more than 90%.The method adopted can effectively solve the application needs of related fields.


Introduction
China is in the process of high-speed urbanization, the change of land attributes is very rapid, the detection of land change is conductive to the government's supervision and policy-making, and is conductive to the development and protection of land and resources. And the classification of land features is conductive to the overall planning of urban and rural areas, so as to promote the process of urbanization.
The traditional remote sensing image classification and change detection generally use the methods of manual annotation and recognition, which is time-consuming and laborious. Although machine learning method can complete the function of remote sensing image feature labeling, there are still performance defects, which can not meet the business application requirements of remote sensing image. With the rapid development of deep learning technology, significant breakthroughs have been made in the fields of automatic recognition, labeling, change detection of remote sensing images, which makes it possible to meet the application requirements of automatic production of remote sensing image products.
In order to effectively improve the utilization rate of remote sensing image, automatically mine the deep information of remote sensing image, and provide high time effective data support for land monitoring, urban planning, surveying and mapping and other fileds. In this paper, the technology of feature classification and change detection based on in-depth learning is proposed, and the highresolution remote sensing image is used to make a demonstration application case for feature classification and change detection to verify the correctness and effectiveness of the technology. In order to reduce the labor cost, improve the automatic degree of remote sensing image information

Deeplab Model
The establishment of Deeplab model is based on Fully convolutional networks (FCN). FCN can input any size of image, and the output of each layer is also an image, which is conductive to the global understanding of the image. Ordinary CNN can only input images of a specified size, and the output of full connection layer is vector. The core idea of FCN is to replace the full connection layer of the traditional CNN network with a convolution layer whose convolution kernel is 1x1,so that all the neuron layers with parameters in the network are convolution layers, so as to allow the input of images of any size. If the output is a two-dimensional image, the spatial information of the input image will be well preserved, while CNN of image level classification completely loses the spatial information of the image.
The results of FCNet not only identify the target, but also show the shape, size and location of the target. Ordinary CNN can only give the category of image. Because FCNet does not need to traverse the whole image, it is much faster than the DCNN algorithm applied in scene annotation.
In this paper, the deeplab model is used to improve the accuracy of FCN by using the improved full convolution neural network and the full connection conditional random field algorithm. Except for CRF, the structure of deeplab is basically the same as that of FCNet, which is improved on the basis of VGG-Net: Increase output heat map size. Because there are 32 times down sampling in FCN model, even if 100 pixels wide filling area is added around the image during input, the input is 500x500, and the size of the output heat map is only 16x16, so the spatial information of the image is seriously lost. In deeplab, the step size of the 4th and 5th lower sampling layer in FCN is changed from 2 to 1, and the width of filling operation is 1, which makes the lower sampling multiple of the whole model reduce to 8 times [3] . Thus, when the input is 500x500, the output is 65x65. This can reduce the loss of spatial information and improve the accuracy of the model.

RES-FCN Model
In order to further improve the expressiveness and detection performace of the network, this paper further studies the idea of ResNet. ResNet requires each layer of network to directly learn its corresponding mapping relationship, while in the residual network, each layer needs to learn its "residual" mapping. Through formula expression, if the corresponging ideal mapping relationship is H(x),the residual mapping F(x) that each layer of network needs to learn now is H(x)-x, and then the output of each layer of network is remapped to F(x)+x. This method comes from "shortcut mapping". Through shortcut mapping, some of the data can directly reach the next layer in the process of each layer transmission, which can effectively alleviate the problem of gradient disappearance/explosion.
According to the idea of residual network, the original convolution layer of the original deeplab network is replaced by the 50 layer network structure of residual network, and the network structure is adjusted according to the characteristic of remote sensing image. In the experimental comparison, the residual network has the following performance: (1) Convergence performance: after a certain number of layers are added to the residual network, the performance is improved to a certain extent. The residual network has a lower convergence loss and does not produce too much over fitting.
(2) Classification performance: as the number of layers of residual network is generally higher than that of previous models, and there is a residual structure as the support premise of extreme depth, the performance of residual network is generally higher than that of previous excellent models, and the performance of residual network increases with the increase of network layers [4] .
The full convolution neural network model used in this paper is shown in The specific processing flow of image segmentation algorithm based on Res-FCN is as follows: firstly, the input remote sensing image gets the thermal map through Res-FCN network, that is, each layer of image represents a category, and the pixel value represents the confidence degree of the location belonging to the category. Then, the thermal graph is input into the all connected random field, and the spatial relation of the original image is used to post process the thermal graph, which can supplement the spatial relation for the segmentation result of neural network and improve the accuracy of segmentation. Finally, the result of conditional random field processing is selected by maximum value and bilinear interpolation, and the corresponding annotation result of the original image is obtained.

Adaptive joint loss function
At present, in the field of deep learning, the loss functions commonly used are Softmax loss and Euclidean loss. Softmax loss is often used in the field of image classification. It provides a smooth and differentiable operation to get the maximum value. It can be seen as an extension of logistic regression. Its function form is as follows: ∑ y log p ( 2 ) Where o is the output of neural network, y is the label of sample, L is Softmax loss. For regression problems, the Euclidean distance loss function, Euclidean loss, is often used. Its function form is as follows: L ‖ ‖ ( 3 ) As you can see, Softmax loss pays more attention to the relative size of the output value, that is, the maximum value. In this mode, the optimization direction of neural network is to increase the output of target class position and reduce the output of non target class position. Euclidean loss is more concerned with the absolute value of the output. In this mode, the optimization direction of neural network is to make the output and label as consistent as possible.
For image classification, obviously Softmax loss is more suitable, because if the output of neural network is correct, but the absolute value is larger, Euclidean loss will punish it, thus reducing the classification ability of neural network. FCN algorithm transforms image segmentation problem into pixel by pixel classification problem, and achieve good results. Considering that Euclidean loss can learn the continuity information of tag image, we propose Self-adaption joint loss function(SAJ loss)to integrate the two loss functions mentioned above. SAJ loss combines Softmax loss and Euclidean loss, so that the neural network can learn not only the spatial relationship of the input image, but also the spatial relationship of the label image. At the same time, the weight relationship between two is obtained through optimization learning, which is more conductive to training to get better network parameters. The equation of adaptive joint loss function is as follows:   Through the above experimental results, it can be seen that the target of Softmax loss has a low separation from the background, and it is easy to confuse the background and the target, thus reducing the generalization ability of the network. SAJ loss can effectively separate the target from the background, and keep the spatial relationship of the label image. In addition, because the full connection condition random field algorithm takes a long time, while SAJ loss only slightly increases the training time of the model, not the testing time, so SAJ loss can replace the full connection condition random field algorithm to some extent, and further increase the accuracy of the image model.

Research on Algorithm of land change detection module
The remote sensing data contains a large amount of information, but the change of ground features is always slow. For multitemporal remote sensing data, it contains a large number of redundant information, which is not conductive to the development of related monitoring work. In order to solve this problem, by introducing the change detection function, the intelligent extraction results of remote sensing images are processed twice to obtain more valuable change information [5] .
The function of land change detection requires high registration accuracy of the front and back phases, so it is necessary to input the remote sensing image extraction results after RPC information correction. Because the multitemporal remote sensing image is often not completely coincident, it is necessary to extract the coincident region from the input image according to its geographic information. After getting the overlapped area, according to the location of the overlapped area, the results of intelligent extraction of time-phase remote sensing images before and after comparision are ortained. If it changes, the pixel is retained. In addition, in order to further improve the accuracy of change extraction, the input image needs to be corroded and expanded first, and the extraction results also need to be processed later. Through the connectivity domain analysis algorithm, the change patterns with an area of less than 2 meters are removed. The flow chart of this function is shown in  Remote sensing image change detection is mainly to monitor the changes between two or more remote sensing images in the same geographical and different periods. The classic remote sensing image change detection method is based on the significance test. Let I x, y and I x, y represent the two remote sensing images before and after the change,let D x, y I x, y I x, y ,in an ideal case,the region with D x, y value of 0 is the region without change,and the region without D x, y value is the region with change. Due to the interference of the equipment imaging noise, D x, y is composed of two parts:the real change of the region and the noise. Generally, it is assumed that the noise obeys the Gaussian distribution, and its mean value and variance can be estimated from the difference graph. In most cases, the change area only accounts for a small part of the whole image, and change detection can be considered as a signal detection problem in noise. Assuming that H represents that the pixel has not changed, and H represents that the pixel(block)has changed,the probability density functions of the unchanged region and the changed region are p D x, y |H ,p D x, y |H respectively. Under the assumption that the noise is Gaussian distribution, the probability density function of the difference graph is: The basic idea of significance test is to determine the threshold according to the probability distribution and false alarm rate, and judge whether the difference image belongs to the changing area or the non changing area by the threshold. Given the false alarm probability α ,the ： , α.
By calculating the exit limit τ , μ σ z ，you can get the final output： , 0 , , 1 , ℎ The region of 1 in the output image Out x, y is the real changed region, and the region of 0 in the output image Out x, y is the non changed region. However, this method is often affected by light, seasons and sensors, so it is difficult to obtain two images that fully meet the requirements. The deep learning method can overcome these effects, so we propose a change detection method based on deep learning and significance test. This method first marks the multi temporal remote sensing image pixel by pixel, then subtracts the labeling results and detects the connected domain. Finally, it uses the significance test to screen the connected domain, and finally retains the larger connected domain, which is considered as the changed region. The flow chart of this method is shown in Fig5.

Change Detection module test result
Through the time-phase pair of the original remote sensing image as the test case, the change detection module test is carried out for the, building, water, road, forest and other types of features, represented by the road. The front time phase diagram and the back time phase diagram of the change detection are shown in Fig 8 and Fig 9,respectively,with the yellow edge as new and the magentaedge as less.

Conclusion
In this paper, the technical ability of satellite remote sensing image classification and change detection based on deep learning is established, and several demonstration cases based on high-resolution image are formed. Relevant research results show that the technology can meet the requirements of remote sensing image classification and change detection, with an average accuracy of more than 90%, and the processing results meet the accuracy requirements. The method used can effectively solve the application requirements of related fields, especially in the face of massive image processing and realtime monitoring applicaitions, this research method has the advantages of high efficiency and high precision, and good versatility, which is convenient for replication and promotion to other regions.