Intelligent classification method and system for complex scene in edge sea area
Through the information entropy measurement and spatial clustering of multimodal data, combined with convolutional neural network, a complex scene classification model in the border and sea areas is constructed, which solves the complex and diverse surface features and data quality problems, and achieves efficient scene classification and accuracy improvement.
Patent Information
- Application Number
- CN202410959517.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-12
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Classification and interpretation of complex scenes in border and sea areas faces the problem of complex and diverse surface features and difficult models to ensure accuracy. The existing methods require a lot of manual optimization and correction, and data quality and sample selection are difficult.
Multimodal data is used for information entropy measurement, geographical partitioning is divided through spatial clustering algorithm, scene classification model is built in combination with convolutional neural network, training and verification data is constructed using geographic partitioning level, landscape pattern index and remote sensing vegetation index are introduced, and depth is carried out. Learn model training and iterative optimization.
It improves the accuracy and efficiency of complex scene classification in border and sea areas, reduces the workload of manual assisted correction, and improves the classification accuracy and data utilization efficiency in complex surface features.
Smart Images

Figure CN118941845A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of spatial scene classification, and in particular to a method and system for intelligently classifying complex scenes in coastal areas. Background Art
[0002] With the advent of the era of widespread coverage of geographic big data represented by remote sensing and social perception, the classic land use / cover classification has gradually developed into scene classification. The scene originated from the scene recognition technology in the field of computer vision, and was later introduced into the field of interpretation and recognition of high-resolution remote sensing images, that is, scene classification, which extracts and classifies various land cover types and ground objects in remote sensing images into their respective semantic categories, especially urban built-up areas.
[0003] At present, scene classification is mostly concentrated in urban built-up areas, and there is little research on coastal and sea areas. In the document "Tang Yuzhi et al., Ecological carrying capacity and sustainable evaluation of coastal and marine areas based on spatial scenes. Acta Geographica Sinica 78(11), 2811-2832", the scene extraction and carrying capacity of the coastal and marine areas in the Greater Bay Area were studied based on the random forest method, but it needs to be supplemented by a lot of manual optimization and correction. Therefore, the accuracy and efficiency of intelligent interpretation in coastal and sea areas still need to be improved. Summary of the invention
[0004] The purpose of the present application is to provide a method and system for intelligent classification of complex scenes in coastal areas, so as to solve or alleviate the problem in the above-mentioned prior art that the classification and interpretation of complex scenes in coastal areas is faced with complex and diverse surface features and it is difficult for one model to ensure accuracy.
[0005] In order to achieve the above objectives, this application provides the following technical solutions:
[0006] This application provides a method for intelligent classification of complex scenes in coastal areas, including:
[0007] Obtain multimodal data of the coastal research area, including hyperspectral images and high-resolution image data obtained by various sensors;
[0008] The study area is divided into independent partition patches, and the spatial information entropy index of each partition patch is calculated based on the multimodal data, and the scene complexity of the surface space within the partition patch is measured using the spatial information entropy index to obtain a complexity measurement index;
[0009] Using the complexity measurement index as the basis for quantitative expression of features, based on the spatial clustering algorithm, the complex scene of the coastal research area is divided into multiple relatively homogeneous regions to obtain geographical divisions;
[0010] In combination with the complexity measurement index, the geographical partitions are graded, and training sample data and validation set data of each geographical partition at different levels are constructed according to the levels of the geographical partitions;
[0011] A convolutional neural network model CNN is used to construct scene classification models for different geographical divisions. Based on the training sample data and the validation set data, the scene classification models are trained and verified to obtain a final intelligent classification model for complex scenes in coastal areas.
[0012] In a possible implementation, after acquiring the multimodal data of the coastal research area, the method further includes:
[0013] The multimodal data is preprocessed, wherein the preprocessing includes denoising, radiation correction, and geometric correction.
[0014] In a possible implementation, the multimodal data further includes: point of interest POI data, DEM data, and basic geographic data;
[0015] Correspondingly, the study area is divided into independent partition patches, specifically:
[0016] Based on the DEM data, for the hilly and mountainous areas in the study area, the sub-basin boundaries are used to extract the ridge lines, and the surface runoff model is used to extract the valley lines to obtain the topographic linear elements in the study area;
[0017] Based on the topographic linear elements, other linear geographical elements in the study area and the constraints of the four-level ecological zoning of China's coastal waters, the study area is divided into independent zoning patches;
[0018] Among them, the other linear geographic elements include roads, rivers, and green belt elements in basic geographic data.
[0019] In a possible implementation, the spatial clustering algorithm is a K-means clustering method;
[0020] Correspondingly, based on the spatial clustering algorithm, the complex scene of the study area is divided into multiple relatively homogeneous regions to obtain geographical partitions, specifically:
[0021] Taking the complexity measurement index as the basis for quantitative expression of features, based on the K-means clustering method, the complex scenes of the border and sea study area are divided into multiple relatively homogeneous areas. The clustering results are evaluated using the Moran index and silhouette coefficient to achieve geographical zoning based on the complexity measurement index.
[0022] In a possible implementation, training sample data and validation set data for each geographical partition at different levels are constructed according to the level of the geographical partition, specifically:
[0023] The level of the geographical partition represents the complexity of the geographical partition. For a geographical partition with high complexity, the sample points selected for its training sample data and verification set data are more than those for a geographical partition with low complexity.
[0024] In a possible implementation, performing stratified selection on the sample points includes:
[0025] Following the principle of starting with the easy and then the difficult, stratified sampling was carried out in the order of "construction land - water body - ecological land - cultivated land";
[0026] The sample points are expanded using a sample time-space migration method.
[0027] In one possible implementation, a convolutional neural network model CNN is used to construct scene classification models for different geographical regions, specifically:
[0028] Use image processing and remote sensing technology to extract features suitable for classification from data obtained from different sensors;
[0029] Use feature selection methods to screen out the most representative and discriminative features and obtain the filtered classification features;
[0030] calculating a landscape pattern index based on the geographic divisions;
[0031] The landscape pattern index is used in combination with the screened classification features to construct a deep learning model for each geographical partition, thereby obtaining a scene classification model for different geographical partitions.
[0032] In a possible implementation, the screened classification features include color, texture, shape, spatial information, and remote sensing vegetation index;
[0033] Wherein, the remote sensing vegetation index includes NDVI index and EVI index, and both NDVI index and EVI index contain time series information;
[0034] The landscape pattern index includes area, density, edge, shape, aggregation, and diversity;
[0035] The scene classification model is built based on the U-Net model.
[0036] In a possible implementation, it further includes:
[0037] Evaluating the accuracy of scene classification output by the scene classification model;
[0038] According to the evaluation results and combined with field sample points, ensemble learning is performed on low-precision scene categories in a semi-supervised classification manner to update their attribute values. At the same time, new dynamic features are recalculated with reference to high-precision scene categories and added to the feature library for iterative ensemble training.
[0039] This embodiment provides an intelligent classification system for complex scenes in coastal areas, including:
[0040] A data acquisition unit configured to acquire multimodal data of the border sea study area, including hyperspectral images and high-resolution image data obtained by a variety of sensors;
[0041] An index calculation unit is configured to divide the coastal research area into mutually independent partition patches, and based on the multimodal data, use the information entropy index to measure the scene complexity of the surface space within the partition patches to obtain a complexity measurement index;
[0042] A geographic partitioning unit is configured to use the complexity measurement index as a basis for quantitative expression of features, and based on a spatial clustering algorithm, divide the complex scene of the border and sea research area into a plurality of relatively homogeneous regions to obtain geographic partitions;
[0043] A data set construction unit is configured to classify the geographical partitions in combination with the complexity measurement index, and construct training sample data and validation set data of each geographical partition at different levels according to the levels of the geographical partitions;
[0044] The model training unit is configured to use a convolutional neural network model CNN to respectively construct scene classification models for different geographical divisions, and to train and verify the scene classification models based on the training sample data and the verification set data to obtain a final intelligent classification model for complex scenes in coastal areas.
[0045] The technical solution of the embodiment of the present application has the following beneficial effects:
[0046] By dividing the study area into independent partition patches, the spatial information entropy index of each partition patch is calculated based on multimodal data, and the spatial information entropy index is used to measure the scene complexity of the surface space within the partition patch to obtain the complexity measurement index; using the complexity measurement index as the basis for quantitative expression of features, based on the spatial clustering algorithm, the complex scenes of the coastal study area are divided into multiple relatively homogeneous regions to obtain geographical partitions; combining the complexity measurement index, the geographical partitions are graded, and the training sample data and validation set data of each geographical partition at different levels are constructed according to the level of the geographical partition; the convolutional neural network model CNN is used to construct the scene classification model of different geographical partitions, and the scene classification model is trained and verified based on the training sample data and validation set data to obtain the final intelligent classification model of complex scenes in the coastal area. This method introduces geological knowledge, uses information entropy to measure scene complexity, and carries out model construction and scene extraction on the basis of geographical partitions, thereby improving the accuracy of scene classification in the case of complex and diverse surface features in the coastal area. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 A schematic diagram of the flow of a method for intelligently classifying complex scenes in coastal areas provided according to some embodiments of the present application.
[0048] Figure 2 A schematic diagram of the structure of an intelligent classification system for complex scenes in coastal areas provided according to some embodiments of the present application. DETAILED DESCRIPTION
[0049] This embodiment relates to the field of geographic information systems and artificial intelligence, and provides a method for intelligently extracting complex scene categories in coastal areas using deep learning methods.
[0050] First, the relevant terms of this embodiment are described in detail below.
[0051] In this embodiment, the scene refers specifically to a space scene.
[0052] Space and scene are two different perspectives in geographic analysis. Space uses the Cartesian coordinate system as a reference system, and uses quantitative methods such as geographic coordinates and geometric measurements to accurately express the location of geographic entities and phenomena, providing a measurable and computable analysis framework for the objective physical environment. The origin of the concept of scene is relatively complex: one source is extended from the scenes and scenes in movies and dramas, and has been widely used in communication, sociology and the Internet. For example, some scenes refer to the combination of urban life and entertainment facilities, while also conveying urban culture and values to different groups of people; the other comes from scene recognition technology in the field of computer vision, which was later introduced into the field of interpretation and recognition of high-resolution remote sensing images, that is, scene classification, using semantic information to classify the scenes in remote sensing images as a whole. The former emphasizes the non-material morphological characteristics of the scene, that is, the metaphysical cultural connotation, value orientation and the identity of lifestyle; the latter emphasizes the material morphological characteristics of the scene, that is, the similarity of land use, surface cover and landscape structure; both explain the connotation of the scene from different latitudes. However, the existing scenario concepts mostly focus on the classification of urban consumption space and urban land use, similar to "urban functional zoning", ignoring the ecological dimension of geographical space. In recent years, ecological environmental protection has received increasing attention, and the concept of ecological zoning has emerged. Ecological use of the sea, protection and improvement of the marine ecological environment have also become the central purpose of marine functional zoning in recent years. Therefore, the ecological dimension needs to be added to the scenario concept, that is, when classifying scenarios, the ecological environmental sensitivity and ecological function of the geographical space, as well as the impact of human activities on the ecological environment, need to be considered.
[0053] The spatial scene in this embodiment is a geographical unit with coordinate location, highly unified socio-economic attributes, surface cover, ecological functions and externalities. The spatial scene has multi-dimensional attributes such as nature, society, economy and ecology, and is a distribution of the differences in functional attributes of different geographical units in space. It is a higher level of cognition and synthesis than traditional land cover / utilization, taking into account the differences in land cover due to different land use functions or locations; it is also different from functional zoning, not guided by planning goals, but based on the quantitative response provided by the social ecological functions of the objective world.
[0054] In this embodiment, neural network technology can be used to automatically identify and classify different scene types according to the content in the multimodal data. The multimodal data can be image or video data, and the image data can be remote sensing data of multi-platform sensors.
[0055] In this embodiment, all scene types, that is, scene categories, can be pre-acquired based on the coastal space scene classification system. Among them, the coastal space scene classification system can adopt a three-level classification structure or a two-level classification structure. In an example of a three-level classification structure, the first-level scene can include, for example, sea and land space, and the second-level scene can include, for example, agricultural space, green space, wetland space, town space, nearshore sea space, etc. The third-level scene is further divided on the basis of the second-level scene. For example, the nearshore sea space in the second-level scene can be further divided into three-level scenes such as woodland, sparse forest, grassland, wetland, marine protected sea area, development and reserved sea area, and special use sea area. In an example of a two-level classification structure, the space scene can be divided into two levels: the first-level scene and the second-level scene, wherein the first-level scene includes agricultural, urban, ecological, and industrial space scenes, and each first-level scene includes multiple second-level scenes. In the first-level space scene, the agricultural function scene can include the following second-level space scenes: economic forest, paddy field, dry land, and aquaculture land. Urban functional scenes can include the following secondary space scenes: industry, services, residence, commerce, and roads; ecological functional scenes can include the following secondary space scenes: other woodlands, grasslands, water bodies, and unused land.
[0056] The goal of scene classification is to classify different areas in coastal areas into different scene categories based on multimodal data. However, coastal areas have different natural environments, complex geopolitical patterns, and are difficult to reach for observation. Therefore, remote sensing surveys of complex scenes in coastal areas face a series of severe challenges. First, coastal areas contain a variety of landforms, vegetation, and water types, as well as human activity areas. Their surface features are diverse and complex, which brings challenges to classification and identification. Secondly, coastal areas usually have harsh geographical environments, with more cloudy and rainy weather, and data coverage and data quality are often difficult to guarantee. The borders of the sea and land are intertwined, and the coastline is changeable. The remote sensing data obtained may have problems such as noise, distortion, or inconsistency. For example, extreme weather and ocean fluctuations may affect data acquisition and data quality. Thirdly, the distribution of different land feature categories in coastal areas may be uneven, and the number of samples in some categories is small, which will cause the training of the classifier to be biased towards categories with a larger number, affecting the balance and comprehensiveness of the classification. Therefore, the classification and interpretation of complex scenes in coastal and coastal areas faces challenges such as complex and diverse surface features, difficulty in ensuring high-quality data, and difficulty in selecting samples. It is necessary to build targeted intelligent interpretation technology for complex scenes in coastal and coastal areas to improve the accuracy of intelligent classification of complex scenes in coastal and coastal areas.
[0057] In view of this, the present embodiment provides a method and system for intelligent classification of complex scenes in coastal and coastal areas. Based on multi-source data (multimodal data), geological knowledge (spatial information entropy and landscape indicators) is introduced and combined with artificial intelligence technology (convolutional neural network model CNN) to construct an intelligent interpretation model (scene classification model). Specifically, spatial information entropy is used to measure scene complexity, and model construction and scene extraction are carried out on the basis of geographical zoning, so as to take into account the complex scenes in coastal and coastal areas, improve the accuracy and efficiency of intelligent extraction of complex scenes in coastal and coastal areas, and at the same time deal with the complex and diverse surface features in coastal and coastal areas.
[0058] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.
[0059] The terms "first", "second", "third" and "fourth" etc. in the specification and claims of the present application and the drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.
[0060] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.
[0061] "Multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the related objects are in an "or" relationship.
[0062] This embodiment provides a method for intelligent classification of complex scenes in coastal areas. The execution subject of this method can be a personal computer, server, mobile phone, tablet and other electronic devices. Figure 1 As shown, the method includes:
[0063] Step S101: Acquire multimodal data of the coastal research area, where the multimodal data includes hyperspectral images and high-resolution image data obtained by various sensors.
[0064] Multimodal data refers to multiple types of data from different sensors or data sources, which can provide multiple aspects of information about the same phenomenon or object. Multimodal data recognition technology can understand and analyze complex phenomena and problems more comprehensively and accurately by integrating these different modal data.
[0065] In remote sensing technology, sensors are devices used to detect and record electromagnetic radiation reflected or emitted from the Earth's surface or atmosphere. Sensors are the core components of remote sensing systems. They can be carried on platforms such as satellites, aircraft, and drones to obtain high-resolution surface or atmospheric information (high-resolution image data).
[0066] The sensors in this embodiment may include various types, such as optical sensors, microwave sensors, thermal infrared sensors, etc., wherein the optical sensor can effectively detect electromagnetic radiation in the visible light, near infrared and short-wave infrared regions, and obtain information such as land cover and vegetation in a large range. In this embodiment, the data obtained by the optical sensor may be, for example, optical images such as Landsat. Microwave sensors are used to detect electromagnetic radiation in the microwave region. They can penetrate clouds and vegetation and work around the clock. Therefore, they are not limited by weather conditions and can obtain data under cloudy, foggy, rainy and snowy conditions. In this embodiment, the data obtained by the microwave sensor may include radar data, such as Sentinel-1 or Sentinel-2 image data.
[0067] In this embodiment, a hyperspectral image refers to a spectral image with a spectral resolution in the micrometer range, which can simultaneously image the target area in the ultraviolet, visible, near-infrared and mid-infrared regions of the electromagnetic spectrum with tens to hundreds of continuous and subdivided spectral bands through hyperspectral sensors (such as imaging spectrometers) carried on different space platforms. In this embodiment, the hyperspectral image can be obtained through platforms such as MODIS (moderate resolution imaging spectrometer), or by carrying hyperspectral sensors on platforms such as drones and space shuttles to obtain the hyperspectral image.
[0068] This embodiment uses a variety of sensors (such as satellites, aircraft payloads, etc.) to obtain multimodal data in complex coastal and coastal areas, including optical images such as Landsat, radar data such as Sentinel, hyperspectral images such as MODIS, and high-resolution image data, etc., which can provide richer data resources for the training of convolutional neural network models and fully understand and process information. By training with multimodal data, the model can learn more comprehensive and complex feature representations, thereby improving its performance and generalization ability.
[0069] Step S102: Divide the study area into independent partition patches, and calculate the spatial information entropy index of each partition patch based on multimodal data, and use the spatial information entropy index to measure the scene complexity of the surface space within the partition patch to obtain a complexity measurement index.
[0070] In this embodiment, taking into account the differences caused by the regular differentiation of various components of the natural geographical environment along the surface, that is, regional heterogeneity, geological knowledge is first introduced, and spatial information entropy is used to measure the complexity of the scene. Based on this, geographical divisions are established, and subsequent spatial scene classification research is carried out on this basis.
[0071] Among them, spatial information entropy is mainly an indicator to measure the degree of chaos and uncertainty in the distribution of geographic spatial data.
[0072] The calculation of spatial information entropy requires that the study area be divided into independent partition patches. In this embodiment, there are many ways to divide the partition patches. For example, the study area can be divided into several regular grid units by drawing a grid. Each grid unit is a partition patch, and each unit can be regarded as a discrete spatial state. It is also possible not to draw a grid, but to directly regard each pixel (grid) in the remote sensing image as an independent partition. In addition, the existing fine-scale land use status data can also be used to extract partition patches. This embodiment does not limit the method of dividing the study area into partition patches.
[0073] After dividing the partition patches, the spatial information entropy index of each partition patch can be calculated based on the multimodal data, specifically as follows: according to the surface type to which each partition patch belongs, the number of different surface types in different partition patches is counted, and the probability distribution of the surface type represented by it is calculated, and then the information entropy value is calculated according to the probability distribution. Among them, the information entropy value can be calculated according to the existing information entropy formula, which will not be described in detail in this embodiment.
[0074] Furthermore, the above-mentioned calculation process of spatial information entropy can be implemented in detail through ArcGIS software. For example, spatial resolution alignment can be performed first, that is, the spatial resolution of multimodal data can be unified so that data from various sources can be operated at the same spatial resolution, that is, the grid sizes can be equal. Then, the grid calculator in ArcGIS software is used to construct the information entropy formula and calculate the information entropy value of each grid.
[0075] This embodiment uses the spatial information entropy index to measure the scene complexity of the surface space in the partition patch to obtain a complexity measurement index. This index represents the surface complexity of each partition patch. The larger the spatial information entropy index, the higher the surface complexity and the higher the difficulty of spatial scene classification.
[0076] Step S103: Using complexity measurement indicators as the basis for quantitative expression of features, based on the spatial clustering algorithm, the complex scenes of the border and sea research area are divided into multiple relatively homogeneous regions to obtain geographical divisions.
[0077] Among them, spatial clustering algorithm is a type of technology used to discover patterns and structures in geographic spatial data. It divides spatial data into different groups or clusters so that data points within the same cluster are more similar than data points between different clusters, thereby revealing the distribution and trends in spatial data.
[0078] In this embodiment, the spatial clustering algorithm may be, for example, a hierarchical clustering algorithm, a K-means clustering method, or the like.
[0079] Specifically, in an optional embodiment, a hierarchical clustering algorithm can be used to implement spatial clustering. The main idea of the hierarchical clustering algorithm is to gradually merge or split clusters by constructing a cluster tree (or dendrogram) to achieve the purpose of clustering. Based on the hierarchical clustering algorithm, the complex scene of the border sea study area is divided into multiple relatively homogeneous regions to obtain geographical partitions, which can be achieved by the following steps:
[0080] (1) Initialization: The partition patches (pixels or grids) are input as samples into the hierarchical clustering algorithm. The hierarchical clustering algorithm regards each partition patch in the data set as a separate cluster. That is, if there are n partition patches in the data set, there are n clusters initially.
[0081] (2) Calculate similarity: Calculate the similarity (or distance) between all clusters. There are many methods for calculating similarity or distance, such as Euclidean distance, Manhattan distance, cosine similarity, etc.
[0082] (3) Merge clusters: Find the two clusters with the highest similarity (or the smallest distance) and merge them into a new cluster. After the merger, the total number of clusters decreases by one.
[0083] (4) Update similarity: Use the weighted sum method to calculate the total complexity measure of all samples in the newly generated cluster, determine the center of the newly generated cluster based on the total complexity measure, and recalculate the similarity between the center of the newly generated cluster and other clusters. The weights of each sample can be set to equal weight or the coefficient of variation of the complexity measure.
[0084] (5) Repeat steps (3) and (4) until a preset termination condition is reached. The termination condition may be that a preset number of clusters is reached, or the similarity of a cluster is lower than a certain threshold, indicating that further merging is not possible.
[0085] (6) Output the final clustering result, that is, the cluster to which each sample point belongs.
[0086] The above is the process of the hierarchical clustering algorithm. Taking into account the high computational complexity of the hierarchical clustering algorithm, especially when the data set is large, the choice of merging or splitting points may affect the clustering results. Therefore, as a preferred solution, this embodiment specifically selects the K-means clustering method to achieve spatial clustering, including the following steps: using complexity metrics as the basis for quantitative expression of features, based on the K-means clustering method, the complex scenes of the border and sea study area are divided into multiple relatively homogeneous regions, and the Moran index and silhouette coefficient are used to evaluate the clustering results to achieve geographical zoning based on complexity metrics.
[0087] Exemplarily, the K-means clustering method can be implemented in the following steps to achieve spatial clustering:
[0088] Step S1031: Consider the partition patch as a data point, and randomly select K data points from the data set consisting of the partition patches as initial cluster centers. These center points represent the initial position of each cluster.
[0089] Step S1032: for each data point in the data set (the data point has a spatial information entropy value), calculate its distance to each cluster center (which can be the Euclidean distance or other suitable distance measurement), and assign each point to the nearest center point to form K clusters;
[0090] Step S1033: Update the cluster center: for each cluster, recalculate the average value or weighted average value of the spatial information entropy of all data points in the cluster, and use this average value as the new cluster center;
[0091] Step S1034: repeating steps S1032 and S1033 until the cluster center no longer changes significantly (i.e., the distance between the new and old cluster centers is less than a preset threshold) or a predetermined number of iterations is reached;
[0092] Step S1035: Evaluate clustering effect: Use Moran's index and silhouette coefficient to evaluate clustering effect, and check whether the clustering result conforms to the expected spatial distribution or characteristic pattern. If so, end the iteration and output the final clustering result. If not, reselect K data points as the initial clustering centers, and iterate steps S1031 to S1035 until the clustering result meets the actual needs, and finally realize geographic partitioning based on complexity measurement indicators.
[0093] Among them, the Moran index and silhouette coefficient can be calculated by ArcGIS software.
[0094] The core idea of the K-means clustering method adopted in this embodiment is to iteratively assign data points to the nearest cluster centers and then update the positions of these cluster centers. This process is simple and computationally efficient. Especially for large data sets such as high-resolution remote sensing images, the K-means clustering method can give clustering results within a reasonable time, which can effectively improve the efficiency of classification of complex coastal and sea scenes.
[0095] Step S104: The geographical partitions are graded in combination with the complexity measurement index, and training sample data and validation set data of each geographical partition at different levels are constructed according to the levels of the geographical partitions.
[0096] In this embodiment, the geographical partitions are graded using complexity metrics, and the geographical partitions are divided into multiple levels of different complexity. According to the level of the geographical partition, training sample data and verification set data of each geographical partition at different levels are constructed. By adopting objective data and scientific calculation methods, the accuracy and rationality of the classification are ensured. In addition, geographical partitions at different levels have significant differences in natural environment, resource distribution, human activities, etc. Through classification, these differences can be identified more accurately, which helps to more accurately describe and understand the structure and characteristics of the geographic space, and provide a more targeted basis for subsequent data processing and model construction. In addition, since the data of geographical partitions at different levels differ in complexity and characteristics, targeted training can ensure that the model achieves optimal performance on data at a specific level, avoids mixing data of all levels together for training, reduces the impact of redundant data on model performance, and thus improves data utilization efficiency.
[0097] Step S105: Use the convolutional neural network model CNN to build scene classification models for different geographical divisions, train and verify the scene classification models based on the training sample data and the validation set data, and obtain the final intelligent classification model for complex scenes in coastal areas. The intelligent classification model for complex scenes in coastal areas can be used to perform scene intelligent classification on the research area to be classified, effectively improving the accuracy and efficiency of intelligent classification of complex scenes in coastal areas.
[0098] In this embodiment, the powerful feature extraction capability of the convolutional neural network model CNN is used to automatically extract feature representations useful for classification tasks from the original image, which helps the model to more accurately identify scene categories in different geographical divisions. By training on the training sample data and verifying on the validation set data, the performance of the model can be comprehensively evaluated, which helps to timely discover problems with the model and make improvements. After sufficient training and verification, the model can more stably handle scene classification tasks in different geographical divisions, improving the reliability and stability of the model in practical applications.
[0099] In summary, the intelligent classification method for complex scenes in coastal and coastal areas provided in this embodiment uses information entropy to measure scene complexity, and carries out targeted model construction and scene extraction based on geographical divisions, thereby reducing the workload of later manual auxiliary corrections and improving the accuracy and efficiency of intelligent interpretation in coastal and coastal areas.
[0100] Based on the above embodiments, some optional embodiments, after acquiring the multimodal data of the border sea research area, further include: preprocessing the multimodal data, the preprocessing includes denoising, radiation correction, and geometric correction to ensure data quality and consistency.
[0101] In the aforementioned embodiments, the method of dividing into regular grid units or dividing into partition patches from fine-scale land use status data is exemplarily adopted. Considering that the regular grid unit division still cannot well represent the spatial heterogeneity of coastal areas under complex scenes, and the land use status data are usually difficult to obtain, the accuracy of geographic partitioning is still insufficient. Based on this, in some embodiments, the multimodal data also includes: point of interest POI data, DEM data, and basic geographic data. These data are used as a supplement to the high-resolution remote sensing image data of various sensors and participate in the construction process of geographic partitioning, which can make full use of the advantages of complementary multi-source data and improve the accuracy of geographic partitioning.
[0102] Among them, high-resolution remote sensing image data can provide large-scale, multi-temporal surface information, including vegetation coverage, water distribution, land use type, etc., providing a macro background for geographical zoning. The advantage of remote sensing data is that it has strong macro observation capabilities, can cover a wide area, and can be updated regularly to reflect surface changes. POI data can contain specific location information, such as schools, hospitals, shopping malls, etc. These locations are usually closely related to human activities and can reflect the socio-economic characteristics of the region. The addition of POI data makes geographical zoning more refined and can accurately identify areas with specific functions, such as commercial areas and residential areas. DEM data can provide surface elevation information, which is used to identify natural geographical elements such as terrain undulations and water system distribution, and provide important terrain basis for geographical zoning. Basic geographic data can include basic geographic information such as administrative divisions, road networks, and water systems. It is the basic framework for geographical zoning. The accuracy of basic geographic data directly affects the accuracy and reliability of geographical zoning.
[0103] On the basis of obtaining point of interest POI data, DEM data, and basic geographic data, the study area is divided into independent partition patches. Specifically, for the hilly and mountainous areas in the study area, the sub-basin boundary is used to extract the ridge line, and the surface runoff model is used to extract the valley line to obtain the topographic linear elements in the study area. Based on the topographic linear elements, other linear geographic elements in the study area (such as roads, rivers, green belts, etc.) and the constraints of the four-level ecological zoning of China's nearshore waters are considered to divide the study area into independent partition patches.
[0104] Among them, the China Coastal Waters Ecological Four-Level Zoning (Trial) was released on December 4, 2023. It is the latest achievement of my country's marine ecological classification and zoning. It established a marine ecological classification framework and standardized marine ecological types. This achievement divides my country's coastal waters into three ecological first-level zones based on latitude zones and geographical contours; on the basis of the first-level zones, 22 ecological second-level zones are divided according to water depth and water mass distribution; on the basis of the second-level zones, 53 ecological third-level zones are further divided according to the first-level landform types and supplemented by water body characteristics. On this basis, 132 ecological fourth-level zones are also divided. These ecological fourth-level zones focus on coastal areas with frequent human activities, laying a high-quality data foundation for the spatial scene classification of coastal areas.
[0105] In order to obtain highly accurate geographic zoning, this embodiment combines multimodal data with the four-level ecological zoning of China's coastal waters. First, for hilly and mountainous areas, sub-basin boundaries are used to extract ridge lines, and surface runoff models are used to extract valley lines and other terrain linear elements. The study area is divided into independent partition patches based on the constraints of linear geographical elements such as terrain lines, roads, rivers, green belts, and the four-level ecological zoning of China's coastal waters. The information entropy index is used to measure the scene complexity of the surface space. Then, based on spatial clustering algorithms such as K-means clustering, the complex scene is divided into multiple relatively homogeneous areas. The clustering results are evaluated using indicators such as the Moran index and silhouette coefficient to achieve geographic zoning based on complexity measurement indicators.
[0106] Among them, the following steps can be included by integrating linear geographical elements such as terrain lines, roads, rivers, green belts and the constraints of the four-level ecological zoning of China's coastal waters: first, the linear geographical elements such as terrain lines, roads, rivers, green belts and the four-level ecological zoning of China's coastal waters are overlaid and analyzed, and the results of the overlay analysis are used to determine whether there are conflicts between the various geographical elements and the impact of different terrains on the marine ecological zoning; then, based on the linear geographical elements such as terrain lines, roads, rivers and green belts, as well as the constraints of the four-level marine ecological zoning, preliminary block division is carried out. When dividing, attention should be paid to maintaining the ecological consistency within the block, while considering the reasonable layout of human activities. According to the preliminary division results, optimization and adjustment are carried out in combination with the results of the overlay analysis and the actual situation. For the blocks with conflicts, the rationality of their division needs to be re-evaluated, and appropriate adjustments should be made if necessary to obtain the final results of the zoning patch division.
[0107] In order to further improve the feature expression ability and model performance of training samples, in some embodiments, the training samples are selected by partitioning and stratification. Specifically, the training sample data and validation set data of each geographical partition at different levels are constructed according to the level of geographical partition, which can include two steps: determining the number of training samples and validation samples according to different geographical partitions and implementing stratified selection of sample points. Among them, the number of training samples and validation samples is determined according to the level of geographical partition. Specifically, the level of geographical partition represents the complexity of geographical partition. For geographical partitions with high complexity, the sample points selected for its training sample data and validation set data are more than those for geographical partitions with low complexity. At the same time, stratified selection is implemented for sample points, following the principle of starting with the easy and then the difficult, and stratified sampling is carried out in the order of "construction land-water body-ecological land-arable land"; the sample points are expanded by using the method of sample spatiotemporal migration.
[0108] In this embodiment, the geographic partitions are graded in combination with complexity metrics, and the weight of the number of training samples is determined according to the level of the geographic partition, that is, different numbers of training samples are allocated based on the complexity of the geographic partition, and the weight of these samples in the model training process is determined accordingly. Since partitions with high complexity often contain more diverse geographical features, environmental conditions or human factors, in high-complexity partitions, the increase in the number of sample points helps to reduce the deviation caused by sparse or uneven data distribution, so that the model can more accurately reflect the real situation of the area. By increasing the number of sample points in these areas and improving the scientificity and rationality of sample selection, it can be ensured that the model can learn more diverse data features during the training process, thereby improving the prediction and classification accuracy of the model in complex environments.
[0109] Among them, the number of training samples and validation samples for each geographical partition at different levels can be determined by the proportional allocation method, specifically, directly setting the sample allocation ratio corresponding to geographical partitions of different complexity levels. For example, the geographical partition with the highest complexity can be allocated 40% of the samples, the second highest 30%, and so on. The empirical judgment method can also be used to directly judge and allocate the number of samples based on the experience and knowledge of domain experts and their understanding of the complexity of each geographical partition. The complexity can also be combined with the spatial distribution of data. The data distribution (the density and drastic changes of geographical elements) under each geographical partition can be analyzed through the historical data of the study area, and the number of samples can be allocated according to the characteristics of the data distribution, such as allocating more samples to geographical partitions with dense or drastic data changes.
[0110] When selecting sample points in layers, the stratified sampling method is adopted, including the following steps:
[0111] (1) The four categories are stratified in the order of “construction land – water body – ecological land – cultivated land”. These four categories represent different land use types and have clear distinction and representativeness.
[0112] Specifically, first, all construction land (such as urban built-up areas, industrial areas, commercial areas, etc.) in the study area are divided into one level. Since construction land usually has clear boundaries and high development intensity, it is relatively easy to identify and divide. Secondly, all water bodies (such as rivers, lakes, reservoirs, ponds, etc.) in the study area are divided into one level. Since water bodies usually have natural fluidity or stillness and have obvious characteristics on remote sensing images or maps, they are also relatively easy to identify. Thirdly, ecological land (such as woodland, grassland, wetland, nature reserve, etc.) in the study area are divided into one level. Ecological land usually has high ecological value and biodiversity, but it may be difficult to accurately divide due to factors such as terrain and vegetation cover. Therefore, when dividing, remote sensing technology, geographic information system (GIS) and other tools can be used for auxiliary identification. Finally, all cultivated land (including paddy fields, dry land, etc.) in the study area are divided into the last level. Cultivated land is the basis of agricultural production, with a wide distribution and clear utilization methods, but it may intersect or overlap with other types of land in some areas.
[0113] (2) According to the number of sample points (including training samples and validation samples) in different geographical partitions, random sampling is performed in each layer, and the label information of the sample points (i.e., the scene type to which the sample points belong) is collected to obtain the training sample data and validation set data of each geographical partition.
[0114] In order to reduce the workload of sample selection, the samples can be expanded by combining methods such as sample spatiotemporal migration. For example, if the objects have strong temporal continuity, the source domain samples can be directly migrated to the target domain for use as samples.
[0115] Among them, sample spatiotemporal migration in the field of remote sensing refers to the process of applying sample data at a specific time (time) or space (space) position to another time or space position after appropriate processing or conversion. This embodiment uses the method of sample spatiotemporal migration to expand the sample, which can make full use of existing data resources, reduce the demand for new data collection, improve data utilization efficiency, and enhance the generalization ability of the model.
[0116] In some embodiments, a convolutional neural network model CNN is used to construct scene classification models for different geographical partitions, and the steps of feature extraction and selection are also included, specifically: using image processing and remote sensing technology to extract features suitable for classification from data obtained by different sensors, including color, texture, shape, spatial information, etc.; using feature selection methods to screen out the most representative and discriminative features, and obtain screened classification features to reduce classification complexity and improve classification accuracy; calculating landscape pattern index based on geographical partitions; using landscape pattern index, combined with screened classification features (such as color, texture, shape, spatial information and remote sensing vegetation index and other indicators) to construct deep learning models for each geographical partition, and obtain scene classification models for different geographical partitions. The above steps make full use of the comprehensive advantages of spatial features and spectral information to achieve intelligent interpretation based on geographical partitions.
[0117] Based on the above embodiment, as an optional scheme, the filtered classification features include color, texture, shape, spatial information and remote sensing vegetation index; wherein the remote sensing vegetation index includes NDVI index and EVI index, and both NDVI index and EVI index contain time series information; the landscape pattern index includes area, density, edge, shape, aggregation and diversity.
[0118] In one example, feature extraction and selection can be implemented in the following steps: First, the landscape pattern index in the study area is obtained based on the geographic partition. These indexes include the area, density, edge, shape, aggregation, diversity, etc. of the landscape index within the partition and the overall landscape for calculating the object, which provide a basis for understanding and analyzing the spatial structure of the geographic region; the calculation of the landscape pattern index within the partition can be combined with the simple classification of specific cover types according to the vegetation index, water body index and impervious index, and then the corresponding landscape index is calculated. Secondly, the landscape pattern index (area, density, edge, shape, aggregation, diversity), color, texture, spatial information (aspect ratio, area / perimeter ratio, etc.) and remote sensing vegetation index (NDVI, EVI, etc.) are further integrated, such as the time series information of the vegetation index, texture index and spectral information. Combining the landscape pattern characteristics of the geographic partition and the vegetation detail information of the remote sensing data, through the powerful learning ability of the deep learning model, the accuracy and efficiency of the intelligent interpretation of the geographic partition can be effectively improved, and more accurate spatial scene classification can be achieved.
[0119] In this embodiment, the scene classification model (also called the intelligent interpretation model) is built based on the U-Net model.
[0120] Specifically, in the construction of the intelligent interpretation model, TensorFlow is used for deep model training through the combination of Google Colab cloud service platform and Google EarthEngine geographic cloud computing platform. Based on the training and validation set data, the U-Net model of the convolutional neural network (CNN) model is used to train and verify the spatial scenes under the geographic partition, and the results are corrected. Among them, the U-Net model is a CNN architecture specially designed for image segmentation. It can effectively combine high-level contextual information and low-level detail information by using jump connections, which helps to efficiently extract large-scale geographic partition features and fine-scale remote sensing vegetation information.
[0121] In the specific model training process, the geographical division based on landscape pattern characteristics is used as the interpretation unit. In order to control the residual back propagation and the stability of the network, the initial value of the learning rate is set to 1×10 -4 , Based on the random initialization parameters of the Gaussian algorithm, the SGD stochastic gradient descent method is used to train the model, the Adam optimizer is used to accelerate the training process of the neural network, and ReLU is used as the activation function.
[0122] In order to further improve the accuracy of the scene classification model, some embodiments also include a scene extraction optimization step, which is as follows: using accuracy (Ac), recall rate (Recall) and Kappa coefficient as accuracy evaluation indicators, the scene category accuracy output by the scene classification model is evaluated; based on the evaluation results, combined with field sample points, ensemble learning is performed on low-precision scene categories in a semi-supervised classification manner to update their attribute values; at the same time, new dynamic features are recalculated with reference to high-precision scene categories, and added to the feature library for iterative ensemble training.
[0123] Among them, the field sample points can be obtained through field surveys or a pre-built sample point database, and the field sample points can include labeled samples (labeled data) and unlabeled samples (unlabeled data).
[0124] Specifically, after collecting field sample points, labeled samples can be used to train multiple initial classifiers, called pre-trained classifiers. It can be understood that this classifier can be a traditional machine learning model (such as SVM, decision tree, etc.) or a deep learning model (such as convolutional neural network CNN). Then, the pre-trained classifier is used to predict the unlabeled samples to obtain pseudo labels. Here, a threshold (such as confidence) can be set to screen pseudo-label samples with high confidence to reduce the impact of noise. Subsequently, iterative training is performed, including: merging the pseudo-label samples with the original labeled samples to form a new training set; retraining the classifier with the new training set and updating the model parameters; repeating the above process, iterating the training continuously until the performance of the classifier is no longer significantly improved or reaches the preset number of iterations, and obtaining the semi-supervised classification results. Finally, a suitable integration strategy (such as voting method, averaging method or stacking method) is selected to combine the prediction results of multiple classifiers to obtain a synthetic classification result, and then based on the classification result, the attribute values of the scene categories with low precision are updated (the updated content may include category labels, feature descriptions, etc.). At the same time, new dynamic features are recalculated based on the scene categories with high accuracy, and added to the feature library for iterative integrated training. The above steps continuously improve the scene classification accuracy.
[0125] In some optional embodiments, after using the intelligent classification model for complex scenes in coastal areas to classify the scenes in the study area to be classified, the step of generating a scene classification map is also included. Specifically: within each geographical block, a pair of boundary smoothing and beautification tools are used, and through the GIS overlay tool and the narrow and small patch processing tool, a layer-by-layer combination is formed to form a scene classification map with full coverage, fine classification, and smooth edges.
[0126] In summary, the intelligent classification method for complex scenes in coastal areas provided by this embodiment adopts information entropy to measure scene complexity, carries out model construction and scene extraction on the basis of geographical partitioning, introduces geographical partitioning, and constructs a deep learning model on the basis of partitioning to solve the problem that the classification and interpretation of complex scenes in coastal areas is faced with complex and diverse surface features, and it is difficult for one model to ensure accuracy. By adopting multimodal data, comprehensively using optical remote sensing, radar remote sensing and other images, POI, DEM and landscape index data, and combining time series information, it can make full use of remote sensing images and other data of past years to solve the problem that high-quality data is difficult to ensure. On the basis of hierarchical partitioning, sample point selection is carried out, and a deep learning model is constructed by integrating indicators such as landscape pattern index, color, texture, spatial information and remote sensing vegetation index to ensure the balance and comprehensiveness of sample points as much as possible, and solve the problem that samples are difficult to select and the representativeness of features is insufficient. At the same time, this embodiment constructs a deep learning model by partitioning to cope with the problem of large workload of manual auxiliary correction in the later stage.
[0127] This embodiment provides an intelligent classification system for complex scenes in coastal areas. Figure 2 As shown, the system includes:
[0128] The data acquisition unit 201 is configured to acquire multimodal data of the border sea research area, including hyperspectral images and high-resolution image data obtained by various sensors;
[0129] The index calculation unit 202 is configured to divide the coastal research area into independent partition patches, and measure the scene complexity of the surface space within the partition patches using the information entropy index based on the multimodal data to obtain the complexity measurement index;
[0130] The geographic partition unit 203 is configured to use the complexity measurement index as a basis for quantitative expression of features, and based on a spatial clustering algorithm, divide the complex scene of the border and sea research area into a plurality of relatively homogeneous regions to obtain geographic partitions;
[0131] The data set construction unit 204 is configured to classify the geographical partitions in combination with the complexity measurement index, and construct training sample data and validation set data of each geographical partition at different levels according to the levels of the geographical partitions;
[0132] The model training unit 205 is configured to use a convolutional neural network model CNN to construct scene classification models for different geographical divisions, and to train and verify the scene classification models based on training sample data and verification set data to obtain a final intelligent classification model for complex scenes in coastal areas.
[0133] The intelligent classification system for complex scenes in coastal areas provided in this embodiment can implement the steps and processes of the intelligent classification method for complex scenes in coastal areas provided in any of the above embodiments, and achieve the same technical effects, which will not be described one by one here.
[0134] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. An intelligent classification method for complex scenes in coastal areas, characterized in that: include: Obtain multimodal data of the coastal research area, including hyperspectral images and high-resolution image data obtained by various sensors; The study area is divided into mutually independent partition patches, and the spatial information entropy index of each partition patch is calculated based on the multimodal data, and the scene complexity of the surface space within the partition patch is measured using the spatial information entropy index to obtain a complexity measurement index; Taking the complexity measurement index as the basis for quantitative expression of features, based on the spatial clustering algorithm, the complex scene of the coastal research area is divided into multiple relatively homogeneous regions to obtain geographical divisions; In combination with the complexity measurement index, the geographical partitions are graded, and training sample data and validation set data of each geographical partition at different levels are constructed according to the levels of the geographical partitions; A convolutional neural network model CNN is used to construct scene classification models for different geographical divisions. Based on the training sample data and the validation set data, the scene classification models are trained and verified to obtain a final intelligent classification model for complex scenes in coastal areas.
2. The method according to claim 1, characterized in that After obtaining multimodal data of the coastal research area, it also includes: The multimodal data is preprocessed, wherein the preprocessing includes denoising, radiation correction, and geometric correction.
3. The method according to claim 1, characterized in that The multimodal data also includes: point of interest POI data, DEM data, and basic geographic data; Correspondingly, the study area is divided into independent partition patches, specifically: Based on the DEM data, for the hilly and mountainous areas in the study area, the sub-basin boundaries are used to extract the ridge lines, and the surface runoff model is used to extract the valley lines to obtain the topographic linear elements in the study area; Based on the topographic linear elements, other linear geographical elements in the study area and the constraints of the four-level ecological zoning of China's coastal waters, the study area is divided into independent zoning patches; Among them, the other linear geographic elements include roads, rivers, and green belt elements in basic geographic data.
4. The method according to claim 1, characterized in that: The spatial clustering algorithm is a K-means clustering method; Correspondingly, based on the spatial clustering algorithm, the complex scene of the study area is divided into multiple relatively homogeneous regions to obtain geographical partitions, specifically: Taking the complexity measurement index as the basis for quantitative expression of features, based on the K-means clustering method, the complex scenes of the border and sea study area are divided into multiple relatively homogeneous areas. The clustering results are evaluated using the Moran index and silhouette coefficient to achieve geographical zoning based on the complexity measurement index.
5. The method according to claim 1, characterized in that According to the level of the geographical partition, the training sample data and the validation set data of each geographical partition at different levels are constructed, specifically: The level of the geographical partition represents the complexity of the geographical partition. For a geographical partition with high complexity, the sample points selected for its training sample data and verification set data are more than those for a geographical partition with low complexity.
6. The method according to claim 5, characterized in that The sample points are selected in layers, including: Following the principle of starting with the easy and then moving on to the difficult, stratified sampling was carried out in the order of "construction land - water bodies - ecological land - cultivated land"; The sample points are expanded using a sample time-space migration method.
7. The method according to claim 1, characterized in that The convolutional neural network model CNN is used to construct scene classification models for different geographical regions, specifically: Use image processing and remote sensing technology to extract features suitable for classification from data obtained from different sensors; Use feature selection methods to screen out the most representative and discriminative features and obtain the filtered classification features; calculating a landscape pattern index based on the geographic divisions; The landscape pattern index is used in combination with the screened classification features to construct a deep learning model for each geographical partition, thereby obtaining a scene classification model for different geographical partitions.
8. The method according to claim 7, characterized in that The selected classification features include color, texture, shape, spatial information and remote sensing vegetation index; Wherein, the remote sensing vegetation index includes NDVI index and EVI index, and both NDVI index and EVI index contain time series information; The landscape pattern index includes area, density, edge, shape, aggregation, and diversity; The scene classification model is built based on the U-Net model.
9. The method according to claim 1, characterized in that: Also includes: Evaluating the accuracy of scene classification output by the scene classification model; Based on the evaluation results and combined with field sample points, ensemble learning is performed on low-precision scene categories in a semi-supervised classification manner to update their attribute values; At the same time, new dynamic features are recalculated with reference to scene categories with high accuracy and added to the feature library for iterative integrated training.
10. An intelligent classification system for complex scenes in coastal areas, characterized in that: include: A data acquisition unit configured to acquire multimodal data of the border sea study area, including hyperspectral images and high-resolution image data obtained by a variety of sensors; An index calculation unit is configured to divide the coastal research area into mutually independent partition patches, and based on the multimodal data, use the information entropy index to measure the scene complexity of the surface space within the partition patches to obtain a complexity measurement index; A geographic partitioning unit is configured to use the complexity measurement index as a basis for quantitative expression of features, and based on a spatial clustering algorithm, divide the complex scene of the border and sea research area into a plurality of relatively homogeneous regions to obtain geographic partitions; A data set construction unit is configured to classify the geographical partitions in combination with the complexity measurement index, and construct training sample data and validation set data of each geographical partition at different levels according to the levels of the geographical partitions; The model training unit is configured to use a convolutional neural network model CNN to respectively construct scene classification models for different geographical divisions, and to train and verify the scene classification models based on the training sample data and the verification set data to obtain a final intelligent classification model for complex scenes in coastal areas.
Citation Information
Patent Citations
Urban land function automatic identification method based on spatial entropy
CN116108122A
Cited By
Remote sensing monitoring method, device and equipment for grassland green returning period in high-altitude complex terrain area
CN119693805A
Urban difficult site space identification method, medium and equipment
CN120071127A
A method, medium and device for identifying urban difficult site spaces
CN120071127B
Geographic scene-based demographic data spatialization method
CN120974173A
Geospatial population data spatialization method
CN120974173B
