CN117272134A - Deep learning model, submarine topography classification model construction method and classification method - Google Patents
Deep learning model, submarine topography classification model construction method and classification method Download PDFInfo
- Publication number
- CN117272134A CN117272134A CN202311132082.0A CN202311132082A CN117272134A CN 117272134 A CN117272134 A CN 117272134A CN 202311132082 A CN202311132082 A CN 202311132082A CN 117272134 A CN117272134 A CN 117272134A
- Authority
- CN
- China
- Prior art keywords
- layer
- submarine
- classification
- network
- topography
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000012876 topography Methods 0.000 title claims abstract description 80
- 238000013136 deep learning model Methods 0.000 title claims abstract description 46
- 238000000034 method Methods 0.000 title claims abstract description 34
- 238000013145 classification model Methods 0.000 title claims abstract description 27
- 238000010276 construction Methods 0.000 title claims abstract description 12
- 238000000605 extraction Methods 0.000 claims description 37
- 238000012549 training Methods 0.000 claims description 33
- 230000006870 function Effects 0.000 claims description 30
- 238000011176 pooling Methods 0.000 claims description 27
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 18
- 238000009499 grossing Methods 0.000 claims description 17
- 238000005259 measurement Methods 0.000 claims description 16
- 238000013138 pruning Methods 0.000 claims description 16
- 238000007781 pre-processing Methods 0.000 claims description 10
- 238000009826 distribution Methods 0.000 claims description 7
- 230000006978 adaptation Effects 0.000 claims description 6
- 230000007246 mechanism Effects 0.000 claims description 6
- 230000009466 transformation Effects 0.000 claims description 6
- 238000012935 Averaging Methods 0.000 claims description 4
- 238000004590 computer program Methods 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 description 9
- 241000204169 Patiria miniata Species 0.000 description 6
- 238000012937 correction Methods 0.000 description 6
- 238000005065 mining Methods 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 244000141353 Prunus domestica Species 0.000 description 4
- 210000002569 neuron Anatomy 0.000 description 4
- 238000012795 verification Methods 0.000 description 4
- 230000003044 adaptive effect Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000005855 radiation Effects 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000012512 characterization method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- -1 depression Substances 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 229910052500 inorganic mineral Inorganic materials 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- NMJORVOYSJLJGU-UHFFFAOYSA-N methane clathrate Chemical compound C.C.C.C.O.O.O.O.O.O.O.O.O.O.O.O.O.O.O.O.O.O.O.O.O.O.O NMJORVOYSJLJGU-UHFFFAOYSA-N 0.000 description 1
- 239000011707 mineral Substances 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000011435 rock Substances 0.000 description 1
- 239000000523 sample Substances 0.000 description 1
- 239000013049 sediment Substances 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
- 239000013598 vector Substances 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention provides a deep learning model, a submarine topography classification model construction method and a classification method, and relates to the field of remote sensing information classification, wherein the deep learning model comprises an input layer, a priori knowledge weight network, a dense connection network group and an output layer, and the dense connection network group comprises a plurality of dense connection networks; the input layer is used for acquiring back scattering data; the priori knowledge weight network is used for extracting submarine classification features of the backscatter data; the dense connection network group is used for sequentially and deeply iterating the submarine classification features through a plurality of dense connection networks to generate deep submarine classification features; the output layer is used for outputting the submarine topography label according to the deep submarine classification characteristic. The features are extracted again on the basis of extracting the submarine classification features with high expression capacity through the dense connection network and the pre-knowledge weight network, so that the deep submarine classification features are more accurate, and the most accurate classification result is obtained.
Description
Technical Field
The invention relates to the field of remote sensing information classification, in particular to a deep learning model, a submarine topography classification model construction method and a classification method.
Background
In recent years, as the great potential of submarine resources in economy is explored, research into the topography of the seabed has become increasingly important. For the classification of submarine topography, the method has extremely important roles and significance in the fields of ocean resource development, submarine engineering, port construction, marine danger prevention and life saving, marine environment monitoring, land frame division and exclusive economic zone division, global climate evolution and sediment process exploration, submarine mineral and natural gas hydrate exploration, navigation and shipping and the like.
However, the current data set of the submarine topography is still quite scarce, and the existing submarine topography classification technology mainly adopts a man-machine interaction visual interpretation method, so that a certain subjectivity exists on the classification result of the submarine topography, the submarine topography classification precision is low, and the real submarine topography classification cannot be reflected.
Disclosure of Invention
The invention solves the problem of how to improve the precision of the classification of the submarine topography, and further reflects the real submarine topography classification.
In order to solve the problems, the invention provides a deep learning model, a submarine topography classification model construction method and a classification method.
In a first aspect, the present invention provides a deep learning model comprising: the device comprises an input layer, a priori knowledge weight network, a dense connection network group and an output layer, wherein the dense connection network group comprises a plurality of dense connection networks, the dense connection networks are connected in series, the input end of the first dense connection network is connected with the output end of the priori knowledge weight network, and the output end of the last dense connection network is connected with the input end of the output layer;
the input layer is used for acquiring back scattering data;
the priori knowledge weight network is used for extracting submarine classification features of the backscatter data;
the dense connection network group is used for sequentially and deeply iterating the submarine classification features through a plurality of dense connection networks to generate deep submarine classification features;
the output layer is used for outputting the submarine topography label according to the deep submarine classification characteristic.
Optionally, the prior knowledge weight network includes a weight adaptive layer and a feature extraction layer;
the weight self-adaptive layer is used for acquiring priori knowledge in training and updating the weight of the feature extraction layer according to the priori knowledge and the weight distribution mechanism;
the updated feature extraction layer is used for extracting the seabed classification features.
Optionally, the dense connection network includes a feature extraction sub-network, a downsampling layer, and a pruning sub-network; the characteristic extraction sub-network is respectively connected with the downsampling layer and the pruning sub-network, the characteristic extraction sub-network comprises a plurality of hidden layers which are sequentially arranged, each hidden layer is respectively connected with all the hidden layers before the hidden layer, and the last hidden layer is connected with the downsampling layer;
the feature extraction sub-network is used for extracting the submarine classified dense features of the submarine classified features;
the downsampling layer is used for reducing the resolution ratio of the submarine classified dense features and generating the deep submarine classified features;
and the pruning sub-network is used for deleting redundant connection and redundant channels of the hidden layer and generating a new feature extraction sub-network.
Optionally, the output layer includes a global average pooling layer and a full connection layer, an input end of the global average pooling layer is connected with an output end of the dense connection network, and an output end of the global average pooling layer is connected with an input end of the full connection layer;
the global average pooling layer is used for pooling the deep seabed classification features and generating low-dimensional deep seabed classification features;
the full connection layer is used for converting the low-dimensional deep seabed classification characteristics into the seabed topography tag.
Optionally, the device further comprises a loss module, wherein the loss module is respectively connected with the dense connection network and the output layer;
the loss module is used for obtaining the prediction probability and the label smoothing function of the submarine topography label, calculating a loss function according to the prediction probability and the label smoothing function, and updating the weight of the dense connection network according to the loss function.
Optionally, the input layer includes an acquisition layer and a preprocessing layer;
the acquisition layer is used for acquiring original back scattering data;
the preprocessing layer is used for removing stripe noise in the original back scattering data by wavelet transformation to generate the back scattering data.
Optionally, the a priori knowledge acquired by the weight adaptation layer includes water depth measurement data and seafloor characteristic data corresponding to the backscatter data.
In a second aspect, the invention provides a method for constructing a submarine topography classification model, which comprises the following steps:
acquiring a training set, wherein the training set comprises back scattering data and corresponding submarine topography tags;
training a pre-established deep learning model by adopting the training set to obtain a submarine topography classification model;
wherein, the deep learning model adopts the deep learning model.
In a third aspect, the present invention provides a method for classifying a submarine topography, comprising:
acquiring back scattering data of a target area;
and inputting the back scattering data into the submarine topography classification model obtained by the submarine topography classification model construction method, and generating a submarine topography label.
In a fourth aspect, the present invention also provides a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the method of subsea topography classification model construction as described above or the method of subsea topography classification as described above.
The deep learning model, the submarine topography classification model construction method, the submarine topography classification method and the computer readable storage medium have the advantages that:
after the back scattering data is acquired by the input layer, extracting the submarine classification features of the back scattering data by a priori knowledge weight network, and then carrying out deep iterative mining on the submarine classification features by a dense connection network group formed by a plurality of dense connection networks to obtain deep submarine classification features with multi-scale expression and high identification. The features are extracted again on the basis of the seabed classification features extracted by the dense connection network and the prior knowledge weight network, so that the deep seabed classification features are more accurate, and the most accurate classification result is obtained. In addition, because the output characteristics of each hidden layer of the dense connection network are transmitted to the next hidden layer as the input characteristics, the number of the input characteristics of each hidden layer can be increased, the training degree is increased, the problem of low classification precision caused by scarcity of a data set is solved, a better fitting effect is achieved, and the generalization capability of a deep learning model is increased.
Drawings
FIG. 1 is a schematic diagram of a deep learning model according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a deep learning model according to another embodiment of the present invention;
FIG. 3 is a schematic flow chart of a method for constructing a classification model of submarine topography according to an embodiment of the invention;
FIG. 4 is a schematic flow chart of a method for classifying seafloor topography according to an embodiment of the present invention;
fig. 5 is a schematic structural diagram of a feature extraction sub-network according to an embodiment of the invention.
Detailed Description
In order that the above objects, features and advantages of the invention will be readily understood, a more particular description of the invention will be rendered by reference to specific embodiments thereof which are illustrated in the appended drawings. While the invention is susceptible of embodiment in the drawings, it is to be understood that the invention may be embodied in various forms and should not be construed as limited to the embodiments set forth herein, but rather are provided to provide a more thorough and complete understanding of the invention. It should be understood that the drawings and embodiments of the invention are for illustration purposes only and are not intended to limit the scope of the present invention.
It should be understood that the various steps recited in the method embodiments of the present invention may be performed in a different order and/or performed in parallel. Furthermore, method embodiments may include additional steps and/or omit performing the illustrated steps. The scope of the invention is not limited in this respect.
The term "including" and variations thereof as used herein are intended to be open-ended, i.e., including, but not limited to. The term "based on" is based at least in part on. The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "alternative embodiments". Related definitions of other terms will be given in the description below. It should be noted that the terms "first," "second," and the like herein are merely used for distinguishing between different devices, modules, or units and not for limiting the order or interdependence of the functions performed by such devices, modules, or units.
It should be noted that references to "one", "a plurality" and "a plurality" in this disclosure are intended to be illustrative rather than limiting, and those skilled in the art will appreciate that "one or more" is intended to be construed as "one or more" unless the context clearly indicates otherwise.
The names of messages or information interacted between the devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of such messages or information.
As shown in fig. 1, an embodiment of the present invention provides a deep learning model, including: the device comprises an input layer, a priori knowledge weight network, a dense connection network group and an output layer, wherein the dense connection network group comprises a plurality of dense connection networks, the dense connection networks are connected in series, the input end of the first dense connection network is connected with the output end of the priori knowledge weight network, and the output end of the last dense connection network is connected with the input end of the output layer;
the input layer is used for acquiring back scattering data;
specifically, the back-scattered data is data obtained from information such as submarine topography, geological structure, and submarine medium parameters from measurement data of submarine scattered waves. The backscatter data can be acquired through the combined use of underwater multibeam sounding devices such as Reson Seabat 7125, reson Seabat 8101, and Reson Seabat 8111, or the Internet.
The priori knowledge weight network is used for extracting submarine classification features of the backscatter data;
specifically, the prior knowledge weight network comprises a convolution layer and a maximum pooling layer, and submarine classification features of the back scattering data are initially extracted through the convolution layer and the maximum pooling layer.
The dense connection network group is used for sequentially and deeply iterating the submarine classification features through a plurality of dense connection networks to generate deep submarine classification features;
specifically, the dense connection network group comprises a plurality of dense connection networks, wherein the input end of a first dense connection network is connected with the output end of a priori knowledge weight network, the output end of a last dense connection network is connected with the input end of an output layer, the first dense connection network receives a seabed classification model of the priori knowledge weight network and performs deep mining, the result of the deep mining is transmitted to a next dense connection network to perform deep mining, and the dense network transmits the result of the deep mining to the next dense connection network until the last dense connection network finally outputs deep seabed classification characteristics. Each hidden layer of each dense connection network is directly connected with all the hidden layers in the front to form a dense connection structure, namely, each layer can directly access the feature images of all the layers in the front, and because each layer is directly connected with all the layers in the front, gradient information and parameters in the dense connection network can be shared by a plurality of layers, the number of parameters calculated by the dense connection network is reduced, the risk of overfitting is reduced, and the problem of gradient disappearance is avoided.
For example, as shown in fig. 2, in order to achieve the highest classification precision and calculation efficiency, three dense connection networks may be set to iteratively extract deep sea-bottom classification features, so as to achieve the highest sea-bottom topography classification precision while reducing the calculation amount.
The output layer is used for outputting the submarine topography label according to the deep submarine classification characteristic.
Specifically, the output layer decodes the deep seabed classification feature, outputs the seabed topography label, and the seabed topography includes land frame, rock, depression, water course and ocean ridge etc., and the corresponding seabed topography label is 1, 2, 3, 4, 5, and specific seabed topography and corresponding seabed topography label can be set according to actual conditions.
The input layer acquires back scattering data, the priori knowledge weight network extracts more expressive and effective seabed classification features, the seabed classification features are deeply and iteratively excavated through a plurality of dense connection networks, the obtained deep seabed classification features have the characteristics of multi-scale expression and high identification degree, the classification is easy, and finally the seabed topography label output by the output layer is more accurate. Because the output characteristics of each hidden layer of the dense connection network are transmitted to the next hidden layer as input characteristics, and the characteristics are continuously extracted by the next hidden layer, the output characteristics of the next hidden layer contain the characteristic information of all the previous layers, the functions of parameter sharing, characteristic multiplexing and gradient sharing are realized, the information mobility is effectively improved, the problem that the classification precision of the submarine topography is not high due to the fact that the data set of the submarine topography is still quite scarce at present is solved, meanwhile, the calculation process is less, the robustness and generalization capability are improved, and the precision of the extracted characteristics is increased.
Optionally, as shown in fig. 2, the prior knowledge weight network includes a weight adaptation layer and a feature extraction layer;
the weight self-adaptive layer is used for acquiring priori knowledge in training and updating the weight of the feature extraction layer according to the priori knowledge and the weight distribution mechanism;
the updated feature extraction layer is used for extracting the seabed classification features.
Specifically, the prior knowledge is knowledge prior to experience, that is, correct knowledge verified in advance, through which the extracted target feature can be better inferred. Updating the weight of the feature extraction layer according to the prior knowledge and the weight distribution mechanism means that before model training is carried out, the weight distribution mechanism is guided to adjust the weight of the model for initialization or adjustment according to the knowledge or the previous experience of a domain expert; the prior knowledge acquired by the weight adaptive layer during training comprises water depth measurement data and seabed characteristic data. The water depth measurement data refers to data for measuring the vertical distance from a certain point in the ocean or water body to the water surface or the seabed, namely seabed topography data; the submarine feature data refers to data describing features such as submarine topography, geological structures, submarine media and the like; further, it is also possible to acquire geomorphic category label data corresponding to the water depth measurement data, the sea bottom feature data, and the backscatter data, that is, the classification result of the determined backscatter data current area has been confirmed. The method comprises the steps of segmenting three data into data sets with the size of 64 x 64 by adopting a python batch segmentation data method through water depth measurement data, seabed characteristic data, back scattering data and landform type label data, and dividing the data sets into a training set, a verification set and a test set according to the ratio of 3:1:1, wherein the training set, the verification set and the test set are used for training a deep learning model, so that the deep learning model can better complete a seabed landform classification task.
The feature extraction layer and the weight self-adaptive layer comprise a convolution layer and a maximum pooling layer structure, and are respectively used for extracting features and adjusting weights.
The weight self-adaptive layer takes the acquired water depth measurement data, the submarine feature data and the landform type tag data as priori knowledge, judges the influence of the water depth measurement data, the submarine feature data and the backscattering data on the landform type tag data, and self-adaptively adjusts the input weight and the proportion of the water depth measurement data, the submarine feature data and the backscattering data in the feature extraction layer and the weights of the convolution layer and the maximum pooling layer of the feature extraction layer according to the influence and the weight distribution mechanism until the submarine classification features extracted by the feature extraction layer reach the highest classification precision.
Illustratively, the multi-beam data of the Moruo Bay Male corner offshore of California, including the backscatter data in TIFF format, the water depth measurement data, the subsea characterization data, and the tag data in shp format, may be downloaded from the U.S. geological office USGS network.
Optionally, as shown in fig. 2, the dense connection network includes a feature extraction sub-network, a downsampling layer, and a pruning sub-network; the characteristic extraction sub-network is respectively connected with the downsampling layer and the pruning sub-network, the characteristic extraction sub-network comprises a plurality of hidden layers which are sequentially arranged, each hidden layer is respectively connected with all the hidden layers before the hidden layer, and the last hidden layer is connected with the downsampling layer;
the feature extraction sub-network is used for extracting the submarine classified dense features of the submarine classified features;
the downsampling layer is used for reducing the resolution ratio of the submarine classified dense features and generating the deep submarine classified features;
and the pruning sub-network is used for deleting redundant connection and redundant channels of the hidden layer and generating a new feature extraction sub-network.
Specifically, the feature extraction sub-network includes a plurality of hidden layers, the hidden layers are called dense blocks, and the hidden layers are all connected with all the hidden layers, so when there are L hidden layers, there are l×1/2 connections, that is, dense connections, through dense connections, parameters required by the feature extraction sub-network are less, calculation steps are reduced, the number of feature graphs generated by the hidden layers is defined as a growth rate, the number of feature graphs generated by the hidden layers is controlled according to a growth rate formula, and the growth rate formula is as follows:
K=2 m-1 *k 0 ;
wherein K is growth rate, m is m-th hidden layer, K 0 Is constant.
The hidden layer comprises a plurality of convolution layers, and the calculation formula of the convolution layers is as follows:
x L =T L ([x 0 ,x 1 ,…,x L-1 ]);
wherein [ x ] 0 ,x 1 ,...,x L-1 ]Output of cascade operation for convolution layer of front L layers, T L To include a set of nonlinear transforms, x L Is the output of the L-th convolutional layer. In this embodiment, the hidden layer includes multiple sets of 1*1 and 3*3 convolutional layers with the same padding for cascading operations to extract a richer feature representation.
In one embodiment, as shown in fig. 5, the feature extraction sub-network includes Identity (Identity mapping block), 2 x 2Pooling (2 x 2Pooling block), 4 x 4Pooling (4*4 Pooling block) and Global Pooling block, where Identity, 2 x 2Pooling and 4 x 4Pooling are combined and/or separately set, and each of the inputs of Identity, 2 x 2Pooling, 4 x 4Pooling and Global Pooling is derived from the outputs of all of Identity, 2 x 2Pooling and 4 x 4Pooling, so as to implement feature multiplexing, and as the structure goes deep, the feature map size generated by each group is smaller and smaller, the number of feature maps is larger and the extracted features are further extracted by multiple Global Pooling and output.
The downsampling layer is used for unifying feature graphs with different sizes or resolutions and generated by the feature extraction sub-network, and generating deep seabed classification features. The pruning sub-network utilizes the sparse network principle, carries out zero setting operation on the convolution kernel according to the L1 value of the convolution kernel during training, namely learningly grouping convolution, sparsifies the weight of the feature extraction sub-network, then prunes connections with smaller weight, realizes pruning operation, and removes redundant connection and redundant channels. The pruning sub-network prunes the weights of the feature extraction sub-network at the beginning of training, so that the pruning sub-network can accurately prune and can continue training the model, and the network weights are smoother. The learning-based grouped convolution is applied to the first 1*1 convolution of each hidden layer, the channel scrambling operation is performed, the feature images are uniformly distributed in each group, and the 3*3 convolution-based standard grouped convolution is further arranged, so that the model can be lightened and efficient by reducing parameters of a deep learning model and improving feature interactivity, and meanwhile, the expression capability of the model is enhanced, so that the model is lighter and more efficient, and the method is suitable for the conditions that a data set of submarine topography is quite scarce and the accuracy requirement is efficient. Illustratively, the pruning sub-network may also employ importance-ranking pruning and sensitivity-ranking pruning methods to prune the weights of the feature extraction sub-network.
Optionally, as shown in fig. 2, the output layer includes a global average pooling layer and a full connection layer, an input end of the global average pooling layer is connected with an output end of the dense connection network, and an output end of the global average pooling layer is connected with an input end of the full connection layer;
the global average pooling layer is used for pooling the deep seabed classification features and generating low-dimensional deep seabed classification features;
the full connection layer is used for converting the low-dimensional deep seabed classification characteristics into the seabed topography tag.
Specifically, the global average pooling layer performs average pooling dimension reduction on the features on each channel of the deep seabed classification features to finally obtain the low-dimension deep seabed classification features on one channel. The low-dimensional deep sea-bottom classification features output by the global averaging pooling layer can be regarded as global summary information of the entire deep sea-bottom classification features. Secondly, the output of the global average pooling layer is connected to the full connection layer; each neuron of the full-connection layer is connected with all neurons of the downsampling layer in the dense connection network, each neuron of the full-connection layer corresponds to one label type, the prediction probability or score of the corresponding type can be obtained by calculating the activation value of the neuron of the full-connection layer, the activation value of the output layer is normalized by using a Softmax function to obtain probability distribution of each type, and then the type with the highest probability is selected as the final prediction label, namely the submarine topography label is generated.
Optionally, the device further comprises a loss module, wherein the loss module is respectively connected with the dense connection network and the output layer;
the loss module is used for obtaining the prediction probability and the label smoothing function of the submarine topography label, calculating a loss function according to the prediction probability and the label smoothing function, and updating the weight of the dense connection network according to the loss function.
Specifically, the loss module calculates the prediction probability according to a geomorphic probability formula, which is shown as follows:
wherein p is i To predict probability, x T Vectors representing deep sea-bottom classification features, w i Weight of the ith class, w l Is the weight of the first category.
Calculating a label smoothing function according to a landform smoothing formula, and bringing the prediction probability and the label smoothing function into a loss function calculation formula to obtain a loss function, wherein the landform smoothing formula is as follows:
y′=(1-∈)y+∈u(I);
where y' is a label smoothing function, e is a smoothing factor, and u (I) is a uniform distribution of obeying class numbers. The loss function calculation formula is as follows:
wherein H (y, p) is a loss function, y i True value when y i When 1, it indicates that the classification is correct, y i When 0, an error is indicated; p is p i To predict probability.
When the prediction probability and the label smoothing function are brought into the loss function calculation formula, the label smoothing function y' in the relief smoothing formula needs to be replaced by the true value y in the loss function calculation formula i . Because the loss function without the tag smoothing function only calculates the value of the correct position of the tag, this can result in the network being overly focused on increasing the probability of correct tags, rather than decreasing the probability of predicting false tags, resulting in poor results when based on other test set experiments, although the model fits the own training set well. In calculating the penalty function, the smoothed label smoothing function will be used instead of the true value of the true label. Therefore, the model can pay more attention to samples with larger uncertainty in the training process, and the robustness and generalization capability of the model are improved.
Optionally, the input layer comprises an acquisition layer and a preprocessing layer;
the acquisition layer is used for acquiring original back scattering data;
the preprocessing layer is used for removing stripe noise in the original back scattering data by wavelet transformation to generate the back scattering data.
In particular, the acquisition layer may download multi-beam data off-shore from the USGS network of the united states geological office, the morobus wave male angle of california, which includes raw back-scattered data in TIFF format. The original backscatter data is then subjected to preprocessing operations such as decoding, geographical correction, radiation correction, wavelet transformation, etc., to generate backscatter data. In order to train the deep learning model, the method of splitting data in python batches is adopted to split the back scattering data into data sets with the size of 64 x 64, and the data sets are divided into a training set, a verification set and a test set according to the ratio of 3:1:1, so that the deep learning model is trained, and the deep learning model can better complete the classification task of the submarine topography.
Optionally, the a priori knowledge acquired by the weight adaptation layer includes water depth measurement data and seafloor characteristic data corresponding to the backscatter data.
Specifically, the prior knowledge includes water depth measurement data and seafloor characteristic data corresponding to the back-scattered data, the weight adaptation layer may obtain the prior knowledge through the input layer, and the input layer may download multi-beam data of Moro Bay wave Male angle offshore of California from the United states geological bureau USGS functional network, which includes raw back-scattered data in TIFF format, raw water depth measurement data, raw seafloor characteristic data, and raw tag data in shp format. And then converting the shp format original tag data into TIFF format original tag data by using an Arcgis surface grid conversion tool through a preprocessing layer, and then carrying out preprocessing operations such as decoding, geographic correction, radiation correction, wavelet transformation and the like on the original backscatter data, the original water depth measurement data, the original seabed characteristic data and the original tag data to generate water depth measurement data, seabed characteristic data and backscatter data. In order to train the deep learning model, three data are segmented into data sets with the size of 64 x 64 by adopting a python batch data segmentation method, and the data sets are divided into a training set, a verification set and a test set according to the proportion of 3:1:1, so that the deep learning model is trained, and the deep learning model can better complete the classification task of the submarine topography.
In one embodiment, to further illustrate the advantages of the present invention, the present invention performs precision evaluation and outcome evaluation on the deep learning model and the existing model of the present invention with the constructed data set. The interpretation results of the deep learning model and other existing models are shown in the following table respectively:
Net/Index | Precision(%) | Recall(%) | F1-score(%) | Iou(%) |
Condensenet+Lss1 | 71.74±0.69 | 50.62±0.72 | 53.34±0.63 | 35.28±0.56 |
Condensenet-C1-wce | 71.92±0.66 | 50.58±0.85 | 53.28±0.82 | 35.80±0.77 |
Condensenet-C3 | 70.96±0.55 | 45.51±0.64 | 53.22±0.65 | 35.82±0.69 |
Condensenet-C1-lss1weight | 72.84±0.78 | 46.63±0.77 | 53.08±0.67 | 36.26±0.57 |
Condensenet-C3-lss1weight | 73.13±0.67 | 48.40±0.66 | 54.74±0.78 | 36.38±0.68 |
to make the comparison more convincing, we iteratively trained each model for 20 rounds over the training set, averaged and calculated the positive and negative bias. The Condensnet-C3-ls 1weight in the table is a deep learning model of the invention, and the overall classification accuracy of the deep learning model of the invention is higher than that of other existing models as can be seen from the table. The deep learning model of the present invention was 73.13% in Precision, 36.38% in IOU and 54.74% in F1-score. Precison increases by approximately one percent compared to the Condensnet-C1-wce model without prior knowledge, and by approximately two percent compared to the weight-free adaptive Condensnet-C3 model; for F1-score, the deep learning model of the present invention improves by approximately 1.5 percent; for IOU, the deep learning model of the present invention improves by about 0.12 percent. The result shows that the deep learning model can improve the precision of classification of the submarine topography.
In one embodiment, as shown in fig. 3, the embodiment of the invention provides a method for constructing a classification model of submarine topography, which comprises the following steps:
step S10, a training set is obtained, wherein the training set comprises back scattering data and corresponding submarine topography tags;
step S20, training a pre-established deep learning model by adopting the training set to obtain a submarine topography classification model;
wherein, the deep learning model adopts the deep learning model.
Specifically, the method can directly acquire the back scattering data used for training the deep learning model and the training set of the corresponding submarine topography label or topography class label data, and can adjust the weight of the model according to actual conditions after training is completed, so as to obtain the submarine topography classification model with highest precision.
In one embodiment, as shown in fig. 4, an embodiment of the present invention provides a method for classifying a submarine topography, including:
step S10, obtaining back scattering data of a target area;
specifically, the back scattering data of the target area needs to be acquired first, so that the back scattering data is input into a submarine topography classification model to complete classification tasks, and preferably, the back scattering data can be acquired through the combined use of underwater multi-beam depth sounders such as a Reson Seabat 7125, a Reson Seabat 8101 and a Reson Seabat 8111.
And S20, inputting the back scattering data into the submarine topography classification model obtained by the submarine topography classification model construction method, and generating a submarine topography label.
Specifically, the acquired back scattering data is input into a trained submarine topography classification model, so that a high-precision submarine topography label can be obtained, because the submarine topography classification model extracts submarine classification features based on priori knowledge, the submarine classification features are deeply excavated by adopting a dense connection network, and the precision of the finally obtained submarine topography label is extremely high through repeated iterative training and parameter adjustment.
Optionally, the acquiring the backscatter data of the target area includes:
acquiring submarine data of the target area;
and removing stripe noise of the submarine data, and acquiring the back scattering data.
Specifically, the underwater multi-beam probe can be used for acquiring the submarine data, and preprocessing such as geographic correction, radiation correction, data cleaning, wavelet transformation and the like can be performed on the submarine data, so that stripe noise of the submarine data is removed, and clean and tidy back scattering data is obtained.
In one embodiment, the present invention provides a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements a subsea topography classification model building method as described above or a subsea topography classification method as described above.
Although the invention is disclosed above, the scope of the invention is not limited thereto. Various changes and modifications may be made by one skilled in the art without departing from the spirit and scope of the invention, and these changes and modifications will fall within the scope of the invention.
Claims (10)
1. A deep learning model, comprising: the device comprises an input layer, a priori knowledge weight network, a dense connection network group and an output layer, wherein the dense connection network group comprises a plurality of dense connection networks, the dense connection networks are connected in series, the input end of the first dense connection network is connected with the output end of the priori knowledge weight network, and the output end of the last dense connection network is connected with the input end of the output layer;
the input layer is used for acquiring back scattering data;
the priori knowledge weight network is used for extracting submarine classification features of the backscatter data;
the dense connection network group is used for sequentially and deeply iterating the submarine classification features through a plurality of dense connection networks to generate deep submarine classification features;
the output layer is used for outputting the submarine topography label according to the deep submarine classification characteristic.
2. The deep learning model of claim 1, wherein the prior knowledge weight network comprises a weight adaptation layer and a feature extraction layer;
the weight self-adaptive layer is used for acquiring priori knowledge in training and updating the weight of the feature extraction layer according to the priori knowledge and the weight distribution mechanism;
the updated feature extraction layer is used for extracting the seabed classification features.
3. The deep learning model of claim 1, wherein the dense connection network comprises a feature extraction sub-network, a downsampling layer, and a pruning sub-network; the characteristic extraction sub-network is respectively connected with the downsampling layer and the pruning sub-network, the characteristic extraction sub-network comprises a plurality of hidden layers which are sequentially arranged, each hidden layer is respectively connected with all the hidden layers before the hidden layer, and the last hidden layer is connected with the downsampling layer;
the feature extraction sub-network is used for extracting the submarine classified dense features of the submarine classified features;
the downsampling layer is used for reducing the resolution ratio of the submarine classified dense features and generating the deep submarine classified features;
and the pruning sub-network is used for deleting redundant connection and redundant channels of the hidden layer and generating a new feature extraction sub-network.
4. The deep learning model of claim 1, wherein the output layer comprises a global averaging pooling layer and a fully connected layer, an input of the global averaging pooling layer being connected to an output of the dense connection network, an output of the global averaging pooling layer being connected to an input of the fully connected layer;
the global average pooling layer is used for pooling the deep seabed classification features and generating low-dimensional deep seabed classification features;
the full connection layer is used for converting the low-dimensional deep seabed classification characteristics into the seabed topography tag.
5. The deep learning model of any of claims 1-4, further comprising a penalty module connected to the dense connection network and the output layer, respectively;
the loss module is used for obtaining the prediction probability and the label smoothing function of the submarine topography label, calculating a loss function according to the prediction probability and the label smoothing function, and updating the weight of the dense connection network according to the loss function.
6. The deep learning model of claim 1, wherein the input layers include an acquisition layer and a preprocessing layer;
the acquisition layer is used for acquiring original back scattering data;
the preprocessing layer is used for removing stripe noise in the original back scattering data by wavelet transformation to generate the back scattering data.
7. The deep learning model of claim 2, wherein the a priori knowledge acquired by the weight adaptation layer includes water depth measurement data and seafloor characteristic data corresponding to the backscatter data.
8. The method for constructing the submarine topography classification model is characterized by comprising the following steps of:
acquiring a training set, wherein the training set comprises back scattering data and corresponding submarine topography tags;
training a pre-established deep learning model by adopting the training set to obtain a submarine topography classification model;
wherein the deep learning model employs the deep learning model as claimed in any one of claims 1 to 7.
9. A method for classifying seafloor topography, comprising:
acquiring back scattering data of a target area;
inputting the back scattering data into the submarine topography classification model obtained by the submarine topography classification model construction method according to claim 8, and generating a submarine topography label.
10. A computer-readable storage medium, on which a computer program is stored which, when executed by a processor, implements the method of subsea topography classification model construction according to claim 8 or the method of subsea topography classification according to claim 9.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311132082.0A CN117272134A (en) | 2023-09-01 | 2023-09-01 | Deep learning model, submarine topography classification model construction method and classification method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311132082.0A CN117272134A (en) | 2023-09-01 | 2023-09-01 | Deep learning model, submarine topography classification model construction method and classification method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN117272134A true CN117272134A (en) | 2023-12-22 |
Family
ID=89209587
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202311132082.0A Pending CN117272134A (en) | 2023-09-01 | 2023-09-01 | Deep learning model, submarine topography classification model construction method and classification method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117272134A (en) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111738332A (en) * | 2020-06-22 | 2020-10-02 | 黄河勘测规划设计研究院有限公司 | Underwater multi-source acoustic image substrate classification method and system based on feature level fusion |
CN114266920A (en) * | 2021-12-14 | 2022-04-01 | 北京邮电大学 | Deep learning image classification method and system based on knowledge driving |
CN115936992A (en) * | 2022-12-28 | 2023-04-07 | 西北工业大学太仓长三角研究院 | Garbage image super-resolution method and system of lightweight transform |
-
2023
- 2023-09-01 CN CN202311132082.0A patent/CN117272134A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111738332A (en) * | 2020-06-22 | 2020-10-02 | 黄河勘测规划设计研究院有限公司 | Underwater multi-source acoustic image substrate classification method and system based on feature level fusion |
CN114266920A (en) * | 2021-12-14 | 2022-04-01 | 北京邮电大学 | Deep learning image classification method and system based on knowledge driving |
CN115936992A (en) * | 2022-12-28 | 2023-04-07 | 西北工业大学太仓长三角研究院 | Garbage image super-resolution method and system of lightweight transform |
Non-Patent Citations (6)
Title |
---|
CHRISTIAN SZEGEDY,ET AL.: "Rethinking the Inception Architecture for Computer Vision", 《ARXIV》, 11 December 2015 (2015-12-11), pages 7 * |
GAO HUANG,ET AL.: "CondenseNet: An Efficient DenseNet using Learned Group Convolutions", 《2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》, 16 December 2018 (2018-12-16), pages 2 - 4 * |
LAURA VON RUEDEN,ET AL.: "Informed Machine Learning - A Taxonomy and Survey of Integrating Prior Knowledge into Learning Systems", 《IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING》, 7 December 2022 (2022-12-07), pages 2 - 5 * |
唐秋华;刘保华;陈永奇;周兴华;: "基于改进BP神经网络的海底底质分类", 海洋测绘, no. 05, 30 September 2009 (2009-09-30) * |
罗伟东;郭军;: "基于多波束背向散射数据的海底底质分类", 海洋地质前沿, no. 08, 31 August 2017 (2017-08-31) * |
阳凡林,翟国君,赵建虎编: "海洋测绘学概论", 31 March 2022, 武汉大学出版社, pages: 176 - 184 * |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111724478B (en) | Point cloud up-sampling method based on deep learning | |
CN109086824B (en) | Seabed substrate sonar image classification method based on convolutional neural network | |
CN109709603B (en) | Seismic horizon identification and tracking method and system | |
CN110361778B (en) | Seismic data reconstruction method based on generation countermeasure network | |
Ji et al. | Seabed sediment classification using multibeam backscatter data based on the selecting optimal random forest model | |
CN107016677A (en) | A kind of cloud atlas dividing method based on FCN and CNN | |
Jin et al. | A point-based fully convolutional neural network for airborne LiDAR ground point filtering in forested environments | |
US9341712B2 (en) | Variable resolution uncertainty expert system for digital bathymetry database | |
CN103020649A (en) | Forest type identification method based on texture information | |
KR102367818B1 (en) | Deep learning-based seabed sediment analysis system for benthic ecological habitat classification | |
CN117953314B (en) | Multi-dimensional feature optimization ocean substrate classification method and system | |
Yin et al. | Mapping high-resolution basal topography of West Antarctica from radar data using non-stationary multiple-point geostatistics (MPS-BedMappingV1) | |
CN114022368A (en) | Pavement disease data enhancement method based on generation of countermeasure network | |
CN116468995A (en) | Sonar image classification method combining SLIC super-pixel and graph annotation meaning network | |
Wang et al. | Automatic mapping of lunar landforms using DEM-derived geomorphometric parameters | |
Chakraborty et al. | Application of hybrid techniques (self-organizing map and fuzzy algorithm) using backscatter data for segmentation and fine-scale roughness characterization of seepage-related seafloor along the western continental margin of India | |
CN111428627B (en) | Mountain landform remote sensing extraction method and system | |
CN115639605B (en) | Automatic identification method and device for high-resolution fault based on deep learning | |
CN116663428A (en) | Neural network-based face wave imaging joint inversion method | |
CN117272134A (en) | Deep learning model, submarine topography classification model construction method and classification method | |
CN115223033A (en) | Synthetic aperture sonar image target classification method and system | |
WO2022093036A1 (en) | Method of estimating polymetallic nodule abundance | |
CN116934603A (en) | Logging curve missing segment completion method and device, storage medium and electronic equipment | |
CN117173548B (en) | Method and device for constructing intelligent classification model of submarine topography and classification method | |
Ganguli et al. | Predicting food security outcomes using cnns for satellite tasking |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |