CN115761513A - Intelligent remote sensing identification method for mountain large landslide based on semi-supervised deep learning - Google Patents
Intelligent remote sensing identification method for mountain large landslide based on semi-supervised deep learning Download PDFInfo
- Publication number
- CN115761513A CN115761513A CN202211584428.6A CN202211584428A CN115761513A CN 115761513 A CN115761513 A CN 115761513A CN 202211584428 A CN202211584428 A CN 202211584428A CN 115761513 A CN115761513 A CN 115761513A
- Authority
- CN
- China
- Prior art keywords
- network model
- training network
- training
- landslide
- student
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A50/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE in human health protection, e.g. against extreme weather
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a mountain large-scale landslide intelligent remote sensing identification method based on semi-supervised deep learning, and belongs to the technical field of surveying and mapping. The invention comprises the following steps: s1: selecting a remote sensing image containing a landslide and digital elevation data of a corresponding area for data synthesis to obtain a data set, and labeling the data to obtain a sample set; s2: establishing a student training network model and a teacher training network model, training and judging a sample set, predicting the sample set to obtain a prediction result, judging the landslide characteristic type of the prediction result and obtaining a training result; s3: and (3) calculating the cross entropy loss value of the training result obtained in the step (S2), calculating the mean square error to obtain an exponential moving average line, updating the weight in the student training network model according to the exponential moving average line, and continuously predicting to perform weighted average to update the weight of the teacher training network model. The method can efficiently and accurately identify the landslide in the remote sensing image under the support of a small amount of samples and limited calculation power.
Description
Technical Field
The invention relates to the technical field of surveying and mapping, in particular to a mountain large-scale landslide intelligent remote sensing identification method based on semi-supervised deep learning.
Background
The wide area of China is boundless, has various landforms and complicated and changeable geographic and geological environments. The overall situation of the southwest region is large in fluctuation, the mountainous region has more landforms, and the situation is particularly the same in the mountainous region of the plateau plain transition region. As the mountainous area is close to the subtropical area, the annual average precipitation is higher, and the precipitation time is longer, geological disasters frequently occur, and the disasters are extremely serious. Wherein, the typical disaster with larger harm and more serious degree belongs to the mountain landslide. In addition, landslide can also damage road traffic, posing a serious threat to critical facilities nearby, such as mountain railroads and other important items.
The frequent occurrence of geological disasters reminds the seriousness of disaster reduction and prevention, and aiming at the situation, a landslide area needs to be mapped during landslide identification, landslide boundaries are distinguished, and the distribution situation of landslides is identified, so that data bases are provided for landslide investigation research, secondary disaster early warning and risk assessment by using the information. The first work of landslide identification is to select an identification method, the landslide identification method is subject to long-term development for years, landslide identification is mostly carried out manually in the early stage, field geological mapping is carried out by using a traditional method, and the obtained result is accurate, time-consuming and labor-consuming. After the development of the aerospace remote sensing technology, a plurality of methods are developed for unmanned aerial vehicle or satellite remote sensing images, and for image interpretation analysis, from visual interpretation to computer interpretation, all landslide identification methods need to consider the identification accuracy. At present, most methods have accuracy reaching a certain degree, and how to improve efficiency and quickly identify a large amount of data while ensuring accuracy becomes a research hotspot. The conventional landslide identification mainly adopts machine learning algorithm classification, and is difficult to cope with complex landslides with high identification difficulty.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a mountain area large-scale landslide intelligent remote sensing identification method based on semi-supervised deep learning, which aims to realize the following steps: under the support of a small amount of samples and limited calculation force, landslide in the remote sensing image is efficiently and accurately identified.
The technical scheme adopted by the invention is as follows:
a mountain large-scale landslide intelligent remote sensing identification method of semi-supervised deep learning comprises the following steps:
s1: selecting a remote sensing image containing landslide and digital elevation data of a corresponding area for data synthesis to obtain a landslide remote sensing image data set containing the digital elevation data, carrying out label labeling on part of the landslide remote sensing image data containing the digital elevation data to obtain a sample set containing label labeling, and forming the rest of the landslide remote sensing image data containing the digital elevation data into a sample set without label labeling;
s2: inputting the sample set containing the label marks and the sample set not containing the label marks obtained in the S1 into a semantic segmentation network to respectively construct a student training network model and a teacher training network model, respectively training the sample set containing the label marks and the sample set not containing the label marks through the student training network model and the teacher training network model and judging whether the training is finished, predicting the sample set which is correspondingly finished with the training through the student training network model and the teacher training network model and obtaining a prediction result, returning the sample set which is not finished with the training to the semantic segmentation network, inputting the prediction result into a discriminator to respectively judge the landslide characteristic types of the prediction results corresponding to the student training network model and the teacher training network model and respectively obtaining a student training result and a teacher training result;
s3: and (3) calculating a cross entropy loss value of the student training result obtained in the step (S2), calculating mean square errors of the student training network model and the teacher training network model, calculating an exponential moving average line according to the cross entropy loss value of the student training network model and the mean square error of the teacher training network model, adjusting and updating the weight in the student training network model according to the exponential moving average line, and continuously predicting and carrying out weighted average to update the weight of the teacher training network model according to the adjusted and updated weight of the student training network model.
Preferably, in S1, the landslide remote sensing image with a spatial resolution of 0.2m to 0.9m and the digital elevation data model of the corresponding area are selected for data synthesis, so as to obtain a landslide remote sensing image data set containing digital elevation data.
Preferably, the image enhancement is performed on the landslide remote sensing image data set containing the digital elevation data in S1, and the image enhancement mode includes: cutting, transforming, rotating, turning, zooming and translating, wherein a sample image in the landslide remote sensing image data set containing the digital elevation data is cut, the size of the cut sample image is consistent with that of the corresponding digital elevation data, the sample image is uniformly zoomed to be square, and a non-square sample image is cut to be square under the condition that the integrity of a landslide is not damaged.
Preferably, the sample normalization process in S1 is:
carrying out image normalization processing on the landslide remote sensing image data set containing the digital elevation data obtained in the step S1, unifying the original images of the input layer, and reducing the size range of the characteristic value, wherein the formula is as follows:
x * the corresponding is the pixel value of the pixel point after the image normalization operation, x is the pixel value of each point of the image to be processed, and min (x) and max (x) are respectively the minimum and maximum pixel values found in the image.
Preferably, in S1, the process of labeling the landslide remote sensing image data partially containing the digital elevation data is as follows: an Arcmap editor is adopted to carry out vector pattern spot drawing on a small number of characteristic regions for label labeling, a boundary box is generated, a minimum wrapping rectangle is directly generated by element envelope rectangle surface conversion in an Arcmap tool, then the minimum wrapping rectangle is output and stored as a slice, and the proportion of a sample containing label labeling to a sample not containing label labeling is 1.
Preferably, the training process of constructing the student training network model and the teacher training network model in S2 is as follows:
the method comprises the steps of training and testing by using a Mask RCNN algorithm based on deep learning framework Tensorflow, carrying out recognition training by adopting manual labeling and label labeling samples of partial public data sets, acquiring and outputting Feature Map based on a main network ResNet, then sampling by a Feature pyramid FPN, combining and outputting large-scale landslide features.
Further, the process of predicting the corresponding training completed sample set through the student training network model and the teacher training network model in S2 is as follows:
the semantic segmentation network in the S2 is a candidate area network, and the large landslide feature obtained in the S2 is used for generating an anchor frame by taking each pixel as a center through the candidate area network, wherein the position coordinates of the real frame are x _ a and y _ a, the width and the height of the real frame are w _ a and h _ a, and the position coordinates of the anchor frame are x * ,y * The width and height of the anchor frame are w * ,h * The offset formula is:
moving on each image of a large-scale landslide data set by virtue of an anchor frame to help candidate regional network training, then performing classification tasks and regression tasks on the anchor frames, calculating the score of each anchor frame by the classification tasks, judging the probability of landslide, representing the positions of the anchor frames by using the upper left point and the lower right point of the regression task, sequencing according to the score conditions of the anchor frames, and dividing positive examples and negative examples according to the probability;
the candidate area network classification is based on a Softmax function, which is as follows:
wherein a is j Representing the probability of forward computation of the j class, S j Is the probability calculated by the Softmax function;
the cross entropy loss function used is as follows:
wherein y is j Representing a genuine tag, S j The probability is indicated.
Preferably, the process of inputting the prediction result into the discriminator in S2 to respectively judge the landslide feature types of the prediction results corresponding to the student training network model and the teacher training network model is as follows:
after a prediction result output by a semantic segmentation network in a student training network model is processed, extracting corresponding large-scale landslide characteristics by using ROI (region of interest), wherein the ROI Align loss function is as follows:
these large landslide feature types are judged in a pre-trained ResNet-50 discriminator while regressing the prediction box before adjustment.
Preferably, the mean square error of the student training network model and the teacher training network model is calculated in S3, an exponential moving average line is calculated according to the cross entropy loss value of the student training network model and the mean square error of the teacher training network model, and the process of adjusting and updating the weights in the student training network model according to the feedback of the exponential moving average line is as follows:
calculating a loss value of a student training network model according to a loss function of a Mask RCNN, generating a Mask by a full-connection network in a Mask branch, comparing information of each pixel point of a real Mask and a predicted Mask, and adopting a two-class cross entropy loss function, wherein the loss value of the student training network model is calculated as follows:
wherein y is i To predict the probability, y1 i For true probability, a loss value of 0 will reach the ideal value if and only if the two values are the same. When the probability difference between the predicted value and the true value is larger, the loss value is larger, and the weight parameter is changed according to the loss value to realize back propagation;
calculating the mean square error between the student training network model and the teacher training network model, namely the expected distance between the prediction result of the student training network model and the prediction result of the teacher training network model, wherein the calculation formula is as follows:
wherein f (x, theta, eta) represents the student training network model, f (x, theta ', eta') represents the teacher training network model, J (theta) represents the expected distance between the student training network model and the prediction result of the teacher training network model,
the method comprises the steps of representing an expected value between a student training network model and a teacher training network model, x representing a training prediction result, theta representing the weight of the student training network model, theta 'representing the weight of the teacher training network model, eta representing the noise of the student training network model, eta' representing the noise of the teacher training network model, calculating a mean square error by inputting the weight (theta, theta ') and the noise (eta, eta') and combining the training prediction result, taking the mean square error as a consistency cost, and judging the precision and reliability of the training results of the student training network model and the teacher training network model according to the mean square error;
calculating an index moving average line according to a cross entropy loss value and a mean square error obtained by the student training network model, adjusting and updating the weight in the student training network model according to the index moving average line, wherein a formula for calculating the index moving average line is as follows:
avg＝loss+εJ(θ) (10)
wherein loss represents the loss value of the student training network model, and epsilon represents the weight number of the mean square error.
Preferably, in S3, the process of continuously predicting and performing weighted average to update the weight of the teacher training network model according to the adjusted and updated weight of the student training network model includes:
the following formula is adopted:
θ' t ＝αθ' t-1 +(1-α)θ t (11)
wherein alpha represents a smoothing coefficient parameter, theta represents the weight of the student training network model, theta' represents the weight of the teacher training network model, t represents the training times, and the weight of the teacher training network model in the t-1 training
And the weight theta of the network model for student training in the t-th training t Updating the weight theta of the teacher training network model in the t-th training t '。
In summary, due to the adoption of the technical scheme, the invention has the beneficial effects that:
1. the method can improve the precision and reliability of remote sensing image recognition and landslide result extraction when the similarity is high or landslide characteristics are not obvious under the condition that label marking samples are few; the landslide information can be effectively extracted only through the input remote sensing image without additional auxiliary data, the landslide image extraction process is based on landslide detection, and in combination with digital elevation data, landslide images with high similarity or unobvious landslide characteristics cannot influence the identification extraction result. The method realizes pixel-level automatic landslide detection and landslide differentiation with high similarity, integrates a semi-supervised learning method, trains a sample with a label mark and a sample without the label mark respectively by constructing a student training network model and a teacher training network model, can obtain a model with good classification effect only by a small number of label samples, can continuously adjust the weight of the training network model, and continuously iteratively adjusts the optimization, saves a large amount of time for manual label marking, and improves the training efficiency;
2. the landslide identification scheme based on large-scale landslide area search is constructed, a Mask RCNN algorithm framework is combined with semi-supervised learning, a characteristic graph is obtained and output through a trunk network ResNet, then sampling is carried out through a characteristic pyramid FPN, large-scale landslide characteristic output is combined, the back propagation of weight parameters is achieved, and a candidate area network is used for auxiliary training, so that the precision and the reliability of training are greatly improved, landslide characteristics can be better extracted, the computational loss is effectively reduced, the speed of extracting landslides is greatly accelerated, and the reliability of identifying and extracting landslides is improved;
3. a discriminator module is established based on ResNet-50, the mean square error between the output of the extracted landslide model and the input of the landslide sample model is adopted as a training target, and the quality of the output results of the student training network model and the teacher training network model can be effectively evaluated through the discriminator module, the loss function of Mask RCNN and the mean square error between the two models. And realizing automatic fine adjustment of the model according to the loss value, the mean square error between the two training network models and the exponential moving average line.
Drawings
The invention will now be described, by way of example, with reference to the accompanying drawings, in which:
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a flow chart of the neural network of the present invention;
FIG. 3 is a block diagram of two training network models of the present invention;
FIG. 4 is a graph comparing the results of the models of the present invention;
FIG. 5 is a comparison of the convergence results of the model training of the present invention;
FIG. 6 is a graph of loss values for the training results of the present invention;
FIG. 7 is a graph of learning rate of training results of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all the embodiments. The components of the embodiments of the present application, as generally described and illustrated in the figures herein, could be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present application, presented in the accompanying drawings, is not intended to limit the scope of the claimed application, but is merely representative of selected embodiments of the application. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present application without making any creative effort, shall fall within the protection scope of the present application.
In the description of the embodiments of the present application, it should be noted that the terms "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", and the like indicate orientations or positional relationships based on orientations or positional relationships shown in the drawings or orientations or positional relationships that the present invention is used to place as usual, and are only used for convenience of description and simplicity of description, but do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and thus, should not be construed as limiting the present application. Furthermore, the terms "first," "second," "third," and the like are used solely to distinguish one from another and are not to be construed as indicating or implying relative importance.
The present invention is described in detail below with reference to fig. 1 to 5.
FIG. 1 shows a flow diagram of the overall architecture of the present invention, which is divided into three parts S1, S2, and S3 according to steps, and the three parts are performed in sequence, each part has a separate output and is the input of the next part;
fig. 2 shows a flow chart of a training network used by the model, which can be divided into three parts of feature extraction, RPN network, classification and regression prediction,
fig. 3 shows the structure and relationship between the student training network model and the teacher training network model, wherein the student training network model training part trains samples containing label labels, and the teacher training network model trains the rest samples without label labels, wherein the ratio of the samples containing label labels to the samples without label labels is 1;
FIG. 4 shows an image comparison of an original image of a landslide remote sensing image, digital elevation data, an artificially identified landslide area, a landslide segmented by training after semi-supervised learning, and a landslide image segmented by training in combination with digital elevation data; a1, A2 and A3 are landslide remote sensing image original images, AA1, AA2 and AA3 are digital elevation data corresponding to landslide images, B1, B2 and B3 are landslide areas identified manually, C1, C2 and C3 are landslides divided through training after semi-supervised learning, and D1, D2 and D3 are landslide images divided by combining the digital elevation data;
fig. 5 shows the prediction results obtained after model individual discrimination Mask RCNN semantic segmentation and semi-supervised model fine tuning, where Original Img is an Original high-resolution remote sensing image, train Label is an artificially established Label, and Pre is the result of model prediction.
Fig. 6 and 7 respectively show the loss value and the learning rate of the landslide remote sensing image recognition finally calculated according to the loss function of the student training network model.
Examples
A mountain area large landslide intelligent remote sensing identification method of semi-supervised deep learning comprises the following steps:
s1: selecting a remote sensing image containing landslide and digital elevation data of a corresponding area for data synthesis to obtain a landslide remote sensing image data set containing the digital elevation data, carrying out label labeling on part of the landslide remote sensing image data containing the digital elevation data to obtain a sample set containing label labeling, and forming the rest of the landslide remote sensing image data containing the digital elevation data into a sample set without label labeling;
s2: inputting the sample set containing the label marks and the sample set not containing the label marks obtained in the S1 into a semantic segmentation network to respectively construct a student training network model and a teacher training network model, respectively training the sample set containing the label marks and the sample set not containing the label marks through the student training network model and the teacher training network model and judging whether the training is finished, predicting the sample set which is correspondingly finished with the training through the student training network model and the teacher training network model and obtaining a prediction result, returning the sample set which is not finished with the training to the semantic segmentation network, inputting the prediction result into a discriminator to respectively judge the landslide characteristic types of the prediction results corresponding to the student training network model and the teacher training network model and respectively obtaining a student training result and a teacher training result;
s3: and (3) calculating a cross entropy loss value of the student training result obtained in the step (S2), calculating mean square errors of the student training network model and the teacher training network model, calculating an exponential moving average line according to the cross entropy loss value of the student training network model and the mean square error of the teacher training network model, adjusting and updating the weight in the student training network model according to the exponential moving average line, and continuously predicting and carrying out weighted average to update the weight of the teacher training network model according to the adjusted and updated weight of the student training network model.
In the S1, a landslide remote sensing image with the spatial resolution of 0.2-0.9 m and a digital elevation data model of a corresponding area are selected for data synthesis, and a landslide remote sensing image data set containing digital elevation data is obtained.
And carrying out image enhancement on the landslide remote sensing image data set containing the digital elevation data in the S1, wherein the image enhancement mode comprises the following steps: and cutting, transforming, rotating, turning, zooming and translating, wherein a sample image in the landslide remote sensing image data set containing the digital elevation data is cut, the cut sample image is consistent with the corresponding digital elevation data in size, the sample image is uniformly zoomed to be square, and a non-square sample image is cut to be square under the condition that the integrity of the landslide is not damaged.
The sample normalization in S1 is:
carrying out image normalization processing on the landslide remote sensing image data set containing the digital elevation data obtained in the S1, unifying the original images of the input layer, and reducing the size range of the characteristic value, wherein the formula is as follows:
x * the corresponding is the pixel value of the pixel point after the image normalization operation, x is the pixel value of each point of the image to be processed, and min (x) and max (x) are respectively the minimum and maximum pixel values found in the image.
The process of labeling the landslide remote sensing image data partially containing the digital elevation data in the S1 comprises the following steps: an Arcmap editor is adopted to carry out vector pattern spot drawing on a small number of characteristic regions for label labeling, a boundary box is generated, a minimum wrapping rectangle is directly generated by element envelope rectangle surface conversion in an Arcmap tool, then the minimum wrapping rectangle is output and stored as a slice, and the proportion of a sample containing label labeling to a sample not containing label labeling is 1. The training process for constructing the student training network model and the teacher training network model in the S2 comprises the following steps:
the method comprises the steps of training and testing by using a Mask RCNN algorithm based on deep learning framework Tensorflow, carrying out recognition training by adopting manual labeling and label labeling samples of partial public data sets, acquiring and outputting Feature Map based on a main network ResNet, then sampling by a Feature pyramid FPN, combining and outputting large-scale landslide features.
In S2, the process of predicting the corresponding training completed sample set through the student training network model and the teacher training network model is as follows:
the semantic segmentation network in the S2 is a candidate area network, and the large landslide feature obtained in the S2 is used for generating an anchor frame by taking each pixel as a center through the candidate area network, wherein the position coordinates of the real frame are x _ a and y _ a, the width and the height of the real frame are w _ a and h _ a, and the position coordinates of the anchor frame are x * ,y * The width and height of the anchor frame are w * ,h * The offset formula is:
moving on each image of a large landslide data set by using an anchor frame to help candidate regional network training, then performing a classification task and a regression task on the anchor frames, wherein the classification task calculates the score of each anchor frame and judges the probability of landslide, the regression task is to find the position of the anchor frame, the positions of the anchor frames are represented by two points, namely the upper left point and the lower right point, and the positive and negative examples are divided according to the probability by sequencing the score conditions of the anchor frames;
the candidate area network classification is based on a Softmax function, which is as follows:
wherein a is j Representing the probability of forward computation of the j class, S j Is the probability calculated by the Softmax function; cross entropy of useThe loss function is given by:
wherein y is j Representing a genuine label, S j The probability is indicated.
In S2, the process of inputting the prediction results into the discriminator to respectively judge the landslide feature types of the prediction results corresponding to the student training network model and the teacher training network model is as follows:
after a prediction result output by a semantic segmentation network in a student training network model is processed, extracting corresponding large-scale landslide characteristics by using ROI (region of interest), wherein the ROI Align loss function is as follows:
these large landslide feature types are judged in a pre-trained ResNet-50 discriminator while regressing the prediction box before adjustment.
S3, calculating the mean square error of the student training network model and the teacher training network model, calculating an index moving average line according to the cross entropy loss value of the student training network model and the mean square error of the teacher training network model, and adjusting and updating the weight in the student training network model according to the index moving average line feedback, wherein the process comprises the following steps:
calculating a loss value of a student training network model according to a loss function of a Mask RCNN, generating a Mask by a full-connection network in a Mask branch, comparing information of each pixel point of a real Mask and a predicted Mask, and adopting a two-class cross entropy loss function, wherein the loss value of the student training network model is calculated as follows:
whereiny i To predict the probability, y1 i For true probability, the loss value is 0 to reach the ideal value if and only if the two values are the same. When the probability difference between the predicted value and the true value is larger, the loss value is larger, and the weight parameter is changed according to the loss value to realize back propagation;
calculating the mean square error between the student training network model and the teacher training network model, namely the expected distance between the prediction result of the student training network model and the prediction result of the teacher training network model, wherein the calculation formula is as follows:
wherein f (x, theta, eta) represents the student training network model, f (x, theta ', eta') represents the teacher training network model, J (theta) represents the expected distance between the student training network model and the prediction result of the teacher training network model,
the method comprises the steps of representing an expected value between a student training network model and a teacher training network model, x representing a training prediction result, theta representing the weight of the student training network model, theta 'representing the weight of the teacher training network model, eta representing the noise of the student training network model, eta' representing the noise of the teacher training network model, calculating a mean square error by inputting the weight (theta, theta ') and the noise (eta, eta') and combining the training prediction result, taking the mean square error as a consistency cost, and judging the precision and reliability of the training results of the student training network model and the teacher training network model according to the mean square error;
calculating an exponential moving average line according to a cross entropy loss value and a mean square error obtained by the student training network model, adjusting and updating the weight in the student training network model according to the exponential moving average line, wherein a formula for calculating the exponential moving average line is as follows:
avg＝loss+εJ(θ) (10)
wherein loss represents the loss value of the student training network model, and epsilon represents the weight number of the mean square error.
In S3, according to the adjusted and updated weight of the student training network model, the process of continuously predicting and carrying out weighted average to update the weight of the teacher training network model comprises the following steps:
the following formula is adopted:
θ' t ＝αθ' t-1 +(1-α)θ t (11)
wherein alpha represents a smoothing coefficient parameter, theta represents the weight of a student training network model, theta' represents the weight of a teacher training network model, t represents the training times, and the weight theta of the teacher training network model in the t-1 training t ' -1 And the weight theta of the network model trained by the student in the t training t Updating the weight theta of the teacher training network model in the t-th training t '。
The above-mentioned embodiments only express the specific embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present application. It should be noted that, for those skilled in the art, without departing from the technical idea of the present application, several changes and modifications can be made, which all belong to the protection scope of the present application.
Claims (10)
1. The intelligent remote sensing identification method for the large landslide in the mountainous area through semi-supervised deep learning is characterized by comprising the following steps of:
s1: selecting a remote sensing image containing landslide and digital elevation data of a corresponding area for data synthesis to obtain a landslide remote sensing image data set containing the digital elevation data, carrying out label labeling on part of the landslide remote sensing image data containing the digital elevation data to obtain a sample set containing label labeling, and forming the rest of landslide remote sensing image data containing the digital elevation data into a sample set without label labeling;
s2: inputting the sample set containing the label marks and the sample set without the label marks obtained in the S1 into a semantic segmentation network to respectively construct a student training network model and a teacher training network model, respectively training the sample set containing the label marks and the sample set without the label marks through the student training network model and the teacher training network model and judging whether the training is finished, predicting the sample set which is correspondingly finished with the training through the student training network model and the teacher training network model and obtaining a prediction result, returning the sample set which is not finished with the training to the semantic segmentation network, inputting the prediction result into a discriminator to respectively judge the landslide characteristic types of the prediction results corresponding to the student training network model and the teacher training network model and respectively obtain a student training result and a teacher training result;
s3: and (3) calculating the cross entropy loss value of the student training result obtained in the S2, calculating the mean square error of the student training network model and the teacher training network model, calculating an exponential moving average line according to the cross entropy loss value of the student training network model and the mean square error of the teacher training network model, adjusting and updating the weight in the student training network model according to the exponential moving average line, and continuously predicting and performing weighted average to update the weight of the teacher training network model according to the adjusted and updated weight of the student training network model.
2. The intelligent remote sensing identification method for large landslides in mountainous areas for semi-supervised deep learning according to claim 1, wherein in S1, landslide remote sensing images with spatial resolution ranging from 0.2m to 0.9m and digital elevation data models of corresponding areas are selected for data synthesis to obtain a landslide remote sensing image data set containing digital elevation data.
3. The intelligent remote sensing identification method for large landslides in mountainous areas based on semi-supervised deep learning according to claim 1, wherein image enhancement is performed on the landslide remote sensing image data set containing digital elevation data in S1, and the image enhancement mode comprises the following steps: cutting, transforming, rotating, turning, zooming and translating, wherein a sample image in the landslide remote sensing image data set containing the digital elevation data is cut, the size of the cut sample image is consistent with that of the corresponding digital elevation data, the sample image is uniformly zoomed to be square, and a non-square sample image is cut to be square under the condition that the integrity of a landslide is not damaged.
4. The intelligent remote sensing identification method for the large landslide in the mountain area based on the semi-supervised deep learning as recited in claim 1, wherein the sample normalization processing in the S1 is as follows:
carrying out image normalization processing on the landslide remote sensing image data set containing the digital elevation data obtained in the S1, unifying the original images of the input layer, and reducing the size range of the characteristic value, wherein the formula is as follows:
x * the corresponding is the pixel value of the pixel point after the image normalization operation, x is the pixel value of each point of the image to be processed, and min (x) and max (x) are respectively the minimum and maximum pixel values found in the image.
5. The intelligent remote sensing identification method for large landslides in mountainous areas based on semi-supervised deep learning as claimed in claim 1, wherein the process of labeling the landslide remote sensing image data partially containing digital elevation data in S1 is as follows: an Arcmap editor is adopted to carry out vector pattern spot drawing on a small number of characteristic regions for label labeling, a boundary box is generated, a minimum wrapping rectangle is directly generated by element envelope rectangle surface conversion in an Arcmap tool, then the minimum wrapping rectangle is output and stored as a slice, and the proportion of a sample containing label labeling to a sample not containing label labeling is 1.
6. The intelligent remote sensing recognition method for mountain large landslides of semi-supervised deep learning according to claim 1, wherein the training process for constructing the student training network model and the teacher training network model in S2 is as follows:
the method comprises the steps of training and testing by using a Mask RCNN algorithm based on deep learning framework Tensorflow, carrying out recognition training by adopting manual labeling and label labeling samples of partial public data sets, acquiring and outputting FeatureMap based on a main network ResNet, then sampling by a feature pyramid FPN, combining large-scale landslide feature output.
7. The intelligent remote sensing identification method for mountain large-scale landslides of semi-supervised deep learning according to claim 1 or 6, wherein the process of predicting the corresponding trained sample set through the student training network model and the teacher training network model in S2 is as follows:
the semantic segmentation network in the S2 is a candidate area network, and the anchor frame is generated by taking each pixel as the center of the large-scale landslide feature obtained in the S2 through the candidate area network, wherein the position coordinates of the real frame are x _ a and y _ a, the width and the height of the real frame are w _ a and h _ a, and the position coordinates of the anchor frame are x * ,y * The width and height of the anchor frame are w * ,h * The offset formula is:
moving on each image of a large-scale landslide data set by virtue of an anchor frame to help candidate regional network training, then performing classification tasks and regression tasks on the anchor frames, calculating the score of each anchor frame by the classification tasks, judging the probability of landslide, representing the positions of the anchor frames by using the upper left point and the lower right point of the regression task, sequencing according to the score conditions of the anchor frames, and dividing positive examples and negative examples according to the probability;
the candidate area network classification is based on the Softmax function, which is given by:
wherein a is j Representing the probability of forward computation of the j class, S j Is the probability calculated by the Softmax function;
the cross entropy loss function used is as follows:
wherein y is j Representing a genuine label, S j The probability is represented.
8. The intelligent remote sensing identification method for mountain large landslides of semi-supervised deep learning according to claim 1, wherein the process of inputting the prediction results into a discriminator in S2 to respectively judge the landslide feature types of the prediction results corresponding to the student training network model and the teacher training network model is as follows:
after the prediction result output by the semantic segmentation network in the student training network model is processed, extracting the corresponding large-scale landslide characteristic by using ROI Align, wherein the ROI Align loss function is as follows:
these large landslide feature types are judged in a pre-trained ResNet-50 discriminator while regressing the prediction box before adjustment.
9. The intelligent remote sensing identification method for mountain landslides with semi-supervised deep learning according to claim 1, wherein the mean square error of the student training network model and the teacher training network model is calculated in S3, an exponential moving mean line is calculated according to the cross entropy loss value of the student training network model and the mean square error of the teacher training network model, and the process of updating the weight in the student training network model according to the feedback adjustment of the exponential moving mean line is as follows:
calculating a loss value of a student training network model according to a loss function of Mask RCNN, generating a Mask by a full-connection network in a Mask branch, comparing information of each pixel point of the real Mask and the prediction Mask, adopting a two-classification cross entropy loss function, and calculating the loss value of the student training network model as follows:
wherein y is i To predict the probability, y1 i For true probability, a loss value of 0 will reach the ideal value if and only if the two values are the same. When the probability difference between the predicted value and the true value is larger, the loss value is larger, and the weight parameter is changed according to the loss value to realize back propagation;
calculating the mean square error between the student training network model and the teacher training network model, namely the expected distance between the prediction result of the student training network model and the prediction result of the teacher training network model, wherein the calculation formula is as follows:
wherein f (x, theta, eta) represents the student training network model, f (x, theta ', eta') represents the teacher training network model, J (theta) represents the expected distance between the student training network model and the prediction result of the teacher training network model,
representing an expectation value between a student training network model and a teacher training network model, x representing a training prediction result, theta representing a weight of the student training network model, theta ' representing a weight of the teacher training network model, eta representing noise of the student training network model, eta ' representing noise of the teacher training network model, and eta ' representing the noise of the teacher training network model by inputting the weights (theta, theta ') and the noise (eta, eta ') and combiningCalculating the mean square error of the training prediction result, taking the mean square error as consistency cost, and judging the precision and reliability of the training results of the student training network model and the teacher training network model according to the mean square error;
calculating an exponential moving average line according to a cross entropy loss value and a mean square error obtained by the student training network model, adjusting and updating the weight in the student training network model according to the exponential moving average line, wherein a formula for calculating the exponential moving average line is as follows:
avg＝loss+εJ(θ) (10)
wherein loss represents the loss value of the student training network model, and epsilon represents the weight number of the mean square error.
10. The intelligent remote sensing identification method for mountain large landslides of semi-supervised deep learning according to claim 1, wherein the process of continuously predicting and performing weighted average to update the weight of the teacher training network model according to the adjusted and updated weight of the student training network model in S3 comprises:
the following formula is adopted:
θ′ t ＝αθ′ t-1 +(1-α)θ t (11)
wherein alpha represents a smoothing coefficient parameter, theta represents the weight of the student training network model, theta ' represents the weight of the teacher training network model, t represents the training times, and the weight theta ' of the teacher training network model in the t-1 training ' t-1 And the weight theta of the network model trained by the student in the t training t Updating the weight theta of the teacher training network model in the t-th training' t 。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211584428.6A CN115761513A (en) | 2022-12-09 | 2022-12-09 | Intelligent remote sensing identification method for mountain large landslide based on semi-supervised deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211584428.6A CN115761513A (en) | 2022-12-09 | 2022-12-09 | Intelligent remote sensing identification method for mountain large landslide based on semi-supervised deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115761513A true CN115761513A (en) | 2023-03-07 |
Family
ID=85345343
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211584428.6A Pending CN115761513A (en) | 2022-12-09 | 2022-12-09 | Intelligent remote sensing identification method for mountain large landslide based on semi-supervised deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115761513A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116227938A (en) * | 2023-04-26 | 2023-06-06 | 四川川核地质工程有限公司 | Mountain landslide intelligent early warning method and system based on Beidou satellite |
| CN116597287A (en) * | 2023-07-17 | 2023-08-15 | 云南省交通规划设计研究院有限公司 | Remote sensing image landslide recognition method based on deep learning method |
| CN116704357A (en) * | 2023-08-09 | 2023-09-05 | 江西省水利科学院(江西省大坝安全管理中心、江西省水资源管理中心) | YOLOv 7-based intelligent identification and early warning method for landslide of dam slope |
| CN118130742A (en) * | 2024-05-06 | 2024-06-04 | 阳光学院 | River and lake water quality remote sensing inversion and evaluation method based on transfer learning |
-
2022
- 2022-12-09 CN CN202211584428.6A patent/CN115761513A/en active Pending
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116227938A (en) * | 2023-04-26 | 2023-06-06 | 四川川核地质工程有限公司 | Mountain landslide intelligent early warning method and system based on Beidou satellite |
| CN116227938B (en) * | 2023-04-26 | 2023-06-30 | 四川川核地质工程有限公司 | Mountain landslide intelligent early warning method and system based on Beidou satellite |
| CN116597287A (en) * | 2023-07-17 | 2023-08-15 | 云南省交通规划设计研究院有限公司 | Remote sensing image landslide recognition method based on deep learning method |
| CN116704357A (en) * | 2023-08-09 | 2023-09-05 | 江西省水利科学院(江西省大坝安全管理中心、江西省水资源管理中心) | YOLOv 7-based intelligent identification and early warning method for landslide of dam slope |
| CN116704357B (en) * | 2023-08-09 | 2023-10-27 | 江西省水利科学院(江西省大坝安全管理中心、江西省水资源管理中心) | YOLOv 7-based intelligent identification and early warning method for landslide of dam slope |
| CN118130742A (en) * | 2024-05-06 | 2024-06-04 | 阳光学院 | River and lake water quality remote sensing inversion and evaluation method based on transfer learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111353413B (en) | Low-missing-report-rate defect identification method for power transmission equipment | |
| CN109977812B (en) | Vehicle-mounted video target detection method based on deep learning | |
| CN115761513A (en) | Intelligent remote sensing identification method for mountain large landslide based on semi-supervised deep learning | |
| CN108596055B (en) | Airport target detection method of high-resolution remote sensing image under complex background | |
| CN112232371B (en) | American license plate recognition method based on YOLOv3 and text recognition | |
| CN111914634B (en) | Automatic detection method and system for well lid class resisting complex scene interference | |
| CN110555841B (en) | SAR image change detection method based on self-attention image fusion and DEC | |
| CN112347895A (en) | Ship remote sensing target detection method based on boundary optimization neural network | |
| CN111985325B (en) | Aerial small target rapid identification method in extra-high voltage environment evaluation | |
| CN110798805B (en) | Data processing method and device based on GPS track and storage medium | |
| EP4105600A2 (en) | Method for automatically producing map data, related apparatus and computer program product | |
| CN110956207B (en) | Method for detecting full-element change of optical remote sensing image | |
| CN114612835A (en) | Unmanned aerial vehicle target detection model based on YOLOv5 network | |
| CN113469278B (en) | Strong weather target identification method based on deep convolutional neural network | |
| CN111027511A (en) | Remote sensing image ship detection method based on region of interest block extraction | |
| CN110532961A (en) | A kind of semantic traffic lights detection method based on multiple dimensioned attention mechanism network model | |
| CN114332473A (en) | Object detection method, object detection device, computer equipment, storage medium and program product | |
| CN111507353B (en) | Chinese field detection method and system based on character recognition | |
| CN111414951B (en) | Fine classification method and device for images | |
| CN116168246A (en) | Method, device, equipment and medium for identifying waste slag field for railway engineering | |
| CN113361528B (en) | Multi-scale target detection method and system | |
| CN114067245A (en) | Method and system for identifying hidden danger of external environment of railway | |
| CN113988222A (en) | Forest fire detection and identification method based on fast-RCNN | |
| CN111145314B (en) | Method for extracting place name symbol of scanning electronic map by combining place name labeling | |
| CN113327253A (en) | Weak and small target detection method based on satellite-borne infrared remote sensing image |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |