Ground disaster hidden danger troubleshooting method and system assisted by multi-source time sequence remote sensing soil water content prediction technology

Abstract

The invention provides a ground disaster hidden danger troubleshooting method and system assisted by a multi-source time sequence remote sensing soil water content prediction technology, and the method comprises the steps: carrying out the preprocessing of obtained multi-source time sequence remote sensing data related to the soil water content, obtaining model basic data, and constructing a model sample database; performing optimization training on the pre-constructed soil water content prediction deep learning model by adopting the sample database to obtain a model parameter initialization model with optimal weight; and initializing the model by using the model parameter with the optimal weight, predicting the soil water content of the next time step of the multi-source time sequence remote sensing data to obtain a prediction result of the soil water content, and then assisting in investigation of the ground disaster hidden danger by using the prediction result. By means of the method, the defects that a traditional soil moisture content monitoring scheme is high in measurement cost and low in efficiency can be overcome, comprehensive analysis can be carried out in combination with other related information, and the precision of ground disaster hidden danger troubleshooting relative to single remote sensing data is improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of soil water content prediction and geological disaster hidden danger investigation. More specifically, it relates to a method and system for geological disaster hidden danger investigation assisted by a multi-source time-series remote sensing soil water content prediction technology. Background Art

[0002] Frequent geological disasters, such as landslides, collapses, debris flows, etc., have brought serious interference and threats to the daily operation of infrastructure in various fields, and have also brought serious threats to the lives and property safety of the people along the line. Therefore, it is of great significance to use technical means to conduct high-precision and regular prior investigations on potential geological disaster hidden danger points.

[0003] Using remote sensing means can achieve large-scale and high-timeliness geological disaster hidden danger investigation. For example, high-resolution optical remote sensing images can interpret the spectral texture and other change features of regional ground objects before the occurrence of geological disasters to identify potential geological disaster hidden danger points. In addition, through InSAR technology, deformation information within the region can be obtained, and the specific situation of geological disaster hidden danger points can be further determined through the deformation information within the region. At present, through remote sensing technology means, geological disaster hidden danger points can be investigated, but the above methods all ignore the connection between soil moisture content and the formation of geological disaster hidden dangers, and only use the spectral change features or deformation features of geological disaster hidden danger points, resulting in room for further improvement in the accuracy and reliability of remote sensing-based geological disaster hidden danger point investigations.

[0004] Soil moisture, also known as soil humidity or soil water content, is an important parameter in the earth's biogeochemical cycle and energy exchange process. Since the increase in soil moisture content will lead to problems such as the reduction of friction between soils and the increase of pore water pressure, the risk of geological disasters such as landslides, collapses, and ground collapses will increase. For example, geological disasters induced by continuous heavy rainfall are like this. Therefore, the change information of soil moisture content has important application value in geological disaster investigation and prevention.

[0005] However, at present, traditional soil moisture content monitoring often determines it through the method of field sampling plus laboratory measurement. The advantage of the field sampling plus laboratory measurement method is that accurate soil moisture content information can be obtained. The disadvantage is that the obtained soil water content can only reflect the single-point soil water content information at the sampling point location. It not only has a small scale but also requires a long laboratory determination time, and its required human and physical costs are relatively large, making it difficult to meet the requirements of large-scale and rapid update of soil water content in the geological disaster hidden danger investigation scenario. Summary of the Invention

[0006] In view of the above problems, the purpose of the present invention is to provide a method for geological disaster hidden danger investigation assisted by a multi-source time-series remote sensing soil water content prediction technology to solve the problems such as low investigation accuracy and poor reliability existing in the current geological disaster hidden danger investigation plan.

[0007] The present invention provides a method for investigating potential geological disaster hazards assisted by a multi-source time-series remote sensing soil water content prediction technology, including:

[0008] Preprocess the acquired multi-source time-series remote sensing data associated with soil water content to obtain model basic data;

[0009] Construct a model sample database based on the model basic data;

[0010] Use the sample database to optimize and train a pre-constructed deep learning model for predicting soil water content to obtain an initialized model with optimal weights for model parameters; wherein, during the training process, deep supervision training is performed on the model using sample category information and supervision training is performed on the model using sample time-series information, the randomly initialized parameters are optimized and adjusted, and the model weights that perform optimally on the validation data set are saved;

[0011] Use the initialized model with optimal weights for model parameters to predict the soil water content at the next time step of the input multi-source time-series remote sensing data of the area to be predicted, and obtain the prediction result of the soil water content of the area;

[0012] According to the prediction result of the soil water content of the area, determine the change rate and second-order change rate of the soil water content; according to the change rate and second-order change rate, determine the risk level of potential geological disasters in the area.

[0013] In addition, an optional solution is that the multi-source time-series remote sensing data includes optical remote sensing time-series data, synthetic aperture radar time-series data, corresponding area land cover time-series data, and meteorological precipitation data.

[0014] In addition, an optional solution is that the preprocessing includes geometric correction, radiometric correction, index analysis, sampling interpolation, normalization, and multi-source fusion.

[0015] In addition, an optional solution is that the index analysis indirectly reflects the soil water content situation by analyzing the vegetation growth status through the normalized difference vegetation index; wherein, the calculation formula of the normalized difference vegetation index is as follows:

[0016]

[0017] Wherein, NDVI represents the normalized difference vegetation index, and NIR and R respectively represent the near-infrared band and red band of the optical remote sensing image.

[0018] In addition, an optional solution is that the normalization is performed by the Z-Score normalization method to unify the value ranges and data distributions from different data sources; wherein, the Z-Score normalization is shown as the following formula:

[0019]

[0020] Among them, Z represents the normalized data, X represents the data to be normalized, and the data to be normalized includes optical remote sensing data, radar remote sensing data, land cover data, and the derived normalized difference vegetation index after max-min normalization; μ and σ respectively represent the mean and standard deviation of the data to be normalized.

[0021] In addition, an optional solution is that the multi-source fusion is used to fuse the normalized data by the method of weighted fusion addition, and the formula is as follows:

[0022]

[0023] Among them, X1, X2, X3, X index respectively represent optical remote sensing data, SAR radar data, land cover type data, and remote sensing index obtained by index analysis; a, b, c, d are the respective proportional weights of X1, X2, X3, X index respectively.

[0024] In addition, an optional solution is to construct a model sample database based on the model basic data, including: performing time-series block cropping on the model basic data, and corresponding the data after time-series block cropping with the soil water content label samples one by one, so that each time-series input sample has a time-series label and a class label; performing data augmentation on the time-series input samples with time-series labels and class labels.

[0025] In addition, an optional solution is that the input samples in the model sample database are multi-source time-series remote sensing data after processing, and the corresponding true label is the soil water content data at the next time step of the time-series node; among them, the production method of the true label includes:

[0026] S201: Regionally match the time-series input data with the soil moisture dataset representing the corresponding ground range according to the longitude and latitude information in the image;

[0027] S202: Obtain the soil moisture data at the next time node (Tn) of the corresponding region as the time-series label according to the time-series node information of the time-series input data (T0, T1, T2…Tn-1);

[0028] S203: Divide the corresponding time-series input samples into arid region samples, semi-arid samples, humid samples, and humid region samples according to the average precipitation meteorological data on the time-series time nodes (T0-Tn-1) of the time-series input data region, and construct regional class label information;

[0029] Match and align the inputs and labels of all samples in sequence according to the data processing method of S201 to S203. The final time-series input sample has a class label information reflecting the average wet-dry condition at the sample time-series node, and a time-series label reflecting the soil water content at the next time step.

[0030] On the other hand, the present invention also provides a geological disaster potential investigation system assisted by a multi-source time-series remote sensing soil water content prediction technology, which is used to conduct geological disaster potential investigations by using the geological disaster potential investigation method assisted by the multi-source time-series remote sensing soil water content prediction technology as described above. The system includes:

[0031] A basic data acquisition unit, which is used to preprocess the acquired multi-source time-series remote sensing data associated with the soil water content to obtain model basic data;

[0032] A database construction unit, which is used to construct a model sample database based on the model basic data;

[0033] A model training unit, which is used to optimize and train a pre-constructed deep learning model for predicting soil water content by using the sample database to obtain an initialized model with optimal weight model parameters; wherein, during the training process, the model is subjected to deep supervision training by using sample category information and supervised training by using sample time-series information, and the randomly initialized parameters are optimized and adjusted, and the model weights showing the best performance on the validation data set are saved;

[0034] A prediction unit, which is used to use the initialized model with the optimal weight model parameters to predict the soil water content at the next time step of the input multi-source time-series remote sensing data of the area to be predicted, and obtain the prediction result of the soil water content of the area;

[0035] A potential investigation unit, which is used to determine the change rate and second-order change rate of the soil water content according to the prediction result of the soil water content of the area; and determine the size of the geological disaster potential risk of the area according to the change rate and second-order change rate in combination with other indicators.

[0036] As can be seen from the above technical solutions, the method and system for geological disaster potential investigation assisted by the multi-source time-series remote sensing soil water content prediction technology provided by the present invention first predict the soil water content based on the selected multi-source time-series remote sensing data and the pre-constructed deep learning model for soil water content prediction, obtain the regional soil water content prediction results, and then comprehensively determine the risk level of geological disaster potential in the region according to the obtained regional soil water content prediction results in combination with other relevant indicators (such as regional deformation obtained by InSAR technology, optical texture features, etc.). It can not only overcome the defects of high measurement cost and low efficiency existing in the traditional soil moisture content monitoring scheme, but also can comprehensively analyze by combining the regional soil water content prediction results, the regional deformation obtained by InSAR technology, and the regional spectral texture provided by optical images, etc., so as to improve the accuracy of geological disaster potential investigation using relatively single remote sensing data.

[0037] To achieve the above and related purposes, one or more aspects of the present invention include the features that will be described in detail later. The following description and the accompanying drawings detail certain exemplary aspects of the present invention. However, these aspects only indicate some of the various ways in which the principles of the present invention can be used. In addition, the present invention aims to include all these aspects and their equivalents. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] By referring to the following description in conjunction with the accompanying drawings, and with a more comprehensive understanding of the present invention, other objects and results of the present invention will become more apparent and easier to understand. In the drawings:

[0039] Figure 1 FIG. is a schematic flow chart of a method for geological disaster potential investigation assisted by multi-source time-series remote sensing soil water content prediction technology according to an embodiment of the present invention;

[0040] Figure 2 FIG. is the overall technical route of a method for geological disaster potential investigation assisted by multi-source time-series remote sensing soil water content prediction technology according to an embodiment of the present invention;

[0041] Figure 3 FIG. is a schematic diagram of the process of adaptive weight fusion for realizing multi-source data according to an embodiment of the present invention;

[0042] Figure 4 FIG. is a schematic diagram of the correspondence between time-series sample data X and two types of labels according to an embodiment of the present invention;

[0043] Figure 5 FIG. is a schematic diagram of the convolutional neural network structure integrating an attention mechanism according to an embodiment of the present invention;

[0044] Figure 6 FIG. is a schematic flow chart of the process of obtaining the deformation result of the target area by InSAR technology according to an embodiment of the present invention;

[0045] Figure 7 It is a schematic diagram of the overall process for investigating geological disaster potential points according to an embodiment of the present invention.

[0046] In all the drawings, the same reference numerals indicate similar or corresponding features or functions. Detailed implementation manners

[0047] In the following description, for the purpose of illustration, in order to provide a comprehensive understanding of one or more embodiments, many specific details are set forth. However, it is obvious that these embodiments can also be implemented without these specific details. In other instances, well-known structures and devices are shown in block diagram form for the convenience of describing one or more embodiments.

[0048] Aiming at the problems of the traditional soil moisture content monitoring with a small scale and long laboratory determination time, large human and physical costs required, and difficulty in meeting the requirements of large-scale and rapid update of soil water content in the geological disaster potential investigation scenario, the present invention provides a method and system for investigating geological disaster potential assisted by multi-source time-series remote sensing soil water content prediction technology, which is used for predicting soil water content, further processing the subsequent obtained soil water content factors, and incorporating them into the process of geological disaster potential investigation to achieve multi-source high-precision investigation of geological disaster potential. This method first proposes a model for predicting soil water content change information based on multi-source time-series remote sensing data, obtains the predicted value of soil water content at future time steps in the region through this model, and then comprehensively analyzes this information after processing with the regional deformation obtained through InSAR technology and the regional spectral texture provided by optical images, etc., to improve the accuracy of investigating geological disaster potential with relatively single remote sensing data.

[0049] The following will describe in detail the specific embodiments of the present invention with reference to the accompanying drawings.

[0050] To illustrate the method for investigating geological disaster potential assisted by multi-source time-series remote sensing soil water content prediction technology provided by the present invention, Figure 1 shows the flow of the method for investigating geological disaster potential assisted by multi-source time-series remote sensing soil water content prediction technology according to an embodiment of the present invention.

[0051] As Figure 1 shown, the method for investigating geological disaster potential assisted by multi-source time-series remote sensing soil water content prediction technology provided by the present invention includes:

[0052] S1: Preprocess the acquired multi-source time-series remote sensing data associated with soil water content to obtain model basic data;

[0053] S2: Construct a model sample database based on the model basic data;

[0054] S3: Optimize and train the pre - constructed deep learning model for predicting soil water content using the said sample database to obtain a model parameter initialization model with optimal weights. Among them, during the training process, deep supervision training is carried out on the model using sample category information and supervision training is carried out on the model using sample time - series information, optimize and adjust the randomly initialized parameters, and save the model weights that perform best on the validation data set.

[0055] S4: Use the model parameter initialization model with optimal weights to predict the soil water content at the next time step of the input multi - source time - series remote sensing data of the area to be predicted, and obtain the prediction result of the soil water content of the said area.

[0056] On this basis, after predicting the soil water content of the area to be predicted, it further includes step S5: Determine the change rate and second - order change rate of the soil water content according to the prediction result of the soil water content of the said area; Determine the size of the geological disaster hidden danger risk of the said area according to the change rate and second - order change rate.

[0057] Based on multi - source time - series remote sensing data, the present invention designs and constructs a hybrid deep learning model structure for time - series prediction of regional soil water content, and then applies the prediction result of the soil water content to the investigation and survey of regional geological disaster hidden danger points, and combines multi - source remote sensing data to improve the investigation accuracy of geological disaster hidden danger points. Figure 2 This is the overall technical route of the geological disaster hidden danger investigation method assisted by the multi - source time - series remote sensing soil water content prediction technology provided by the present invention.

[0058] As Figure 2 shown, soil water content prediction is the premise for geological disaster hidden danger investigation. During the process of soil water content prediction:

[0059] The first step: Obtain multi - source time - series remote sensing data, and perform necessary processing on it including geometric radiation correction, index analysis, sampling interpolation, and multi - source fusion and other steps to prepare data for subsequent steps.

[0060] The second step: Perform block cropping and data augmentation on the data obtained and processed in the first step, and correspond it one - by - one with the soil water content label samples to construct a sample library for deep learning model training.

[0061] The third step: Construct a deep learning model for soil water content prediction. In a specific embodiment of the present invention, this model is composed of a convolutional neural network (CNN) and an attention network (Transform) in combination, and randomly initialize its model parameters.

[0062] Fourth step: Using the constructed sample library, optimize and train the designed model in the third step. During the training process, use the sample category information to perform deep supervised training on the model and use the sample time series information to perform supervised training on the model. Optimize and adjust the randomly initialized parameters, and save the model weights that perform best on the validation dataset.

[0063] Fifth step: Initialize the model based on the model parameters with the optimal weights, input multi-source time series remote sensing data, and use the model to predict the soil water content at the next time step to obtain the prediction result of the soil water content.

[0064] Sixth step: After obtaining the prediction result of the regional soil water content, combine multi-source remote sensing data including optical and SAR remote sensing images. First, perform deformation analysis on the region, and then further perform geological disaster feature analysis and extraction on key regions. In this process, combine the soil water content information obtained by prediction to improve the inspection accuracy of regional geological disaster hidden danger points.

[0065] The following will provide a more detailed technical description of each technical detail in the above steps.

[0066] The present invention uses multi-source time series remote sensing data and a hybrid deep learning model to perform the time series prediction task of soil moisture content. Remote sensing technology realizes the differentiation of different ground objects and the inversion of refined relevant parameters by receiving and analyzing the different electromagnetic wave reflection curves of ground objects.

[0067] Among them, when using remote sensing technology to obtain multi-source data, since soils with different water contents have different reflection curves for electromagnetic waves of the same band, and soils with the same water content also have different reflection or scattering curves for electromagnetic waves of different bands. Therefore, in the present invention, multi-source remote sensing data (different electromagnetic wave bands) are used for comprehensive modeling and analysis of the relationship between the electromagnetic spectrum and soil moisture content, which is of great help for obtaining high-precision soil moisture content information. Since soil samples with different water contents have different optical characteristics, dry soil samples tend to be sandy and yellowish, while wet sample soils are darker in color and lower in brightness, which are reflected as brighter or darker image features on optical images. The near-infrared band is sensitive to water content changes. Therefore, it is also of great value to use optical remote sensing images in the visible-near-infrared band. In addition, the entire physical process of obtaining soil scattering characteristics by SAR radar images is clear, the electromagnetic transmission model therein has strong certainty, and different polarization modes of radar waves are also of great significance for describing soil moisture content information from multiple angles and directions.

[0068] In addition, considering that the amount of soil water content is also highly correlated with the type of ground cover on the soil surface. For example, plant roots can help preserve water, while the water retention capacity in desert areas is relatively poor, and so on. Therefore, in the embodiments provided by the present invention, land cover utilization data is also used as an important type of auxiliary data for the inversion and prediction of soil water content.

[0069] In addition, in a specific embodiment of the present invention, considering that the growth state of vegetation is highly correlated with the water content in the soil and there is a certain causal relationship, the growth state of vegetation can also indirectly reflect the situation of soil water content. The prediction of soil water content can be assisted by calculating remote sensing indices that reflect the vegetation state. The data reflecting soil water content includes meteorological precipitation data, etc.

[0070] Therefore, in a specific embodiment provided by the present invention, the multi-source remote sensing raw data and related auxiliary data used in training the model include visible light remote sensing time series data in the (red - green - blue - near-infrared) band, synthetic aperture radar (SAR) time series data in the full polarization mode (HH, VV, VH, HV), land cover time series data in the corresponding area, and meteorological precipitation data, etc. Based on the direct correlation between the above raw remote sensing data and soil water content, the relationship between them can be directly modeled. Therefore, when preprocessing the multi-source time series remote sensing data, it can be preprocessed and used as part of the subsequent model input.

[0071] After obtaining the raw remote sensing data, first perform geometric correction and radiometric correction on the raw remote sensing data. Then, through the remote sensing index analysis process, calculate indirect indicators such as vegetation indices that can reflect soil water content, which are also used as one of the data sources for soil water content prediction. In the embodiments provided by the present invention, by performing index analysis on optical remote sensing images, the normalized difference vegetation index (NDVI) is calculated as a derivative data from the raw remote sensing data to assist in the inversion and prediction of soil water content.

[0072] After that, perform unified sampling and interpolation operations on the obtained different data sources (optical remote sensing data, radar remote sensing data, land cover data, etc.), the raw remote sensing data with various resolution sizes, and the remote sensing index data obtained by index calculation to unify the spatial dimensions and resolution sizes among different data. Perform normalization processing on the data, specifically including performing Z-Score normalization operation to adjust all data into the form of a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0073] Specifically, for the raw optical and radar remote sensing data, it is necessary to perform:

[0074] 1. Radiometric correction: Eliminate radiometric distortion caused by factors such as sensor characteristics, atmospheric conditions, and changes in solar elevation angle, ensuring that images acquired by different sensors at different acquisition times have consistent radiometric characteristics, facilitating subsequent analysis of multi-source and multi-temporal data.

[0075] 2. Geometric correction: Correct geometric deformations of images caused by factors such as the movement of the sensor platform, terrain undulation, earth curvature, and atmospheric refraction during image acquisition, accurately positioning the pixels of the image in the geographic coordinate system, which is of great significance for the matching and precise overlay of subsequent multi-source and multi-temporal remote sensing images.

[0076] 3. Remote sensing index calculation:

[0077] After radiometric correction and geometric correction, further perform feature analysis on the optical image and calculate the corresponding remote sensing index. The remote sensing index used for soil water content inversion prediction in this embodiment is the Normalized Difference Vegetation Index (NDVI), and the calculation formula is as follows:

[0078]

[0079] As shown in the above formula (1), it is the calculation formula of the Normalized Difference Vegetation Index NDVI, where NIR and R respectively represent the near-infrared band and the red band of the optical remote sensing image. The Normalized Difference Vegetation Index (NDVI) can be calculated by the ratio of subtracting the red band from the near-infrared band to the sum of the near-infrared band and the red band. This index reflects the current state of vegetation.

[0080] The NDVI Normalized Difference Vegetation Index used in this embodiment reflects the growth state of vegetation at the current time node and has a certain relationship with soil water content. The normalized vegetation remote sensing index can indirectly reflect the situation of soil water content by analyzing the growth state of vegetation, which is of great help for the inversion prediction of soil water content.

[0081] 4. Sampling interpolation and normalization processing:

[0082] In order to perform meaningful multi-source data fusion processing on the above data subsequently, it is necessary to first perform unified sampling interpolation on data of different size resolutions. The interpolation method used in this embodiment is the bilinear interpolation method, achieving consistency in spatial size and resolution between optical remote sensing data, radar remote sensing data, land cover data, and remote sensing index data through sampling interpolation.

[0083] After that, perform normalization processing on data from different data sources to unify the value ranges and data distributions of different source data for subsequent meaningful weighted fusion processes.

[0084] By means of Z-Score normalization, the data distributions from different data sources are unified. The Z-Score normalization is shown in the following formula:

[0085]

[0086] Where X is the data to be normalized. In this embodiment, the data to be normalized includes optical remote sensing data, radar remote sensing data, land cover data, and the derived NDVI index after min-max normalization. μ and σ are the mean and standard deviation of the above data respectively. Through the processing of the above formula, all kinds of different source data are Z-score standardized to unify the data distribution.

[0087] After the above processing, the value ranges and data distributions of different source data are unified.

[0088] After completing the unification of the data spatial dimensions and the normalization of the data distribution, the consistency of the spatial dimensions and data distributions of different data can be achieved. Then, the fusion of multi-source data is carried out. In a specific embodiment of the present invention, the method adopted for the fusion of multi-source data is: by assigning different weights to different data sources, the fusion of multi-source data is realized by using the method of band weighted summation, as shown in the following formula (3):

[0089]

[0090] Where, X1, X2, X3, X index respectively represent optical remote sensing data, SAR radar data, land cover type data, and remote sensing index obtained through index analysis; a, b, c, d are their respective proportional weights. The size of the proportional weight parameters of different data sources is specifically adaptively determined by a learnable model weight parameter. During the training process of this learnable model, according to the size of the loss function and the optimization situation of the optimizer, the importance and contribution degree of each part of the data source to the soil moisture content prediction task are determined, and finally the weights of a, b, c, d are assigned.

[0091] In addition, in this embodiment, the a, b, c, d weight parameters of different source data are also restricted by regularization with prior knowledge to artificially limit and adjust the contribution ratio range of each different data source, so as to play a role in highlighting or restricting the importance degree of a certain type of data in the prediction of soil water content. As shown in (b) and (c) in the above formula 3, the regularization rule in this embodiment is:

[0092] First of all, it is necessary to ensure that a + b + c + d = 1, so that the sum of the weights of each source data is equal to 1;

[0093] Secondly, considering the remote sensing index data (X index) is calculated from the optical remote sensing data (X1), so X derived from X1 index (Remote sensing index) The weight d should not exceed the weight a of X1 (optical remote sensing data) to ensure that the information in the original data makes a greater contribution.

[0094] Figure 3 shows a schematic diagram of the adaptive weight fusion process of multi-source data according to an embodiment of the present invention. As Figure 3 shown is the data fusion process of a time series node in the above embodiment (taking the T0 moment as an example). Subsequently, the data of other time series nodes are processed in the same steps, and finally the overall time series sample data X is obtained.

[0095] After completing the preprocessing and fusion of multi-source data, the processed data is further used to construct a high-quality deep learning training sample library.

[0096] In the optimization process based on the deep learning supervision model, in order to obtain accurate soil water content prediction results, high-quality training samples are required to train the model. The objective of this embodiment is to predict the soil water content information at the next time step of the time series node based on time series multi-source remote sensing data. Based on the above tasks, the input samples in the sample library are the processed multi-source time series remote sensing data X, and the corresponding true label is the soil water content data at the next time step of the time series node. In the production of true labels, considering the difficulty of collecting real soil water content data, the publicly available high-quality China Soil Humidity / Soil Moisture Dataset is used as the time series label data Y series . In addition, in this embodiment, each time series input sample X has a class sample label Y in addition to the time series label class . In this embodiment, the correspondence between the input time series sample data X and the two types of labels is as Figure 4 shown.

[0097] In the process of constructing the deep learning training sample library, first, the time series input data is regionally matched with the soil humidity dataset representing the corresponding ground range according to the latitude and longitude information in the image. Secondly, the preprocessed data from T0 to T n is arranged in time series to form the time series sample input X; according to the time series node information of the time series input data (T0, T1, T2...T n-1 ), the soil humidity data of the next time node (T n ) of the corresponding region is obtained as the time series label. In addition, according to the time series time node corresponding to the time series input data region (T0 - T n-1) The average precipitation meteorological data on is used to set a precipitation threshold to construct class labels for the entire time series input. The corresponding time series sample data X is divided into arid area samples, semi-arid samples, humid samples, and humid area samples to construct regional class label information. In this embodiment, the precipitation threshold division is set as shown in the following table:

[0098] Sample Threshold Classification Table

[0099] Classification mm / month Arid sample <10 Semi-arid sample 10–25 Humid sample 25–100 Moist sample 100>

[0100] After that, the soil water content data product at the T n+1 th moment is obtained from the soil water content product corresponding to the sample area in the above manner as the time series label. The inputs and labels of all samples are matched and aligned in sequence. The final time series input sample has a class label information reflecting the average dry-wet condition at the sample time series node, and a time series label reflecting the soil water content at the next time step. As shown in the schematic diagram of input X and label Y, the final training sample in this embodiment should include the time series input data X and the corresponding class label Y class and the time series label Y series . Further, a training sample library of the deep learning model is constituted by a number of sample pairs.

[0101] After the sample library is constructed, data augmentation processing is performed according to the class labels and time series labels of the samples respectively to further enrich the diversity of the samples in the sample library.

[0102] Since there are two types of labels in this embodiment: class labels and time series labels, different data augmentation methods need to be used for different types of sample and label data. Among them, for the input data and class label data, the commonly used data augmentation methods in computer vision tasks are adopted in this embodiment, including methods such as randomly simulating light intensity, rotation, cropping, and contrast enhancement to expand the samples. For the input sample and time series label data pairs, methods such as time translation, injecting noise, and amplitude perturbation are used for data augmentation.

[0103] To achieve the high-precision soil water content prediction task based on multi-source remote sensing images, in a specific embodiment of the present invention, a hybrid deep learning model structure for soil moisture content prediction, which integrates a convolutional neural network (CNN) and an attention mechanism model (Transformer), is specially designed and constructed. Among them, the convolutional neural network is used to extract the features of the input data. The convolutional neural network has certain advantages in obtaining the features of raster remote sensing images. Then, the attention mechanism model is used to process the features to capture the change relationship of the features at the time series nodes. Finally, the soil water content time series prediction result is output through the feedforward neural network and the fully connected layer. Figure 5Shows a fused convolutional neural network structure according to an embodiment of the present invention.

[0104] As Figure 5 shown, the input of the hybrid deep learning model for soil moisture content prediction is time-series multi-source remote sensing raster data (T0 - Tn), and the final task is to perform time-series prediction of soil water content at time Tn+1, including a convolutional module responsible for raster feature extraction and a multi-head attention module responsible for time-series feature extraction.

[0105] Among them, the convolutional module responsible for raster feature extraction includes multiple basic convolutional modules with shared weights, which are composed of a convolutional layer, a batch normalization layer, a downsampling layer, and an activation function layer. In specific applications, the configurations of different modules can be modified according to the application scenario. The content that can be modified includes the number of basic modules used, the number of convolutional channels in each module, the downsampling ratio, etc. The convolutional module in this embodiment is responsible for extracting basic features such as optics, texture, and shape of the input data. Since there are significant commonalities in the basic features of the input samples at different time-series nodes, a strategy of shared weights is adopted to save training costs. In addition, for the differential features in the time-series input data, a branch module for multi-path feature processing is also constructed to capture the differential feature patterns therein. Finally, the feature fusion module is used to fuse the multi-branch features to obtain the features extracted by the convolutional module. And the sample class loss loss1 is calculated through the cross-entropy loss function using this feature and the class loss label of the sample.

[0106] After that, the features are flattened and position-encoded as the input of the time-series attention module. The time-series attention module in this embodiment of the present disclosure is composed of multiple multi-head attention modules, and a fully connected layer is used to implement the mapping of features to the output space.

[0107] In this hybrid deep learning model, the multi-head attention module is responsible for performing time-series feature analysis and processing tasks. The attention mechanism has a strong ability to capture global features and can handle long-distance feature dependence patterns in time-series input data, which is of great significance for time-series tasks with long input sequences. In addition, the multi-head attention module can capture long-distance feature dependence relationships in multiple different modes by setting the number of attention heads. Finally, a fully connected layer is used to map the features to the output space, and the mean squared error loss function (Mean Squared Error, MSE) is used to calculate the time-series loss loss2 between this output value and the sample time-series label. Finally, the total loss loss of the model is further calculated by summing loss1 and loss2.

[0108] After constructing the overall structure of the hybrid deep learning model, the model training process is carried out based on the constructed sample library. Through the Stochastic Gradient Descent (SGD) algorithm, the gradient information of the overall Loss of the model with respect to the network weight parameters is calculated. A small batch of data randomly sampled from the sample library is used to gradually optimize and update the model parameters through the backpropagation algorithm, and finally the training process of the deep learning model is completed. The method of randomly sampling small batches can reduce the computational amount and speed up the training speed.

[0109] When using the Stochastic Gradient Descent (SGD) algorithm in this embodiment, considering that the learning rate scheduling strategy is one of the key factors affecting model convergence during optimization training, a polynomial decay learning rate scheduling strategy is adopted to adjust the learning rate. The specific formula is as follows:

[0110]

[0111] In the above formula, p is the degree of the polynomial, t represents the current training epoch, t max is the maximum number of training epochs, η t and η0 are the learning rate of the current epoch and the initial learning rate respectively. In this embodiment, the polynomial decay learning rate scheduling strategy helps the model converge quickly in the initial stage of training and helps the model stabilize at a better solution in the later stage of training.

[0112] In this embodiment, the design model is trained based on the constructed sample library by the above method. After the total loss of the model converges and stabilizes without further decrease, the optimal model weights are saved for subsequent prediction.

[0113] After the training of the hybrid deep learning model is completed, the model weight parameters with the optimal accuracy on the validation set during the training process are saved. So that the soil water content in the unknown area can be predicted based on the model in the future.

[0114] In the subsequent prediction of soil water content, first initialize and construct the overall model structure, load the optimal model parameters obtained through model training, and start the forward propagation mode of the model. The data to be predicted is first preprocessed and used as the input of the hybrid deep learning model. After being processed layer by layer by the hybrid deep learning model, the prediction result of the soil water content is output.

[0115] Specifically, as an example, the specific steps for predicting the soil water content in the unknown area based on the above hybrid deep learning model are as follows:

[0116] First, obtain the original data of the area to be predicted, which is specifically the same as the input data in the training process, and perform the same preprocessing as in the training process. Secondly, construct the same model structure as in the training process, and load the optimal model weight parameters obtained through training optimization. Finally, use the preprocessed input to be predicted as the model input, feed it into the model and start the forward propagation process of the model, and obtain the prediction result of the soil water content of the area through model processing.

[0117] After obtaining the information of the regional soil water content through the above steps, it can be used as an important indicator to jointly verify the hidden geological disaster hazards with multi-source remote sensing data. First, obtain the surface deformation of the area through InSAR technology, and preliminarily analyze potential hidden geological disaster points in combination with the surface deformation information. Then, based on this, further verify the risk level of the hidden geological disaster points through information such as spectral texture in optical images and the soil water content information at the corresponding location.

[0118] Specifically, as an example, use the information of the obtained regional soil water content as an important indicator to jointly verify the hidden geological disaster hazards with multi-source remote sensing data, including:

[0119] S501: Obtain the surface deformation of the area through InSAR technology, and preliminarily analyze potential hidden geological disaster points in combination with the surface deformation information. First, predict the soil water content value of the target area through the above steps, and obtain the predicted value of the soil water content at the next time step, denoted as S(t i ). Since in this embodiment, considering that in addition to the absolute value of the soil water content, there is also strong correlation information between the change rate of the soil water content and the hidden geological disaster hazards, the change value index of the soil water content at a certain time step is calculated through the following formula (5), that is, the change rate and the second-order change rate of the soil water content, which respectively reflect the change speed of the soil water content and the speed of change of the change speed at this time step.

[0120]

[0121] In formula 5(a), V i is the change rate of the soil water content within one time step, and S(t i ) and S(t i-1 ) are the soil water content values at the t i th and t i-1 th moments respectively. Through formula 5(a), the change rate V i-1 of the soil water content within the time period [t i , t i can be obtained. Then, further calculate the change rate index A i-1 of the soil water content rate within the time period [t i according to formula 5(b). i .

[0122] Among them, V i and A i respectively reflect the change rate of soil water content and the rate of change of the change rate at this time step. These two indicators more comprehensively describe the change of soil water content in the region. When the value of the change rate of soil water content V i is large, or when the change rate changes sharply (that is, A i value is large), it usually indicates that there is a high risk of geological disasters in this area. In the following of this disclosure, the absolute value S(t i ) of the soil water content at time i obtained by model prediction, the change rate V i of the soil water content, and the change of the change rate of the soil water content A i are further combined with other remote sensing data to implement the investigation of geological disaster hidden danger points.

[0123] After determining the change rate and the second-order change rate of the soil water content, further determine the size of the geological disaster hidden danger risk of the region according to the change rate and the second-order change rate. Considering that before the formation of geological disaster hidden dangers, there is often a certain ground deformation situation, so after determining the change rate and the second-order change rate of the soil water content, it is also necessary to re-divide the high-risk areas and low-risk areas in the region according to the change rate and the second-order change rate of the soil water content in the region. Specifically, the deformation result of the target area can be obtained by InSAR technology, and the target area is re-divided into risk areas according to the deformation result information. Figure 6 shows the process of obtaining the deformation result of the target area by InSAR technology according to an embodiment of the present invention, as Figure 6 shown, the process of obtaining the deformation result of the target area by InSAR technology includes:

[0124] First, register the SAR image data of the area and generate an interference pair, then perform interference processing on the interference pair to obtain a differential interference map. The differential interference map contains the phase information of the surface deformation. Subsequently, filter out the noise through phase filtering, and then perform phase unwrapping on the wrapped phase. After unwrapping, convert the unwrapped phase into the true surface deformation amount, and finally obtain the deformation map of the area after geocoding. The regional deformation map reflects the basic situation of the deformation in the region. Areas with certain deformations often have a high risk of forming geological disasters, and vice versa, the risk is small.

[0125] After dividing the hidden danger points in the region into high-risk areas and low-risk areas according to the deformation information of the region obtained by InSAR technology, further divide the hidden danger points in the region into high-risk areas, medium-high-risk areas, medium-risk areas, low-risk areas and risk-free areas through re-division of risk areas.

[0126] In addition, in a specific embodiment of the present invention, considering that the formation of geological disaster hazards often leads to certain changes in the vegetation around the hazard points, and that human engineering activities and regional soil types will also affect the formation of geological disaster hazard points. Therefore, the normalized difference vegetation index is further calculated through optical images to indicate the regional vegetation change situation. And the regional human engineering activities and soil types are interpreted through optical images.

[0127] Finally, the decision tree is used to analyze the regional deformation map, and the multi-source regional soil moisture content information and spectral texture features are used to detect the geological disaster hazard points in the region.

[0128] Figure 7 shows the overall process of detecting geological disaster hazard points according to an embodiment of the present invention. As Figure 7 shown, the overall process of the geological disaster hazard detection process based on soil water content information is as follows:

[0129] First, regional deformation information is obtained. Since the occurrence of geological disaster hazards is often accompanied by ground deformation, the hazard points in the region can be divided into high-risk areas and low-risk areas according to the threshold. Subsequently, considering the role of soil water content change in the formation of geological disaster hazards, two different sets of threshold indicators are set in the high-risk area and the low-risk area for further dividing the high-risk area and the low-risk area. Finally, through the regional optical image, by calculating the normalized index (NDVI), the soil type and human engineering activities are identified, etc., to further detect the level of regional geological disaster hazards, and finally the detection result of regional hazard points is obtained.

[0130] In specific applications, the thresholds a, (b1, b2, b3), and (c1, c2, c3) need to be set in sequence by combining the regional historical geology and hydrology and experience and other conditions. Among them, the threshold a is the regional deformation amount threshold, and (b1, b2, b3) is a set of soil water content-related thresholds for the high-risk area, which are the threshold of the absolute value of the soil water content, the threshold of the soil water content change rate, and the threshold of the change speed of the soil water content change rate respectively. (c1, c2, c3) is a set of soil water content thresholds for the low-risk area, which are also the threshold of the absolute value of the soil water content, the threshold of the soil water content change rate, and the threshold of the change speed of the soil water content change rate respectively.

[0131] When the regional deformation value is greater than the threshold a, it is divided into a high-risk area and a low-risk area. Subsequently, the region is further verified according to the soil water content index thresholds (b1, b2, b3) of the high-risk area and the soil water content index thresholds (c1, c2, c3) of the low-risk area. When judging the high-risk area samples against the soil water content index thresholds (b1, b2, b3), if all three corresponding indicators are greater than the thresholds (b1, b2, b3), it is classified as a high-risk hidden danger point; if two indicators are greater than the thresholds, it is classified as a medium-high risk area; when any one indicator exceeds the threshold, it is classified as a medium-risk area; if there is no indicator exceeding the threshold, it is classified as a medium-low risk area. Additionally, in the low-risk area, the same verification is carried out according to the soil water content index thresholds (c1, c2, c3) of the low-risk area, and the regional samples are successively classified into medium-risk, medium-low risk, low-risk, and risk-free areas.

[0132] In addition, in the present invention, it is also considered that geological disaster hidden danger areas are often accompanied by changes in surface covered vegetation, the risk of geological disasters occurring in areas with different soil types is also different, and human engineering activities will exacerbate the risk of forming geological disaster hidden dangers. Considering the above factors, through the spectral texture information provided by high-resolution optical images, the normalized difference vegetation index is calculated to identify factors related to the formation of geological disaster hidden dangers, such as soil types and traces of human engineering activities.

[0133] The above factors are used to further comprehensively verify and judge the classified medium-high risk areas. The specific method is as follows: if the NDVI index calculated multiple times has significant changes, there is a greater risk of geological disaster hidden dangers; if more than one trace of human engineering activities is identified, the risk level of geological disaster hidden dangers is correspondingly increased. Finally, the comprehensive investigation of geological disaster hidden danger points in the region is realized by combining regional soil information.

[0134] Corresponding to the above geological disaster hidden danger investigation method assisted by multi-source time-series remote sensing soil water content prediction technology, the present invention also provides a geological disaster hidden danger investigation system for multi-source time-series remote sensing soil water content prediction assistance technology, which is used to investigate geological disaster hidden dangers by using the above method. The system includes:

[0135] A basic data acquisition unit, which is used to preprocess the acquired multi-source time-series remote sensing data associated with soil water content to obtain model basic data;

[0136] A database construction unit, which is used to construct a model sample database based on the model basic data;

[0137] A model training unit for optimizing and training a pre-constructed deep learning model for predicting soil water content using the sample database to obtain an initialized model with the optimal weight model parameters. During the training process, deep supervision training is performed on the model using sample category information, and supervision training is performed on the model using sample time series information. The randomly initialized parameters are optimized and adjusted, and the model weights that perform best on the validation dataset are saved.

[0138] A prediction unit for predicting the soil water content at the next time step of the input multi-source time series remote sensing data of the area to be predicted using the initialized model with the optimal weight model parameters to obtain the prediction result of the soil water content of the area.

[0139] A hidden danger investigation unit for determining the change rate and second-order change rate of the soil water content according to the prediction result of the soil water content of the area; and determining the size of the geological disaster hidden danger risk of the area according to the change rate and second-order change rate.

[0140] For the embodiments of the geological disaster hidden danger investigation system assisted by the multi-source time series remote sensing soil water content prediction technology provided by the present invention, since it is basically similar to the embodiments of the geological disaster hidden danger investigation method assisted by the multi-source time series remote sensing soil water content prediction technology, for the relevant parts, refer to the partial description of the method embodiments, and details will not be repeated here.

[0141] As can be seen from the specific embodiments provided by the present invention above, the geological disaster hidden danger investigation method and system assisted by the multi-source time series remote sensing soil water content prediction technology provided by the present invention predict the soil water content through the selected multi-source time series remote sensing data and the pre-constructed deep learning model for predicting soil water content, obtain the prediction result of the soil water content of the area, and then determine the size of the geological disaster hidden danger risk of the area according to the obtained prediction result of the soil water content of the area. It can not only overcome the defects of high measurement cost and low efficiency existing in the traditional soil moisture content monitoring scheme, but also can comprehensively analyze by combining the prediction result of the regional soil water content, the regional deformation obtained by the InSAR technology, and the regional spectral texture provided by the optical image, etc., improve the accuracy of geological disaster hidden danger investigation using relatively single remote sensing data, and realize the rapid investigation of regional geological disaster hidden danger points with the help of multi-source remote sensing data and deep learning technology.

[0142] As described above, the geological disaster hidden danger investigation method and system assisted by the multi-source time series remote sensing soil water content prediction technology according to the present invention are described by way of example with reference to the drawings. However, those skilled in the art should understand that various improvements can be made to the above-mentioned geological disaster hidden danger investigation method and system assisted by the multi-source time series remote sensing soil water content prediction technology of the present invention without departing from the content of the present invention. Therefore, the protection scope of the present invention should be determined by the content of the appended claims.

Claims

1. A method for investigating potential geological disaster hazards assisted by a multi-source time-series remote sensing soil water content prediction technology, characterized in that, Including: Preprocessing the acquired multi-source time-series remote sensing data associated with soil water content to obtain model basic data; Constructing a model sample database based on the model basic data; Using the sample database to optimize and train a pre-constructed deep learning model for predicting soil water content to obtain an initialized model with optimal weight model parameters; wherein, during the training process, deep supervision training is performed on the model using sample category information and supervision training is performed on the model using sample time-series information, the randomly initialized parameters are optimized and adjusted, and the model weights that perform optimally on the validation dataset are saved; Using the initialized model with the optimal weight model parameters to predict the soil water content at the next time step of the multi-source time-series remote sensing data of the area to be predicted as input, and obtaining the prediction result of the soil water content of the area; According to the prediction result of the soil water content of the area, determining the change rate and the second-order change rate of the soil water content; according to the change rate and the second-order change rate, determining the magnitude of the geological disaster hazard risk of the area.

2. The geological disaster hidden danger investigation method assisted by the multi-source time-series remote sensing soil water content prediction technology according to claim 1, characterized in that, The multi-source time-series remote sensing data includes optical remote sensing time-series data, synthetic aperture radar time-series data, corresponding area land cover time-series data, and data reflecting soil moisture content.

3. The geological disaster hidden danger investigation method assisted by the multi-source time-series remote sensing soil water content prediction technology according to claim 2, wherein The preprocessing includes geometric correction, radiometric correction, index analysis, sampling interpolation, normalization, and multi-source fusion.

4. The method for investigating geological disaster hazards assisted by the multi-source time-series remote sensing soil water content prediction technology according to claim 3, wherein, The index analysis indirectly reflects the soil water content situation by analyzing the vegetation growth status through the normalized difference vegetation index; wherein, the calculation formula of the normalized difference vegetation index is as follows: wherein, NDVI represents the normalized difference vegetation index, and NIR and R respectively represent the near-infrared band and the red band of the optical remote sensing image.

5. The geological disaster hidden danger investigation method assisted by the multi-source time-series remote sensing soil water content prediction technology as described in claim 3, characterized in that The normalization is performed in the way of Z-Score normalization to unify the value range and data distribution from different data sources; wherein, the Z-Score normalization is shown in the following formula: wherein, Z represents the normalized data, X represents the data to be normalized, and the data to be normalized includes the optical remote sensing data, radar remote sensing data, surface cover data, and the derived normalized difference vegetation index after maximum-minimum normalization processing; μ and σ respectively represent the mean and standard deviation of the data to be normalized.

6. The geological disaster potential hazard investigation method assisted by the multi-source time-series remote sensing soil water content prediction technology according to claim 5, wherein The multi-source fusion is used to fuse the normalized data by the way of weighted fusion addition, and the formula is as follows: Among them, X1, X2, X3, and X index respectively represent optical remote sensing data, SAR radar data, land cover type data, and remote sensing indices obtained through index analysis; a, b, c, and d are the respective proportional weights of X1, X2, X3, and X index respectively.

7. The method for investigating hidden geological disaster hazards assisted by the multi-source temporal remote sensing soil water content prediction technology according to claim 1, characterized in that Constructing a model sample database based on the model basic data, including: Performing time-series block cropping on the model basic data, and corresponding the time-series block cropped data with the soil water content label samples one by one, so that each time-series input sample has a time-series label and a category label; Performing data augmentation on the time-series input samples with time-series labels and category labels.

8. The geological disaster hidden danger investigation method assisted by the multi-source time-series remote sensing soil water content prediction technology according to claim 7, characterized in that, The input samples in the model sample database are the processed multi-source time-series remote sensing data, and the corresponding true label is the soil water content data at the next time step of the time-series node; wherein, the production method of the true label includes: S201: Regionally matching the time-series input data with the soil moisture dataset representing the corresponding ground range according to the longitude and latitude information in the image; S202: Obtain the soil moisture data of the next time node (Tn) of the corresponding area as the time series label according to the time series node information of the input data (T0, T1, T2... Tn-1) in time series; S203: According to the average precipitation meteorological data at the time series time nodes (T0 - T n-1 ) of the time series input data region, divide the corresponding time series input samples into arid region samples, semi-arid samples, humid samples, and humid region samples, and construct regional category label information; Match and align the inputs and labels of all samples in sequence according to the data processing methods of S201 to S203. The final time series input sample has a class label information reflecting the average dry-wet condition at the time series node of the sample, and a time series label reflecting the soil water content at the next time step.

9. The method for investigating hidden geological disaster hazards assisted by the multi-source time-series remote sensing soil water content prediction technology according to claim 8, wherein During the training process of the soil water content prediction deep learning model, Calculate the cross-entropy class loss Loss1 through the class label of the input sample and the output features of the convolutional neural network; and calculate the time series loss Loss2 by using the time series label of the input sample and the time series features output by the attention mechanism model; Take the sum of Loss1 and Loss2 as the overall loss of the soil water content prediction deep learning model; Through the stochastic gradient descent method, calculate the gradient information of the overall loss of the soil water content prediction deep learning model with respect to the network weight parameters, and gradually optimize the network parameters through the backpropagation algorithm, and finally complete the training of the soil water content prediction deep learning model.

10. A geological disaster hidden danger investigation system assisted by a multi-source time-series remote sensing soil water content prediction technology, characterized in that, A method for geological disaster hidden danger investigation using the multi-source time series remote sensing soil water content prediction technology assisted as described in any one of claims 1 to 9, the system includes: A basic data acquisition unit, configured to preprocess the acquired multi-source time series remote sensing data associated with the soil water content to obtain model basic data; A database construction unit, configured to construct a model sample database based on the model basic data; A model training unit, configured to optimize and train a pre-constructed soil water content prediction deep learning model using the sample database to obtain an initialized model with the optimal weight model parameters; wherein, during the training process, deep supervision training is performed on the model using the sample class information and supervision training is performed on the model using the sample time series information, and the randomly initialized parameters are optimized and adjusted, and the model weights that perform best on the validation data set are saved; A prediction unit, configured to use the initialized model with the optimal weight model parameters to predict the soil water content at the next time step of the input multi-source time series remote sensing data of the area to be predicted, and obtain the prediction result of the soil water content of the area; A hidden danger investigation unit, configured to determine the change rate and the second-order change rate of the soil water content according to the soil water content prediction result of the area; and determine the size of the geological disaster hidden danger risk of the area according to the change rate and the second-order change rate.

Images