CN-121561612-B - Method for constructing sea island sediment classification model, classification method and device

CN121561612BCN 121561612 BCN121561612 BCN 121561612BCN-121561612-B

Abstract

The invention provides a method for constructing a sea island reef substrate classification model, a classification method and a device, and relates to the technical field of model classification, wherein the method for constructing the sea island reef substrate classification model comprises the steps of obtaining a data set of an island reef region, wherein the data set comprises remote sensing data, water depth data and chlorophyll concentration which are inverted based on the remote sensing data; and respectively carrying out feature extraction on the data set based on a preset neural network model, constructing a positive sample pair according to the obtained spectral features and the water depth-chlorophyll features at the same preset position in the island reef area, constructing a whole sample pair according to the obtained spectral features at the selected preset position in the island reef area and the water depth-chlorophyll features at all preset positions, and constructing a sample feature pair according to the positive sample pair and the whole sample pair. The island reef substrate classification method and the island reef substrate classification device can improve the precision of island reef substrate classification under a small amount of label data.

Inventors

CHEN YIFU
WU LIN
YUE YUAN

Assignees

中国地质大学（武汉）

Dates

Publication Date: 20260508
Application Date: 20260126

Claims (8)

1. The method for constructing the island reef substrate classification model is characterized by comprising the following steps of: acquiring a data set of an island reef area, wherein the data set comprises remote sensing data, water depth data inverted based on the remote sensing data and chlorophyll concentration; Inputting the remote sensing data into a first network in a preset neural network model to obtain spectral characteristics; Inputting the water depth data and chlorophyll concentration into a second network in the preset neural network model to obtain water depth-chlorophyll characteristics; Constructing a first positive sample pair based on the spectral features and the water depth-chlorophyll features of the same substrate region in a preset position, constructing a first global sample pair based on the spectral features of a selected substrate region and the water depth-chlorophyll features of all the substrate regions, and constructing a first sample feature pair based on the first positive sample pair and the first global sample pair; Constructing a second positive sample pair based on the spectral features and the water depth-chlorophyll features of the same location point under the same substrate region in the preset location, constructing a second whole sample pair based on the spectral features of the selected location point under the same substrate region and the water depth-chlorophyll features of all location points under the same substrate region, and constructing a second sample feature pair according to the second positive sample pair and the second whole sample pair; Wherein the sample feature pairs comprise the first sample feature pair and a second sample feature pair; Constructing a global contrast loss function according to the first sample characteristic pairs, wherein the global contrast loss function is used for shortening the distance of the first positive sample pairs of the same substrate region in a characteristic space and expanding the distance of first negative sample pairs of different substrate regions in the characteristic space; Constructing a local contrast loss function according to the second sample characteristic pairs, wherein the local contrast loss function is used for pulling the distance of the second positive sample pairs of the same position points of the same substrate region in the characteristic space, expanding the distance of the second negative sample pairs of different position points in the same substrate region in the characteristic space, and constructing the second negative sample pairs for the spectral characteristics and the water depth-chlorophyll characteristics of different positions in the same substrate region in the second whole sample pairs; The global contrast loss function and the local contrast loss function are synthesized, and training is conducted on the first network and the second network until preset training completion conditions are met, so that a fusion feature extractor is obtained; and constructing a sea island reef substrate classification model according to the fusion characteristic extractor and the pre-trained classifier so as to classify the substrate region of the island reef region.
2. The method of constructing an island reef bed classification model according to claim 1, wherein the acquiring the data set of the island reef area comprises: Acquiring remote sensing data of the island reef area, wherein the remote sensing data comprises laser radar data and corresponding remote sensing images; And inverting the water depth data and the chlorophyll concentration according to the remote sensing data, wherein the water depth data is obtained based on the laser radar data and the corresponding remote sensing image, and the chlorophyll concentration is obtained based on the reflectivity inversion of different wave bands of light under the remote sensing image.
3. The method of constructing an island reef matrix classification model according to claim 2, wherein inverting the water depth data from the remote sensing data comprises: determining signal photon density distribution at different elevations of the water body according to the laser radar data under single photons; Determining the upper and lower boundaries of the initial water surface photons according to the difference of the density distribution of the signal photons, and dividing the original single photons into photons above the water surface, photons on the water surface and photons under water according to the upper and lower boundaries of the initial water surface photons; Performing filtering processing on the underwater photons, wherein a variable elliptic filtering core is used for traversing the underwater photons, reserving the underwater photons of which the number of photons exceeds a preset threshold value in the elliptic filtering core, and determining the reserved underwater photons as underwater photons; The elevation of the water surface photon is different from Gao Chengzuo of the water bottom photon, and local water depth data corresponding to the laser radar data are obtained; And determining the water depth data corresponding to the remote sensing data according to the local water depth data corresponding to the laser radar data under the single photon and the corresponding relation between the laser radar data and the remote sensing data.
4. The method for constructing an island reef matrix classification model according to claim 1, wherein the synthesizing the global contrast loss function and the local contrast loss function, training the first network and the second network until a preset training completion condition is satisfied comprises: the global contrast loss function and the local contrast loss function are weighted and summed to construct an overall loss function, and the first network and the second network are trained according to the overall loss function until a preset training completion condition is met; Wherein the overall loss function is: ; Representing an overall loss function; a represents the global loss function Is used for the weight of the (c), 2N represents the number of all samples, and, For a first positive sample pair constructed from the same substrate region spectral features and water depth-chlorophyll features, A first population of sample pairs constructed from the spectral features of the selected substrate region, and the water depth-chlorophyll features of all of the substrate regions; Representing local loss function Is used for the weight of the (c), , Representing the number of location points under all samples; For a second positive sample pair constructed from the spectral features and water depth-chlorophyll features of the same location point under the same substrate region, A second population of sample pairs constructed from the spectral features of selected location points and the water depth-chlorophyll features of all location points under the same substrate region; sim represents a cosine similarity operation, and k represents a cyclic variable of the sample feature pair.
5. The method of constructing an island reef matrix classification model according to claim 1, wherein the first network comprises an image convolution feature extractor, the second network comprises a numerical feature extractor, the predetermined neural network model further comprises a cross-modal fusion module, the spectral features are spatial-spectral joint features, the water depth-chlorophyll features are water depth-chlorophyll joint features, the image convolution feature extractor is used for extracting the spatial-spectral joint features, the numerical feature extractor is used for extracting the water depth-chlorophyll joint features, the cross-modal fusion module is used for fusing the spatial-spectral joint features and the water depth-chlorophyll joint features and introducing a cross-modal attention mechanism to generate an enhanced feature representation, the enhanced feature representation is used for inputting a pre-trained classifier for classification processing, and the pre-trained classifier comprises a classification network introducing attention gating.
6. The method for constructing a sea island reef matrix classification model according to claim 5, wherein the cross-modal fusion module is specifically configured to perform layer normalization processing on cross-modal attention features to obtain a first intermediate feature, input the first intermediate feature representation into a feedforward neural network to perform nonlinear transformation to obtain a second intermediate feature, and fuse the first intermediate feature with the second intermediate feature to generate an enhanced fusion feature representation.
7. An island reef matrix classification method constructed based on the island reef matrix classification model construction method according to any one of claims 1 to 6, comprising: and acquiring a real-time data set of the island reef area, and inputting the real-time data set into the island reef substrate classification model to output island reef substrate classification of the substrate area.
8. An island reef matrix sorting apparatus comprising a memory and a processor, wherein the memory is configured to store a computer program, and wherein the processor is configured to implement the method of constructing an island reef matrix sorting model according to any one of claims 1 to 6 or the method of sorting island reef matrix according to claim 7 when the computer program is executed.

Description

Method for constructing sea island sediment classification model, classification method and device Technical Field The invention relates to the technical field of model classification, in particular to a method for constructing a sea island reef substrate classification model, a classification method and a classification device. Background The island reef substrate type is an important component part of a marine ecological system, has important significance in maintaining marine organism diversity, supporting ecological system service functions and resource development activities, is various in island reef substrate type, mainly comprises various components such as coral scraps, shell sand, gravels, rocks and fine sand, and the spatial distribution change of the island reef substrate type directly influences the habitat structure of the marine organism, underwater sound wave propagation characteristics and coastal landform evolution process, so that classification and distribution investigation of the island reef substrate type are carried out, and the island reef substrate type has important value for ecological protection, resource utilization and environment monitoring. In the related art, the main means for acquiring the island reef substrate information are in-situ investigation and remote sensing technologies, such as manual interpretation based on expert knowledge, supervision classification based on traditional machine learning and the like, wherein in practical application, the manual interpretation based on expert knowledge is limited by expert experience, has high subjectivity and low efficiency although the classification precision is high, and the traditional machine learning method such as a support vector machine, a random forest and the like has high dependence on characteristic dimensions, is easy to generate "salt and pepper noise" in classification, so that a large number of pixels (pixels) are erroneously classified, and the underwater optical environment is complex, the interpretation faces technical bottlenecks, so that tag data are less, and the method is difficult to adapt to the complicated island reef substrate classification. Disclosure of Invention The invention aims to solve the problems that the classification precision of the related technology is insufficient and a large amount of label data is often relied on. In order to solve the above problems, in a first aspect, the present invention provides a method for constructing a sea island reef substrate classification model, including: acquiring a data set of an island reef area, wherein the data set comprises remote sensing data, water depth data inverted based on the remote sensing data and chlorophyll concentration; respectively carrying out feature extraction on the data set based on a preset neural network model, constructing a positive sample pair according to the obtained spectral features and the water depth-chlorophyll features at the same preset position in the island reef area, constructing a whole sample pair according to the obtained spectral features at the selected preset position in the island reef area and the water depth-chlorophyll features at all preset positions, and constructing a sample feature pair according to the positive sample pair and the whole sample pair; Constructing a loss function according to the sample feature pairs, and performing global contrast learning training and local contrast learning training corresponding to the preset positions on the preset neural network model to obtain a fusion feature extractor; and constructing a sea island reef substrate classification model according to the fusion characteristic extractor and the pre-trained classifier so as to classify the substrate region of the island reef region. Optionally, the acquiring the data set of the island reef area includes: acquiring the remote sensing data, wherein the remote sensing data comprises laser radar data and corresponding remote sensing images; And inverting the water depth data and the chlorophyll concentration according to the remote sensing data, wherein the water depth data is obtained based on inversion of the laser radar data and the corresponding remote sensing image, and the chlorophyll concentration is obtained based on inversion of reflectivities of different wave bands of light under the remote sensing image. Optionally, said inverting said water depth data from said remote sensing data comprises: determining signal photon density distribution at different elevations of the water body according to the laser radar data under single photons; Determining the upper and lower boundaries of the initial water surface photons according to the difference of the density distribution of the signal photons, and dividing the original single photons into photons above the water surface, photons on the water surface and photons under water according to the upper and lower boundaries of the initial water surface photons;