Disclosure of Invention
The invention aims to overcome the defects and provide a processing method for a multi-mode data intermediate layer fusion full-connection geological map prediction model, which solves the technical problem of how to deeply mine geological evidence data.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: a processing method for a multi-modal data intermediate layer fusion full-connection geological map prediction model is characterized by comprising the following steps:
performing complement and normalization processing on the evidence data;
reading data information from the processed evidence data according to the training sampling point coordinates;
acquiring basic data and remote sensing data of training sampling points;
and judging whether the data is read into the memory once, if so, directly inputting the model training, and if not, grouping the data from the data storage path to input the model training.
Further, the supplementing the data includes:
reading remote sensing data as a data matrix, and marking the data as A;
creating a data matrix consistent with the basic data processing range, wherein the initial values of data in the matrix are all 0, and the data matrix is marked as B;
reading data consistent with the basic data processing range row by row from the lower left corner of A and storing the data to the corresponding position in B;
if the upper half of the matrix in B has data of 0, the data is replaced by a first value which is not 0 in the column where the data is located, and if the lower half of the matrix in B has data of 0, the data is replaced by a last value which is not 0 in the column where the data is located;
if the left half part of the matrix in B has data of 0, the data is replaced by a first value which is not 0 in the row where the data is located, and if the right half part of the matrix in B has data of 0, the data is replaced by a last value which is not 0 in the row where the data is located;
b is evidence data after the complement.
Further, the normalization process employs an algorithm:
where X represents the data that the underlying data or telemetry data read on each channel or class of chemical elements, x_min represents the minimum value in the data matrix X, and x_max represents the maximum value in the data matrix X.
Further, the model includes:
input layer: the basic data and the remote sensing data of the training sampling points are simultaneously input into a full-connection hybrid input model through an input layer;
basic data feature extraction layer: the feature extraction layer is composed of three full-connection layers, full-connection operation is carried out through a full-connection neural network, the network features, namely basic data features, are extracted by utilizing weight values, and original data are converted into high-dimensional feature expression;
remote sensing data feature extraction layer: the feature extraction layer is composed of three full-connection layers, full-connection operation is carried out through a full-connection neural network, the network features, namely remote sensing data features, are extracted by utilizing weight values, and original data are converted into high-dimensional feature expression;
fusion layer: carrying out feature fusion on the basic data features and the remote sensing data features;
full tie layer: the method comprises the steps of obtaining common characteristics of basic data and remote sensing data in a high-dimensional space by using a fully connected neural network;
output layer: and outputting the probability values of the sampling points predicted as various types.
The beneficial effects of the invention are as follows:
the method is simple to realize, not only converts the original data into high-dimensional characteristic expression, but also performs characteristic fusion on the basic data characteristics and the remote sensing data characteristics, and utilizes the fully connected neural network to maximally acquire the commonality characteristics of the basic data and the remote sensing data in the high-dimensional space, thereby greatly improving the capability and the precision of the prediction object, and enabling the most basic geological object of the geological map, namely the filling unit lithology, to be maximally reflected and expressed in terms of the indexes such as the spatial distribution form, the distribution direction, the distribution position, the adjacent relation among the geological objects and the like. After the technology breaks through, the existing geological investigation working mode is thoroughly changed, and the novel geological mapping mode integrating geological knowledge map, geological big data and deep learning algorithm is innovated, so that the optimal geological route and geological map prediction cycle progressive advancing mapping mode is realized, and a fine-granularity prediction geological map is formed.
Detailed Description
Certain terms are used throughout the description and claims to refer to particular components. Those of skill in the art will appreciate that a hardware manufacturer may refer to the same component by different names. The description and claims do not take the form of an element differentiated by name, but rather by functionality. As used throughout the specification and claims, the word "comprise" is an open-ended term, and thus should be interpreted to mean "include, but not limited to. By "substantially" is meant that within an acceptable error range, a person skilled in the art is able to solve the technical problem within a certain error range, substantially achieving the technical effect. The description hereinafter sets forth the preferred embodiment for carrying out the present application, but is not intended to limit the scope of the present application in general, for the purpose of illustrating the general principles of the present application. The scope of the present application is defined by the appended claims.
Referring to fig. 1 to 5, a method for processing a multi-modal data middle layer fusion full-connection geological map prediction model according to the present invention includes:
step S101, performing complement and normalization processing on the data of the evidence;
step S102, reading data information from the processed evidence data according to the training sampling point coordinates;
step S103, basic data and remote sensing data of training sampling points are obtained;
step S104, judging whether the data is read into the memory once, if yes, directly inputting model training, and if no, inputting model training from the data storage path group.
In one embodiment, the supplementing the data includes:
reading remote sensing data as a data matrix, and marking the data as A;
creating a data matrix consistent with the basic data processing range, wherein the initial values of data in the matrix are all 0, and the data matrix is marked as B;
reading data consistent with the basic data processing range row by row from the lower left corner of A and storing the data to the corresponding position in B;
if the upper half of the matrix in B has data of 0, the data is replaced by a first value which is not 0 in the column where the data is located, and if the lower half of the matrix in B has data of 0, the data is replaced by a last value which is not 0 in the column where the data is located;
if the left half part of the matrix in B has data of 0, the data is replaced by a first value which is not 0 in the row where the data is located, and if the right half part of the matrix in B has data of 0, the data is replaced by a last value which is not 0 in the row where the data is located;
b is evidence data after the complement.
In one embodiment, the normalization process employs an algorithm:
where X represents the data that the underlying data or telemetry data read on each channel or class of chemical elements, x_min represents the minimum value in the data matrix X, and x_max represents the maximum value in the data matrix X.
In one embodiment, the model comprises:
input layer: the basic data and the remote sensing data of the training sampling points are simultaneously input into a full-connection hybrid input model through an input layer;
basic data feature extraction layer: the feature extraction layer is composed of three full-connection layers, full-connection operation is carried out through a full-connection neural network, the network features, namely basic data features, are extracted by utilizing weight values, and original data are converted into high-dimensional feature expression;
remote sensing data feature extraction layer: the feature extraction layer is composed of three full-connection layers, full-connection operation is carried out through a full-connection neural network, the network features, namely remote sensing data features, are extracted by utilizing weight values, and original data are converted into high-dimensional feature expression;
fusion layer: carrying out feature fusion on the basic data features and the remote sensing data features;
full tie layer: the method comprises the steps of obtaining common characteristics of basic data and remote sensing data in a high-dimensional space by using a fully connected neural network;
output layer: and outputting the probability values of the sampling points predicted as various types.
Method for processing spatial boundary consistency of various evidence data and supplementing evidence data
Because the data is multi-modal, and the source and processing mode of the data are various software, boundary data may not fill the whole frame area, data mismatch is caused, and blank areas (non-valued areas) appear, so as to fully utilize the used data, ensure that the predicted area can fill the whole frame area, and various evidence data space boundary consistency processes are required. In this example, 4 data of 1:50000 frames are taken as an example, and a method and a flow of data processing are described:
the data processing ranges of the evidence data after the four pictures are combined are recorded in the table 1, 20 ten thousand geochemical data and high magnetism 2500 are taken as basic data, remote sensing high three SAR data, a Digital Elevation Model (DEM) and remote sensing Landsat8 are taken as remote sensing data, and the processing ranges of the basic data and the remote sensing data are inconsistent. If the data processing ranges are inconsistent, the corresponding basic data and remote sensing data obtained according to the coordinate information of the sampling point may not coincide with the sampling point. As shown in fig. 3, the black area represents a data processing range, and fig. 3 (a) the data processing range is (4, 4); in fig. 3 (b), the data processing range is (5, 5), the white area is the position of the sampling point coordinate information, and under normal conditions, when the data is read, the position coordinate information of the white area in (a) is consistent and the data is correctly read when the data is read by taking the lower left corner as the origin and the position coordinate information of the white area in (a) and (b) is inconsistent when the data is read by taking the upper left corner as the origin and the data is read according to the sampling point coordinate information. To ensure consistency of data reading, the processing range of evidence data needs to be kept consistent.
TABLE 1 four-breadth-to-aggregate evidence data processing scope
Data
|
Data processing range
|
Geochemistry of 1:20 ten thousand
|
Wide 8114 pixels and high 7191 pixels
|
High magnetic_2500
|
Wide 8114 pixels and high 7191 pixels
|
ALOS SAR data
|
Wide 8112 pixels and high 7189 pixels
|
Digital Elevation Model (DEM)
|
Wide 8116 pixels and high 7190 pixels
|
Remote sensing Landsat8
|
Wide 8116 pixels and high 7190 pixels |
And carrying out data complement processing on the remote sensing data by taking the basic data as a reference, so that the processing range of the remote sensing data is consistent with the basic data.
The data complement process is as follows:
(1) reading remote sensing data as a data matrix, and marking the data as A;
(2) creating a data matrix consistent with the basic data processing range, wherein the initial values of data in the matrix are all 0, and the data matrix is marked as B;
(3) reading data consistent with the basic data processing range row by row from the lower left corner of A and storing the data to the corresponding position in B;
(4) if the upper half of the matrix in B has data of 0, the data is replaced by a first value which is not 0 in the column where the data is located, and if the lower half of the matrix in B has data of 0, the data is replaced by a last value which is not 0 in the column where the data is located;
(5) if the left half part of the matrix in B has data of 0, the data is replaced by a first value which is not 0 in the row where the data is located, and if the right half part of the matrix in B has data of 0, the data is replaced by a last value which is not 0 in the row where the data is located;
(6) b is evidence data after the complement.
(2) Geological evidence data normalization
Geological evidence relates to a variety of specialized data, the dimensions of each of which are also substantially inconsistent. It is therefore desirable to change a dimensionless expression to a dimensionless (physical unit without actual data) expression, which facilitates the comparison and weighting of different units or magnitudes of the indicators, a process called normalization. The normalization changes the data into the decimal between (0, 1) or (-1, 1), which is a linear transformation, can ensure that the data can not be invalid after normalization processing, and can improve the expressive force of the data and realize data enhancement. The invention adopts the following normalization formula:
where X represents the data that the underlying data or telemetry data read on each channel or class of chemical elements, x_min represents the minimum value in the data matrix X, and x_max represents the maximum value in the data matrix X.
Normalization processing is carried out on the evidence data, firstly, floating point overflow or underflow during calculation is prevented from being caused by an excessively large or excessively small numerical range; secondly, different numerical ranges can cause different importance of different attributes to the model, and equalization of training of the model by the different attributes is achieved after data are normalized; thirdly, after normalization processing is carried out on the data, the solving speed of gradient descent can be increased, namely the convergence speed of the model is improved.
Referring to fig. 4, the basic data in the evidence data is the same as the remote sensing data in the processing manner, and the normalization processing data flow is specifically described below by taking normalization processing Digital Elevation Model (DEM) data as an example:
(1) obtaining a tif format file of the digital elevation model data from a specified evidence data storage path;
(2) reading the tif format file into a data matrix by using an imread function under an io module in the image packet, and recording the data matrix as A and the data matrix as a numpy array form;
(3) solving a maximum value B and a minimum value C in A by using a max method and a min method in a numpy library;
(4) calculation ofThe result is a result after normalization of Digital Elevation Model (DEM) data, the result is marked as D, and D is in a data matrix format;
(5) the D is a npy format file, i.e., a numpy array format file, and is stored under the folder of normalized data, and is stored under the name "dem.
(3) Data packet storage
Because the radius of the second sampling is larger, the obtained data volume of the training sampling point is overlarge, the problem that the data cannot be read into the memory for model training at one time may exist, when the problem occurs, the basic data and the remote sensing data of the training sampling point are required to be firstly put into the hard disk for storage, and the data is read into the memory from the hard disk during model training. And if the memory is allowed, directly entering the fourth step.
The method for storing data in groups is as follows:
(1) in order to prevent data reading errors, firstly, storing basic data and remote sensing data of training set sampling points obtained by expanding various mapping units and lithology after secondary sampling, and storing the data packets in basic and remote sensing data folders (for example, file storage paths:/datasets/PRB/training sampling point data/basic (remote sensing) data/Qh 3al fourth series riverbed/interval_data_1 (remote_data_1). Npy) corresponding to each mapping unit and lithology; each npy file stores basic data or remote sensing data of training set sampling points of the geologic body;
(2) and (3) grouping and synthesizing the basic and remote sensing data of the various mapping units and lithology obtained in the step (1) to obtain final training sampling point data, storing the final training sampling point data in a corresponding folder (a file storage path such as/data/PRB/training sampling point data/basic (remote sensing) data/inter_data_1 (remote_data_1). Npy), and storing the basic data or remote sensing data of the training sampling points of each mapping unit and lithology participating in training in each npy file.
And according to the mode of grouping and storing the data, the basic data and the remote sensing data of the training sampling points are stored, the basic data and the remote sensing data are grouped and read from a storage path during model training, and the data are read into a memory for model training.
(4) Modeling of multi-modal data middle layer fusion full-connection geological map prediction model
(1) Multimode data middle layer fusion full-connection geological map prediction model structure
Multimodal fusion refers to the process of integrating information from two or more modalities to make predictions. In the prediction process, a single mode cannot generally contain all effective information required for generating an accurate prediction result, and the multi-mode fusion process fuses information from two or more modes, so that information supplementation is realized, the coverage range of information contained in input data is widened, the accuracy of the prediction result is improved, and the robustness of a prediction model is improved. Currently, there are three main fusion modes for multi-modal data fusion: front-end fusion or data level fusion, back-end fusion or decision level fusion, and intermediate fusion. The intermediate fusion is to convert the data of different modes into high-dimensional characteristic expression and then fuse the high-dimensional characteristic expression with the intermediate layer of the model. The invention adopts an intermediate fusion mode. The network structure diagram of the multi-mode data middle layer fusion full-connection geological map prediction model is shown in fig. 5. Wherein the meaning of each layer is as follows:
input layer: the basic data and the remote sensing data of the training sampling points are simultaneously input into a full-connection hybrid input model through an input layer;
basic data feature extraction layer: the feature extraction layer is composed of three full-connection layers, full-connection operation is carried out through a full-connection neural network, the network features, namely basic data features, are extracted by utilizing weight values, and original data are converted into high-dimensional feature expression (3 layers of the full-connection layers are determined by the following (2) th point description);
remote sensing data feature extraction layer: the feature extraction layer is composed of three full-connection layers, full-connection operation is carried out through a full-connection neural network, the network features, namely remote sensing data features, are extracted by utilizing weight values, and original data are converted into high-dimensional feature expression;
fusion layer: carrying out feature fusion on the basic data features and the remote sensing data features;
full tie layer: the method comprises the steps of obtaining common characteristics of basic data and remote sensing data in a high-dimensional space by using a fully connected neural network;
output layer: and outputting the probability values of the sampling points predicted as various types.
(1) Full-connection layer number optimization determination for converting different-mode data into high-dimensional feature expression
The selection of the number of layers of the fully connected layer is an important index for improving modeling accuracy. Increasing the number of layers can lead to the increase of operation time, lower efficiency and overfitting, and insufficient layers can lead to insufficient extraction of characteristic information so as to influence model precision and recall rate. For geological maps predicted by map units and lithology, one project is typically on the scale of 2-4 joint tests, with the map units and lithology classification trees being between substantially 100-200. The prediction is compared with the calculation of the number of different full-connection layers, and the advantages of the three full-connection layers are obvious in the conversion of different mode data into high-dimensional feature expression extraction and the improvement of model precision and recall rate. Table 2 is a comparison table of the effect of using different full connection layer numbers on model accuracy under the same sample data conditions.
TABLE 2
From table 2, it can be found that the difference between the accuracy and the average recall value of the initial model obtained by training the 3 model structures on the same training set sampling point is obvious at the original sampling point and the training set sampling point. Because the accuracy of the initial model can have a certain influence on the training sampling point of the secondary modeling, if the accuracy of the initial model is too low, the reliability of the training sampling point of the secondary sampling and the label thereof can be low. As can be seen from table 2, the feature extraction of the basic data and the remote sensing data by using the 2-layer fully connected neural network is insufficient, and the fitting degree of the training set is insufficient, so that the model has the phenomenon of under fitting. The 3-layer fully-connected neural network is adopted to extract the characteristics of the basic data and the remote sensing data, the accuracy and average recall rate of the original sampling points in the initial model are improved by more than 15%, the accuracy and average recall rate of the original sampling points in the secondary model are improved by more than 13% -15%, and the recall rate of the training set of the final model is improved by more than 8% although the accuracy of the training set of the final model is improved by more than 95%. The effect of the 4-layer fully-connected neural network on the feature extraction of the basic data and the remote sensing data is obviously improved in comparison with that of the 2-layer fully-connected neural network, but compared with that of the 3-layer fully-connected neural network, all corresponding indexes are about 4% different, which shows that the increase of the number of layers of the fully-connected layers in a single branch for independently extracting the features of the basic data and the remote sensing data does not obviously increase the corresponding accuracy and average recall value. Therefore, in the geological map prediction model based on PRB data deep learning, the adoption of a 3-layer fully connected neural network for extracting the characteristics of the basic data and the remote sensing data is a proper choice.
Principle and method for determining network structure parameters
During modeling, the network structure parameters are selected according to the principle and method of selecting the network structure parameters in table 2 according to the number of the labels of the predicted objects and the number of geological evidence. Table 3 is an example of a parameter determination method and principle given according to a specific application.
TABLE 3 Table 3
As shown in table 3, the meaning of each parameter is as follows:
a. input-1 and Input-2 represent basic data and remote sensing data Input, respectively. The corresponding output dimension represents the corresponding data category. If the basic data comprises 39 kinds of chemical element data and ground high magnetic data, the output dimension is 40 dimensions. The remote sensing data comprise high-resolution third satellite data, landsat8 satellite data and digital elevation model data, and the total number of the data is 16, so that the output dimension is 16 dimensions.
b. The three full connection layers of Dense-1, dense-2 and Dense-3 are used as basic data extraction features, and the three full connection layers of Dense-4, dense-5 and Dense-6 are used as remote sensing data extraction features.
c. The Concate layer is a fusion layer.
d. Dense-7 acquires the commonality characteristic of the basic data and the remote sensing data in a high-dimensional space.
e. Output is the Output layer. The number of neurons of the first full-connection layer of each single mode extraction feature, namely parameters, is determined by the number of geologic body categories to be classified, the category number is set as A, the number of neurons of the first full-connection layer is set as B, and the relation between A and B meets the following formula: b=2n > a.
f. The activation function is also one of key elements of network model training, and is mainly like adding a nonlinear factor into a neural network, so that the network can solve more complex nonlinear problems, meanwhile, the activation function filters out some useless information in the forward propagation process, and updates network training parameters in the backward propagation process. The ReLU function has the characteristics of single-side inhibition, single-side straight-through, wider activation boundary, sparse network and the like, so that the activation function has a relatively high convergence rate; the Softmax activation function is generally used as an activation function of an output layer, is often used in combination with a cross entropy loss function, is also called a normalized exponential function, is generalized on multi-classification tasks by a two-classification function Sigmoid function, and aims to display multi-classification results in the form of probability. Therefore, in the present invention, the ReLU activation function is used in the fully connected layer, and the activation function generally used as the output layer is used in the output layer, often in combination with the cross entropy loss function. The function is also called a normalized exponential function, is generalized on multi-classification tasks by a classification function Sigmoid function, and aims to display multi-classification results in a probability form.
The fully connected hybrid input model selects SGD (random gradient descent algorithm) as a parameter optimizer, and compared with the classical gradient descent algorithm which traverses all training data when updating parameters, the random gradient descent algorithm approximates the average loss of all training samples by using the loss of a single training sample, thereby greatly accelerating the network training speed. In the selection of the learning rate, 0.001 is selected as the initial learning rate, and a method for attenuating the learning rate is adopted in the network training, namely, a larger learning rate is adopted at first, after each parameter update, the learning rate is reduced to make finer adjustment on the parameter when the parameter is updated next time.
In a specific application, the change of the number of the filling units can be flexibly changed according to the processing parameter requirements.
The beneficial effects of the invention are as follows:
the method is simple to realize, not only converts the original data into high-dimensional characteristic expression, but also performs characteristic fusion on the basic data characteristics and the remote sensing data characteristics, and utilizes the fully connected neural network to maximally acquire the commonality characteristics of the basic data and the remote sensing data in the high-dimensional space, thereby greatly improving the capability and the precision of the prediction object, and enabling the most basic geological object of the geological map, namely the filling unit lithology, to be maximally reflected and expressed in terms of the indexes such as the spatial distribution form, the distribution direction, the distribution position, the adjacent relation among the geological objects and the like. After the technology breaks through, the existing geological investigation working mode is thoroughly changed, and the novel geological mapping mode integrating geological knowledge map, geological big data and deep learning algorithm is innovated, so that the optimal geological route and geological map prediction cycle progressive advancing mapping mode is realized, and a fine-granularity prediction geological map is formed.
While the foregoing description illustrates and describes the preferred embodiments of the present application, it is to be understood that this application is not limited to the forms disclosed herein, but is not to be construed as an exclusive use of other embodiments, and is capable of many other combinations, modifications and environments, and adaptations within the scope of the teachings described herein, through the foregoing teachings or through the knowledge or skills of the relevant art. And that modifications and variations which do not depart from the spirit and scope of the present invention are intended to be within the scope of the appended claims.