Hyperspectral remote sensing intelligent identification method and system based on uncertainty constraint

Abstract

The application provides a hyperspectral remote sensing intelligent identification method and system based on uncertainty constraint, and belongs to the field of optical detection and surveying and mapping. A high-resolution hyperspectral remote sensing full-link digital simulation model and an indoor physical simulation system are constructed, combined with sky-ground actual measurement data, to form a sky-based and air-based hyperspectral remote sensing ground reflectivity dataset containing data uncertainty and a ground measurement hyperspectral ground reflectivity dataset; then, cross-scale correlation mapping is carried out to construct a comprehensive dataset of sky-ground cooperation; a hyperspectral remote sensing intelligent perception network based on uncertainty constraint is established, and a comparative distillation constraint is constructed to output an identification result and comprehensive uncertainty, so as to solve the problems of unclear interference aliasing mechanism, unknown spectral identification benchmark, insufficient cross-scale representation of sky-ground spectral measurement data and low quality stability of data products of high spatial resolution hyperspectral remote sensing, improve the stability of identification results under variable conditions of optical spectral detection, and realize reliability measurement.

Description

Technical Field

[0001] This invention belongs to the fields of next-generation information technology, optical detection and surveying technology, and specifically relates to a hyperspectral remote sensing intelligent identification method, system, electronic device and non-transitory computer-readable storage medium based on uncertainty constraints. Background Technology

[0002] Hyperspectral remote sensing can acquire ground feature information that combines image and spectral data, representing an important development direction in the fields of optical detection and surveying. However, hyperspectral imaging data is subdivided in both the spectral and spatial dimensions, has weak radiative energy, and suffers from a large number of mixed pixels. The challenge in its application is the high data uncertainty and unstable reconstructed spectrum, which severely restricts the stability of the data products produced by the detection system.

[0003] To quantitatively assess the quality and reliability of remote sensing observation data, uncertainty analysis of observation data has become a recent international research hotspot. After decades of development, diverse remote sensing systems have accumulated massive amounts of data, which continue to grow. Considering the differences in performance parameters of different remote sensing systems, quantitatively assessing the quality and reliability of global and long-term survey and monitoring results, and forming comparable and comprehensively usable "big data" of Earth observation, has become a recent international research trend. In 2014, the European Union proposed the Fidelity and Uncertainty in Climate data records from Earth Observations (FIDUCEO) project, which established a quantitative traceability framework for atmospheric remote sensing data research to enhance the credibility of multi-source climate datasets. In 2017, the United States proposed the Surface Biology and Geology (SBG) ten-year research plan, which focuses on rapid global surface surveys using hyperspectral remote sensing and is tackling cutting-edge challenges in the integrated inversion of atmospheric and surface parameters based on Bayesian estimation and the assessment of their uncertainties. In 2021, the International Committee for Satellite Earth Observation proposed a framework for the processing and quality assurance of analysis-ready data products, requiring that satellite remote sensing data undergo unified processing including radiometric correction, atmospheric correction, and geometric correction, and that pixel-by-pixel quality labels be provided to support future interoperability between datasets from different times and payloads.

[0004] Regarding uncertainty analysis of ground measurement data, the European Space Agency (ESA) has launched a series of Fiducial Reference Measurements (FRM) projects in recent years. This has changed the previous approach of using synchronous ground measurements as the reference truth value for verifying the authenticity of satellite products. Instead, it is conducting independent metrological traceability and quality assessment of ground measurement data in order to obtain the uncertainty of ground measurement data and provide a ground reference benchmark for verifying the authenticity of satellite remote sensing products.

[0005] my country has also carried out a lot of work in related fields, and has initially established a remote sensing product authenticity verification network system, and has initially developed some typical ground object spectral datasets, such as the Chinese typical ground object spectral library; remote sensing image datasets and corresponding database platforms for artificial intelligence applications, such as the Natural Resources Satellite Remote Sensing Cloud Service Platform and the National Integrated Earth Observation Data Sharing Platform; however, existing databases or data platforms lack data uncertainty, making it difficult to assess the stability of data quality.

[0006] With the explosive growth of artificial intelligence technology applications across various fields, research and application exploration of intelligent remote sensing based on Earth observation "big data" and improving the stability of remote sensing application products have become common expectations in the Earth observation field. In recent years, hyperspectral remote sensing ground object recognition technology has evolved from traditional machine learning algorithms to the era of deep learning and large models, aiming to cope with complex scenarios by enhancing feature extraction and expression capabilities. In 2024, the Chinese Academy of Sciences proposed the Spectral Remote Sensing Foundation Model (SpectralGPT), which, based on the three-dimensional characteristics of hyperspectral data, established a pre-training framework based on mask modeling, attempting to improve the accuracy of ground object interpretation by capturing spatial-spectral coupling features. The Hyperspectral Intelligence Comprehension Foundation Model (HyperSIGMA) proposed by Wuhan University further constructed a large-scale hyperspectral pre-training dataset and designed a sparse sampling attention mechanism to train a general-purpose foundation model with over one billion parameters, aiming to provide a model foundation for various downstream visual tasks such as classification and recognition, object detection, change detection, and semantic segmentation under a unified framework. However, existing mainstream hyperspectral remote sensing intelligent identification models typically assume no uncertainty differences in the observation data, failing to adequately consider the instability of the observation data caused by the inherent uncertainties in remote sensing data and the spectra of ground objects under varying conditions, thus affecting the stability of the identification results. The aforementioned shortcomings of existing technologies lead to the following technical problems:

[0007] 1. Space-based and airborne hyperspectral remote sensing data exhibit "different spectra for the same object" under variable imaging conditions, and the underlying mechanisms are unclear: Hyperspectral remote sensing data possesses both high spatial resolution and rich spectral dimensions. It is affected not only by factors such as illumination conditions, atmospheric conditions, topographic relief, observation geometry, payload imaging parameters, and sensor performance differences, but also by factors such as shading, spatial heterogeneity, and coupled radiative transmission from multiple atmospheric and topographic elements. Recently, the spatial resolution of my country's hyperspectral payloads has reached the meter level. At this high spatial resolution, pixel-to-pixel differences become more pronounced. The same ground feature corresponds to different parts with different surface materials and morphological combinations in different pixels, thus exhibiting different spectral characteristics, manifesting as the "different spectra for the same object" phenomenon with unstable reflectance. Existing imaging models have not yet deeply characterized its mechanism. Furthermore, existing measured remote sensing data cannot systematically cover the aforementioned variable conditions and often lacks corresponding uncertainty labels, resulting in insufficient characterization of the spectral variation range of the same ground feature under variable conditions.

[0008] 2. Limited ground-measured spectra are insufficient to represent the true reference values ​​of ground object spectral reflectance characteristics, leading to unclear benchmarks for hyperspectral remote sensing ground object identification: Typical ground object spectral datasets mostly originate from indoor pure material sample measurements under ideal conditions, or from limited field-measured spectra obtained under varying lighting, observation geometry, and environmental conditions. Limited by sampling conditions, scene settings, and measurement costs, existing ground-measured spectral data cannot systematically cover the influence of different lighting conditions, observation angles, scene scales, ground object states, terrain undulations, shadow occlusion, and mixed pixels on the spectral reflectance characteristics of ground objects. This results in limited ground-measured spectra failing to adequately represent the spectral diversity and uncertainty of ground objects in complex and variable scenes. Furthermore, existing image datasets use category labels, typical ground object data slices, or ground object classification maps to provide prior information on ground object categories, lacking measured reference spectra corresponding to specific imaging scenes. This makes it difficult to form unified and traceable spectral criteria for ground object categories, resulting in unclear benchmarks for hyperspectral remote sensing ground object identification.

[0009] 3. Insufficient High-Reliability Cross-Scale Correlation Mapping Between Space-Based and Space-Based Hyperspectral Remote Sensing Surface Reflectance Data and Ground-Measured Hyperspectral Surface Reflectance Data: Space-based and space-based hyperspectral remote sensing surface reflectance data have spatially continuous observation capabilities, but are easily affected by imaging conditions and scale effects; ground-measured hyperspectral surface reflectance data have high spectral measurement accuracy, but usually exist in the form of point samples. The two types of data differ in spatial scale, observation conditions, ground cover conditions, and data quality. Existing correlation mapping methods, including spatial registration based on geographic location, direct matching based on spectral similarity, and spectral unmixing based on endmembers and spectral libraries, struggle to establish stable cross-scale correspondences between space-based and space-based hyperspectral data, ground-measured spectra, condition parameters, and uncertainty quantification results. This lack of a high-reliability cross-scale correlation mapping between space-air and ground limits the effective utilization of space-air-ground collaborative data in intelligent recognition models.

[0010] 4. Insufficient quantitative assessment of the recognition stability and reliability of existing hyperspectral remote sensing intelligent recognition methods: Existing deep learning-based hyperspectral remote sensing intelligent recognition methods typically use the entire scene or its local samples as input, outputting pixel-by-pixel category labels or category probabilities. Influenced by factors such as lighting conditions, observation conditions, changes in ground cover conditions, and data quality conditions expressed as data uncertainty, the statistical characteristic distribution of the input data is prone to change. However, existing hyperspectral remote sensing intelligent recognition methods rarely use uncertainty for feature extraction constraints and reliability measurement of recognition results, leading to significant fluctuations in the feature extraction process and recognition results. Therefore, they often suffer from insufficient recognition stability and difficulty in providing recognition results with reliable metrics, thus restricting the reliable application of hyperspectral remote sensing intelligent recognition in complex environments. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a hyperspectral remote sensing intelligent identification method and system based on uncertainty constraints. This method utilizes physical intelligence to improve the quality and stability of data products from hyperspectral remote sensing systems, thereby solving the following technical problems:

[0012] 1. By establishing a high spatial resolution hyperspectral remote sensing full-link digital simulation model and constructing a space-based and airborne hyperspectral remote sensing surface reflectance dataset with data uncertainty, the problem of "same object, different spectrum" phenomenon and unclear change mechanism in space-based and airborne hyperspectral remote sensing data under multivariable imaging conditions is solved.

[0013] 2. By establishing a multi-angle hyperspectral remote sensing indoor physical simulation system with continuous spatial illumination and controllable illuminance, and constructing a hyperspectral surface reflectance dataset containing data uncertainty, the problem of insufficient reference values ​​for characterizing the spectral reflectance characteristics of ground objects due to a small number of ground measurement spectra is solved, and the benchmark for hyperspectral remote sensing ground object identification is unclear.

[0014] 3. By using a cross-scale correlation mapping method based on multimodal correlation representation, a comprehensive dataset integrating space, air, and ground was constructed to address the problem of insufficient high-reliability cross-scale correlation mapping between space-based and airborne hyperspectral remote sensing surface reflectance data and ground-measured hyperspectral surface reflectance data.

[0015] 4. By constructing a hyperspectral remote sensing intelligent perception network based on uncertainty constraints, the identification results and comprehensive uncertainty are presented, the confidence range of the identification results under a given confidence level is characterized, the identification stability is improved, and the problem of insufficient quantitative evaluation of the identification stability and reliability of existing hyperspectral remote sensing intelligent identification methods is solved.

[0016] A first aspect of the present invention is to provide a hyperspectral remote sensing intelligent identification method based on uncertainty constraints, comprising:

[0017] Step S100: Establish a high spatial resolution hyperspectral remote sensing end-to-end digital simulation model and generate a space-based and air-based hyperspectral remote sensing surface reflectance dataset containing data uncertainty; wherein, the end-to-end digital simulation model is used to generate digital simulation data;

[0018] Step S200: Establish a multi-angle hyperspectral remote sensing indoor physical simulation system with continuous spatial illumination and controllable illuminance, and generate a ground measurement hyperspectral surface reflectance dataset containing data uncertainty; wherein, the multi-angle hyperspectral remote sensing indoor physical simulation system is used to acquire physical simulation data;

[0019] Step S300: Based on the space-based and airborne hyperspectral remote sensing surface reflectance dataset containing data uncertainty and the ground-based hyperspectral surface reflectance dataset containing data uncertainty, perform cross-scale correlation mapping to form a comprehensive dataset that integrates space, air, and ground.

[0020] Step S400: Based on the comprehensive dataset of the sky-ground collaboration, construct a hyperspectral remote sensing intelligent sensing network based on uncertainty constraints, and output pixel-by-pixel recognition results and comprehensive uncertainty.

[0021] Preferably, step S100 includes:

[0022] S110, Establish a high spatial resolution hyperspectral remote sensing end-to-end digital simulation model, including:

[0023] Based on the controllable target components of the triangular mesh and the real surface of the grid, and combined with the background reflectivity and the anisotropic reflectivity of the target, a high spatial resolution three-dimensional reflectivity scene of the surface is constructed.

[0024] Based on imaging geometry, the multiple scattering and shadow occlusion between target components and background within the scene are calculated to generate a scene with arbitrary surface radiance at any angle.

[0025] A high spatial resolution three-dimensional atmospheric field is established, and atmospheric radiative transfer is calculated based on the high spatial resolution three-dimensional atmospheric field to obtain remote sensing entrance pupil radiance data.

[0026] Based on the payload imaging system and parameters, spatial, spectral and radiative response simulations are performed on the entrance pupil radiance data to achieve high spatial resolution hyperspectral imaging simulation and generate digital simulation data.

[0027] S120: Acquire space-based and air-based hyperspectral remote sensing data and corresponding variable condition parameters, and combine them with the digital simulation data generated in S110 to construct space-based and air-based hyperspectral remote sensing datasets under variable conditions.

[0028] S130, the space-based and space-based hyperspectral remote sensing datasets under the aforementioned variable conditions are calibrated, reflectance inverted, quantitatively traced, and uncertainty assessed to form a space-based and space-based hyperspectral remote sensing surface reflectance dataset containing data uncertainty.

[0029] Preferably, step S200 includes:

[0030] S210, establish a multi-angle hyperspectral remote sensing indoor physical simulation system with continuous spatial illumination and controllable illuminance, including: a sunlight simulation subsystem, a skylight simulation subsystem, an observation simulation subsystem, and a ground scene simulation subsystem;

[0031] By coordinating the adjustment of irradiance intensity, illumination geometry, and observation geometry, the ground scene reflectance spectral data under different illumination conditions are collected by using the observation simulation subsystem equipped with imaging spectral detection equipment and non-imaging spectral detection equipment respectively.

[0032] When the sunlight simulation subsystem is turned on and the sky light simulation subsystem is turned off, the biaxial reflectivity factor and biaxial reflectance distribution function of the ground scene are obtained. When the sunlight simulation subsystem and the sky light simulation subsystem are turned on at the same time, the hemispherical conical reflectivity factor of the ground scene is obtained, and physical simulation data under varying conditions are acquired.

[0033] S220: Obtain ground-measured hyperspectral surface reflectance data and corresponding variable condition parameters, and combine them with the physical simulation data and corresponding variable condition parameters obtained in S210 to construct a ground-measured hyperspectral dataset under variable conditions.

[0034] S230, perform metrological traceability and uncertainty assessment on the ground-measured hyperspectral dataset to form the ground-measured hyperspectral surface reflectance dataset containing data uncertainty.

[0035] Preferably, step S300 includes:

[0036] S310, extract the space-based and space-based hyperspectral remote sensing surface reflectance data and the ground-measured hyperspectral surface reflectance data from the space-based and space-based hyperspectral remote sensing surface reflectance dataset and the ground-measured hyperspectral surface reflectance dataset, as well as the lighting conditions, observation conditions, ground cover conditions and data quality conditions expressed in terms of data uncertainty for each data.

[0037] S320, the lighting conditions, observation conditions, ground object status, and data quality conditions expressed as data uncertainty are encoded into multimodal conditional features through linear mapping, and embedded in the spectral feature encoding process to obtain the multimodal conditional spectral features corresponding to the space-based and air-based hyperspectral remote sensing surface reflectance data and the ground-measured hyperspectral surface reflectance data, respectively, forming a multimodal correlation characterization;

[0038] S330, Based on the multimodal correlation characterization, the similarity of the multimodal conditional spectral characteristics of space-based and airborne hyperspectral remote sensing surface reflectance data and ground-measured hyperspectral surface reflectance data is measured to obtain the similarity measurement results;

[0039] S340, Based on the similarity measurement results, determine the relationship between space-based and airborne hyperspectral remote sensing surface reflectance data and ground-measured hyperspectral surface reflectance data corresponding to the same type of land cover, establish a cross-scale correlation mapping between the scale of space-based and airborne hyperspectral remote sensing surface reflectance data and the scale of ground-measured hyperspectral surface reflectance data, and form a comprehensive dataset that coordinates space, air and ground.

[0040] Preferably, step S400 includes:

[0041] The S410 employs a dual-branch distillation architecture for both teachers and students.

[0042] S420: Extract ground-measured hyperspectral surface reflectance data containing data uncertainty from the integrated dataset of the sky-ground collaboration, input it into the teacher branch, and generate ground feature category sample features;

[0043] S430, extract space-based and air-based hyperspectral remote sensing surface reflectance data containing data uncertainty from the integrated dataset of the space-air-ground collaboration, input it into the student branch, and use its data uncertainty as a constraint to extract pixel-level spatial-spectral features, so that the higher the data uncertainty, the lower the attention weight of the corresponding feature, and the lower the data uncertainty, the higher the attention weight of the corresponding feature.

[0044] S440, construct the comparative distillation constraint of the teacher branch and the student branch, so that the pixel-level spatial spectral features are aligned with the features of the land cover category samples, and the aligned pixel-level spatial spectral features are obtained.

[0045] S450, based on the aligned pixel-level spatial spectral features, performs pixel-by-pixel ground object identification, while simultaneously assessing the uncertainty of the intelligent sensing network, and fusing it with the data uncertainty of space-based and air-based hyperspectral remote sensing surface reflectance data, finally outputting the pixel-by-pixel identification result and the comprehensive uncertainty.

[0046] Preferably, S440 includes:

[0047] S441, map the features of the local land cover category samples generated by the teacher branch and the pixel-level spatial spectral features extracted by the student branch to the same feature space, and perform feature normalization;

[0048] S442, for each pixel-level spatial spectral feature, the land cover category sample feature that is the same as its land cover category is determined as the positive sample feature;

[0049] S443, calculate the feature similarity between the pixel-level spatial spectral features and the positive sample features respectively;

[0050] S444, Contrastive distillation loss is constructed based on the feature similarity, and the pixel-level spatial spectral features are aligned with the corresponding land cover category sample features by increasing the feature similarity between the pixel-level spatial spectral features and the positive sample features.

[0051] Preferably, S450 includes:

[0052] S451, based on Monte Carlo random deactivation and evidence-based deep learning methods, assesses the uncertainty of intelligent sensing networks during the forward propagation process in the inference phase;

[0053] S452, Extract the data uncertainty of space-based and air-based hyperspectral remote sensing surface reflectance data from the integrated dataset of the space-air-ground collaboration;

[0054] S453, Based on the weighted fusion method, the uncertainty of the intelligent sensing network and the data uncertainty of the space-based and air-based hyperspectral remote sensing surface reflectance data are fused to obtain the comprehensive uncertainty.

[0055] A second aspect of the present invention provides a hyperspectral remote sensing intelligent identification system based on uncertainty constraints, for performing the hyperspectral remote sensing intelligent identification method based on uncertainty constraints described in the first aspect, comprising:

[0056] The digital simulation model and space-based and airborne dataset construction module are used to establish a high spatial resolution hyperspectral remote sensing end-to-end digital simulation model and form a space-based and airborne hyperspectral remote sensing surface reflectance dataset containing data uncertainty; wherein, the end-to-end digital simulation model is used to generate digital simulation data;

[0057] The physical simulation system and ground measurement dataset construction module are used to establish a multi-angle hyperspectral remote sensing indoor physical simulation system with continuous spatial illumination and controllable illuminance, and to form a ground measurement hyperspectral surface reflectance dataset containing data uncertainty; wherein, the multi-angle hyperspectral remote sensing indoor physical simulation system is used to acquire physical simulation data;

[0058] The integrated dataset construction module for the sky-ground collaboration is used to form an integrated dataset for the sky-ground collaboration based on the space-based and air-based hyperspectral remote sensing surface reflectance dataset containing data uncertainty and the ground-based measurement hyperspectral surface reflectance dataset containing data uncertainty.

[0059] The intelligent recognition module is used to construct a hyperspectral remote sensing intelligent perception network based on uncertainty constraints based on the comprehensive dataset of the sky-ground collaboration, and output pixel-by-pixel recognition results and comprehensive uncertainty based on the hyperspectral remote sensing intelligent perception network.

[0060] A third aspect of the present invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the hyperspectral remote sensing intelligent identification method based on uncertainty constraints of the first aspect.

[0061] A fourth aspect of the present invention is to provide a non-transitory computer-readable storage medium, characterized in that the non-transitory computer-readable storage medium stores computer instructions for causing the computer to execute the hyperspectral remote sensing intelligent identification method based on uncertainty constraints of the first aspect.

[0062] The method and system of the present invention have the following beneficial technical effects:

[0063] (1) An imaging model for hyperspectral remote sensing data under multivariable imaging conditions was established, the coupling mechanism of signal and interference was characterized, and for the first time, a pixel-scale uncertainty label for hyperspectral data was given, clarifying the generation mechanism of "different spectra for the same object" and improving the characterization ability of spectral variation range:

[0064] This invention proposes for the first time a full-link digital simulation model for high spatial resolution hyperspectral remote sensing, revealing the mechanism and laws of changes in space-based and airborne hyperspectral remote sensing data under multi-variable imaging conditions, and explaining the essence of the phenomenon of "different spectra for the same object". This invention constructs a high spatial resolution three-dimensional surface reflectance scene, imaging geometry model, high-resolution three-dimensional atmospheric field, and payload imaging model to characterize the variation mechanism of ground object and background radiation coupling under multivariable imaging conditions. This includes factors such as the three-dimensional structure and surface material distribution of ground objects, three-dimensional atmospheric radiation transmission, and illumination observation geometry. It achieves full-link digital simulation of high spatial resolution hyperspectral remote sensing, generating hyperspectral remote sensing digital simulation data under multivariable imaging conditions, increasing sample diversity. Furthermore, by combining space-based and airborne measured hyperspectral remote sensing data, a hyperspectral remote sensing surface reflectance dataset is formed. This dataset is calibrated, reflectance inverted, quantitatively traced, and its uncertainty is assessed, resulting in corresponding reflectance uncertainty labels. The confidence interval of reflectance at a given confidence level is characterized, enhancing the quantitative characterization ability of the spectral variation patterns of ground objects under complex and variable conditions. This provides a theoretical and data foundation for the systematic analysis of hyperspectral remote sensing data.

[0065] (2) An indoor ground object spectral measurement system was established that can achieve the same parameters as field measurements (hemispherical conic reflectance factor). A ground-based hyperspectral measurement dataset combining indoor and outdoor ground measurement data was constructed. Based on rich samples, multi-parameter characterization and uncertainty evaluation of the ground-based hyperspectral benchmark were realized, solving the problems that a small number of ground measurement spectra are difficult to characterize the true reference value of ground object reflectance characteristics and the hyperspectral remote sensing ground object identification benchmark is unclear.

[0066] This invention proposes combining indoor physical simulation of hyperspectral remote sensing with ground-measured hyperspectral data to enrich the spectral samples of ground objects and collaboratively construct a spectral reference true value that can serve as a benchmark for ground object identification. This invention acquires physical simulation data under controlled conditions of different lighting, observation geometry, atmospheric environment, spatial scale, and ground object heterogeneity through an indoor physical simulation system. Combined with ground-measured hyperspectral data, the invention performs metrological traceability and uncertainty assessment on the data acquisition, radiometric calibration, and reflectance calculation processes. This constructs a ground-measured hyperspectral surface reflectance dataset containing measurement conditions, ground object attributes, reflectance data, and uncertainty information, forming a reference true value for the spectral reflectance characteristics of ground objects under different conditions and their reasonable variation ranges. Through the complementarity of indoor controllable simulation data and ground-based real-scene data, the invention enhances the systematic characterization capability of ground object spectral characteristics under complex lighting, observation geometry, scale effects, and changes in ground object state, providing a unified, traceable reference benchmark with quantifiable uncertainty information for hyperspectral remote sensing ground object identification.

[0067] (3) A cross-scale correlation mapping method based on multimodal correlation representation was proposed, forming a comprehensive dataset of sky-ground collaboration, providing a new approach for reliable correlation mapping of sky-ground data, and solving the problem of highly reliable cross-scale correlation mapping between hyperspectral data:

[0068] This invention proposes a cross-scale correlation mapping method based on multimodal correlation representation, forming a comprehensive dataset for sky-ground collaboration and providing a new approach for reliable correlation mapping of sky-ground data. Based on multimodal correlation representation of illumination conditions, observation conditions, ground cover status, and data quality conditions expressed as data uncertainty, this invention performs similarity measurement to achieve cross-scale correlation mapping between space-based and airborne hyperspectral remote sensing surface reflectance data and ground-measured hyperspectral surface reflectance data. This forms a comprehensive dataset for sky-ground collaboration, reducing correlation bias caused by scale differences, changes in observation conditions, and data quality differences, improving the stability and reliability of the correspondence between sky-ground data, and providing a new mapping paradigm for the collaborative utilization of sky-ground hyperspectral data.

[0069] (4) A hyperspectral remote sensing intelligent sensing network based on uncertainty constraints is proposed, which realizes intelligent interpretation and recognition of physical constraints and solves the problem of insufficient quantitative evaluation of recognition stability and reliability of existing hyperspectral remote sensing intelligent recognition methods:

[0070] This invention proposes a hyperspectral remote sensing intelligent sensing network based on uncertainty constraints, which improves the intelligent recognition stability of hyperspectral data products and achieves, for the first time, a reliability measurement of recognition results based on uncertainty constraints. The invention employs a teacher-student two-branch distillation architecture. Ground-based hyperspectral surface reflectance data containing data uncertainty is used to generate land cover category samples. Space-based and airborne hyperspectral remote sensing surface reflectance data containing data uncertainty are input into the student branch. Through a negative feedback attention mechanism based on uncertainty constraints, the higher the data uncertainty, the lower the attention weight of the corresponding feature; conversely, the lower the data uncertainty, the higher the attention weight of the corresponding feature. Simultaneously, by combining contrastive distillation constraints, pixel-level spatial-spectral features are aligned with land cover category sample features. Finally, by combining data uncertainty and intelligent sensing network uncertainty, pixel-by-pixel recognition results and comprehensive uncertainty are output, describing the confidence interval of pixel-by-pixel category probability at a given confidence level. This improves the stability of recognition results and provides a quantitative reliability assessment, enhancing the quantitative characterization ability of recognition result reliability. Attached Figure Description

[0071] Figure 1 The diagram shown is a flowchart of a hyperspectral remote sensing intelligent identification method based on uncertainty constraints according to a preferred embodiment of the present invention.

[0072] Figure 2The diagram shown is a flowchart of the acquisition process for space-based and airborne hyperspectral remote sensing surface reflectance datasets, including data uncertainty, according to a preferred embodiment of the present invention.

[0073] Figure 3 The diagram shown is a flowchart of obtaining a hyperspectral surface reflectance dataset containing data uncertainty based on a preferred embodiment of the present invention.

[0074] Figure 4 The diagram shown is a flowchart of the construction process of a comprehensive dataset for sky-ground collaboration according to a preferred embodiment of the present invention.

[0075] Figure 5 The diagram shows a flowchart of a computation method for a hyperspectral remote sensing intelligent sensing network based on uncertainty constraints according to a preferred embodiment of the present invention. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to specific embodiments.

[0077] This embodiment uses hyperspectral remote sensing data for land cover identification as an example to illustrate a hyperspectral remote sensing intelligent identification method based on uncertainty constraints. The overall framework of this method is as follows: Figure 1 As shown, the process includes: constructing space-based and airborne hyperspectral remote sensing surface reflectance datasets containing data uncertainty; constructing ground-measured hyperspectral surface reflectance datasets containing data uncertainty; performing cross-scale correlation mapping between space-based and airborne hyperspectral remote sensing surface reflectance data and ground-measured hyperspectral surface reflectance data to form a comprehensive dataset that integrates space, air, and ground; constructing a hyperspectral remote sensing intelligent sensing network based on uncertainty constraints based on the comprehensive dataset, and outputting pixel-by-pixel recognition results and comprehensive uncertainty.

[0078] Step S100: Construct a space-based and air-based hyperspectral remote sensing surface reflectance dataset containing data uncertainty.

[0079] First, establish such as Figure 2 The high spatial resolution hyperspectral remote sensing full-link digital simulation model is shown. Specifically, based on the triangular mesh controllable target components and the grid real surface, combined with the background reflectivity and the target anisotropic reflectivity, a high spatial resolution three-dimensional reflectivity scene of the surface is constructed. The triangular mesh controllable target components can be used to describe the geometric structure, surface orientation and component material composition of the ground target, and the grid real surface can be used to describe the surface background, spatial texture and terrain undulation.

[0080] Subsequently, based on the imaging geometry, multiple scattering and shadow occlusion between the target components and the background within the scene are calculated to generate a scene of surface radiance at any angle. Through the above processing, the spectral characteristics of the same ground feature can be simulated under different angles, different component combinations, and different background conditions.

[0081] Furthermore, a high-resolution three-dimensional atmospheric field is established, and atmospheric radiative transfer is calculated based on the surface outgoing radiance scene to obtain remote sensing entrance pupil radiance data.

[0082] Subsequently, based on the payload imaging system and parameters, spatial, spectral, and radiative dimension response simulations were performed on the remote sensing entrance pupil radiance data to obtain digital simulation data and its corresponding variable condition parameters.

[0083] Simultaneously, space-based and airborne hyperspectral remote sensing data and their corresponding variable condition parameters are acquired; the variable condition parameters include illumination conditions, observation conditions, and ground object status; the digital simulation data, together with the space-based and airborne measured data and the corresponding variable condition parameters, constitute a space-based and airborne hyperspectral remote sensing dataset under variable conditions.

[0084] Finally, the space-based and airborne hyperspectral remote sensing datasets under the aforementioned variable conditions are calibrated, reflectance inverted, quantitatively traced, and uncertainty assessed. Specifically, the hyperspectral remote sensing data can be calibrated for radiometric and spectral properties and inverted for reflectance to obtain surface reflectance data. Then, uncertainty tracing and propagation analysis are performed on the surface reflectance data to obtain the data uncertainty corresponding to the surface reflectance data.

[0085] This results in a dataset of space-based and airborne hyperspectral remote sensing surface reflectance containing data uncertainty.

[0086] Step S200: Construct a dataset of ground-measured hyperspectral surface reflectance containing data uncertainty.

[0087] Establish a multi-angle hyperspectral remote sensing indoor physical simulation system with continuous spatial illumination and controllable illuminance. This system includes a solar light simulation subsystem, a sky light simulation subsystem, an observation simulation subsystem, and a ground scene simulation subsystem.

[0088] like Figure 3 As shown, the multi-angle hyperspectral remote sensing indoor physical simulation system with continuous spatial illumination and controllable illuminance includes a solar light simulation subsystem, a sky light simulation subsystem, an observation simulation subsystem, and a ground scene simulation subsystem.

[0089] The process of acquiring a ground-measured hyperspectral surface reflectance dataset that includes data uncertainty consists of four stages: ground-measured hyperspectral data acquisition, indoor physical simulation hyperspectral data acquisition, construction of a ground-measured hyperspectral dataset under varying conditions, and metrological traceability and uncertainty assessment.

[0090] In the ground-based hyperspectral data acquisition phase, the site was first selected, and the observation angle was adjusted according to the preset measurement conditions. Then, the condition of the ground features and the environment were recorded, and the ground-based spectral measurement equipment was calibrated. After equipment calibration, ground-based spectral measurements were performed on the selected ground features to obtain the corresponding ground-based hyperspectral data.

[0091] During the indoor physical simulation hyperspectral data acquisition phase, the indoor physical simulation system is activated. Based on the solar illumination angle and weather conditions to be simulated, the sky light simulation subsystem and the solar light simulation subsystem are activated separately, and the corresponding illumination modes are selected. By adjusting the light source radiant power and the solar illumination angle, preset solar-sky light illumination conditions are constructed. Simultaneously, the observation simulation subsystem is activated, and the observation angle and observation distance are adjusted. The ground scene simulation subsystem is used to adjust the angle and select the ground scene, setting the corresponding ground observation states. After completing the lighting conditions, observation conditions, and ground scene settings, an imager is selected to perform imaging measurements on the ground scene, acquiring the hemispherical conic reflectance factor physical simulation hyperspectral data.

[0092] The measured hyperspectral data and the physical simulation hyperspectral data are collected and organized to construct a ground measurement hyperspectral dataset under various conditions, including different lighting conditions, observation conditions, ground cover conditions and environmental conditions.

[0093] Further metrological traceability and uncertainty assessment were performed on the ground-measured hyperspectral dataset under the aforementioned variable conditions, ultimately forming a ground-measured hyperspectral surface reflectance dataset containing data uncertainty.

[0094] Step S300: Construct a comprehensive dataset for sky-ground collaboration.

[0095] After acquiring space-based and airborne hyperspectral remote sensing surface reflectance datasets containing data uncertainty, as well as ground-measured hyperspectral surface reflectance datasets containing data uncertainty, according to Figure 4 The process shown demonstrates cross-scale correlation mapping between the two types of data;

[0096] Specifically, it acquires ground-based hyperspectral surface reflectance data and space-based and airborne hyperspectral remote sensing surface reflectance data, as well as the lighting conditions, observation conditions, ground cover conditions, and data quality conditions expressed in terms of data uncertainty for each data set.

[0097] The lighting conditions, observation conditions, ground object conditions, and data quality conditions expressed as data uncertainty are encoded into multimodal conditional features through a multilayer perceptron. These multimodal conditional features are then embedded into the spectral feature encoding process of the corresponding data. This ensures that the spectral feature encoding results simultaneously contain ground object spectral information and their corresponding conditional information. This yields multimodal conditional spectral features corresponding to ground-based hyperspectral surface reflectance data and space-based and airborne hyperspectral remote sensing surface reflectance data, forming a multimodal correlation representation.

[0098] Based on the aforementioned multimodal association characterization, the multimodal conditional spectral features of ground-based hyperspectral surface reflectance data and space-based and airborne hyperspectral remote sensing surface reflectance data are measured for similarity using attention similarity to obtain similarity measurement results.

[0099] Based on the similarity measurement results, the relationship between space-based and airborne hyperspectral remote sensing surface reflectance data and ground-measured hyperspectral surface reflectance data for corresponding land features is determined. For data with high similarity, the correspondence is strengthened; for data with low similarity, the correspondence is weakened, thereby establishing a cross-scale correlation mapping between the scale of space-based and airborne hyperspectral remote sensing surface reflectance data and the scale of ground-measured hyperspectral surface reflectance data.

[0100] Finally, the space-based and airborne hyperspectral remote sensing surface reflectance data and their data uncertainty after cross-scale correlation mapping, the ground-measured hyperspectral surface reflectance data and their data uncertainty, as well as the land cover category labels and condition parameters corresponding to each data, are jointly used to construct a space-air-ground collaborative sample unit, thereby forming a comprehensive space-air-ground collaborative dataset.

[0101] Step S400: Construct a hyperspectral remote sensing intelligent sensing network based on uncertainty constraints.

[0102] Using the integrated dataset of sky-ground collaboration formed in step S300 as the training set, a hyperspectral remote sensing intelligent perception network based on uncertainty constraints is constructed. A teacher-student dual-branch distillation architecture is adopted, including teacher branch, student branch, comparative distillation constraint and integrated uncertainty calculation process.

[0103] In the teacher branch, ground-measured hyperspectral surface reflectance data with data uncertainty is used as input. Specifically, the ground-measured hyperspectral surface reflectance data are first grouped according to land cover categories, and multiple ground-measured hyperspectral surface reflectance data within the same category are encoded using the Transformer method for spectral feature encoding. Then, aggregation weights are set based on the data uncertainty corresponding to the ground-measured hyperspectral surface reflectance data, giving higher weights to samples with lower data uncertainty and lower weights to samples with higher data uncertainty. Finally, the spectral features within the same category are weighted and aggregated to generate the corresponding land cover category sample features.

[0104] In the student branch, space-based and airborne hyperspectral remote sensing surface reflectance data containing data uncertainty are used as input to extract pixel-level spatial-spectral features. Specifically, a local spatial neighborhood is first constructed around the pixel to be identified, and the spectral response within this local spatial neighborhood is combined to form a spatial-spectral feature unit. Subsequently, the spatial-spectral feature unit is input into a spatial-spectral joint encoder to jointly encode the space- and spectral dimensions of the space-based and airborne hyperspectral remote sensing surface reflectance data to obtain the initial spatial-spectral features.

[0105] Furthermore, the data uncertainty corresponding to space-based and airborne hyperspectral remote sensing surface reflectance data is introduced into the feature extraction process of the student branch, forming a negative feedback attention constraint. Specifically, let the first... The query feature corresponding to each spatial-spectral feature unit is , No. The bond features and value features corresponding to each spatial-spectral feature unit are respectively and Its data uncertainty is Then the negative feedback attention weights and the corresponding feature aggregation results are expressed as follows:

[0106]

[0107] in, This indicates the number of spatial-spectral feature units involved in attention calculation. This represents the index of the query feature unit. This indicates the index of the currently viewed feature unit. This represents the summation index that traverses all spatial-spectral feature units during the normalization process. Represents the attention feature dimension. This represents the negative feedback coefficient for data uncertainty. Indicates the first The data uncertainty corresponding to each feature unit This represents the attention weight after introducing data uncertainty constraints. This represents the result of feature aggregation. The higher the data uncertainty, the stronger the negative feedback term. The larger the value, the lower the attention weight of the corresponding feature; the lower the data uncertainty, the weaker the negative feedback suppression it receives, thereby reducing the impact of high uncertainty features on the pixel-level spatial spectral feature extraction results.

[0108] Between the teacher and student branches, a comparative distillation constraint is constructed based on the land cover category sample features generated by the teacher branch and the pixel-level spatial spectral features extracted by the student branch. Specifically, the pixel-level spatial spectral features and the features of each land cover category sample are projected onto the same feature space and normalized. For the ... For each pixel sample, features of samples from the same land cover category as the land cover category are used as positive sample features, and the following contrastive distillation loss is constructed:

[0109]

[0110] in, This indicates the number of pixel samples participating in the training. Indicates the number of land cover categories. Represents the land cover category index. Indicates the first Pixel-level spatial spectral features of individual pixel samples after projection and normalization Indicates the first Features of land cover categories after projection and normalization. Indicates the relationship with the first Individual pixel true category Corresponding, positive sample features of land cover categories after feature mapping and normalization. Indicates the temperature coefficient. Indicates distillation loss, The feature similarity function can be represented, and cosine similarity can be used in particular. Represented by natural constant The natural logarithm with base 0. By minimizing the contrastive distillation loss, the feature similarity between pixel-level spatial spectral features and corresponding land cover category sample features is increased, so that the pixel-level spatial spectral features extracted by the student branch are aligned with the corresponding land cover category sample features. Thus, the ground-based hyperspectral surface reflectance data information is used to constrain the student branch's recognition process of space-based and airborne hyperspectral remote sensing surface reflectance data.

[0111] Simultaneously, during the network training phase, a comprehensive training loss is constructed based on the land cover category sample features generated by the teacher branch, the pixel-level spatial spectral features extracted by the student branch, and the evidence deep learning framework. This comprehensive training loss includes a contrastive distillation loss to constrain the alignment of pixel-level spatial spectral features with land cover category sample features, and an evidence learning loss to constrain category evidence learning. The hyperspectral remote sensing intelligent perception network is then trained using this comprehensive training loss.

[0112] Furthermore, during the network inference phase, the uncertainty of the intelligent perception network is assessed during forward propagation based on Monte Carlo random deactivation and evidence deep learning methods. Specifically, Monte Carlo random deactivation is enabled and multiple forward samples are performed. For each forward sample, the non-negative evidence quantity corresponding to each object category is output through the evidence deep learning framework, and Dirichlet distribution parameters are constructed based on the non-negative evidence quantity to obtain the category probability and evidence uncertainty corresponding to that forward sample. Subsequently, the mean of the evidence uncertainty obtained from multiple forward samples is calculated to form the evidence uncertainty statistics; simultaneously, based on the differences between the category probability distributions obtained from multiple forward samples, the posterior prediction divergence is formed using a mutual information metric. Furthermore, based on the mean of the evidence uncertainty and the posterior prediction divergence, the uncertainty of the intelligent sensing network is formed using the weighted sum-of-squares method. Finally, the data uncertainty of the space-based and air-based hyperspectral remote sensing surface reflectance data and the uncertainty of the intelligent sensing network are scaled and normalized, and the comprehensive uncertainty, which comprehensively represents the data uncertainty and the uncertainty of the intelligent sensing network, is calculated using a weighted linear fusion method. Simultaneously, the intelligent sensing network calculates and outputs pixel-by-pixel land cover category recognition results based on the evidence deep learning framework.

[0113] Finally, the hyperspectral remote sensing intelligent sensing network based on uncertainty constraints outputs pixel-by-pixel recognition results and their comprehensive uncertainty, forming a result as follows: Figure 5 The hyperspectral remote sensing intelligent perception network framework shown is based on uncertainty constraints, which enables the recognition results to include both category determination information and reliability measurement information.

[0114] The embodiments described above are only used to explain the preferred implementation of the present invention and are not intended to limit the scope of protection of the present invention. Any substitutions, combinations or modifications that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be covered within the scope of protection of the present invention.

Claims

1. A hyperspectral remote sensing intelligent identification method based on uncertainty constraints, characterized in that, include: Step S100: Establish a high spatial resolution hyperspectral remote sensing end-to-end digital simulation model and generate a space-based and air-based hyperspectral remote sensing surface reflectance dataset containing data uncertainty; wherein, the end-to-end digital simulation model is used to generate digital simulation data; Step S200: Establish a multi-angle hyperspectral remote sensing indoor physical simulation system with continuous spatial illumination and controllable illuminance, and generate a ground measurement hyperspectral surface reflectance dataset containing data uncertainty; wherein, the multi-angle hyperspectral remote sensing indoor physical simulation system is used to acquire physical simulation data; Step S300: Based on the space-based and airborne hyperspectral remote sensing surface reflectance dataset containing data uncertainty and the ground-based hyperspectral surface reflectance dataset containing data uncertainty, perform cross-scale correlation mapping to form a comprehensive dataset that integrates space, air, and ground. Step S400: Based on the comprehensive dataset of the sky-ground collaboration, construct a hyperspectral remote sensing intelligent sensing network based on uncertainty constraints, and output pixel-by-pixel recognition results and comprehensive uncertainty.

2. The hyperspectral remote sensing intelligent identification method based on uncertainty constraints according to claim 1, characterized in that: Step S100 includes: S110, Establish a high spatial resolution hyperspectral remote sensing end-to-end digital simulation model, including: Based on the controllable target components of the triangular mesh and the real surface of the grid, and combined with the background reflectivity and the anisotropic reflectivity of the target, a high spatial resolution three-dimensional reflectivity scene of the surface is constructed. Based on imaging geometry, the multiple scattering and shadow occlusion between target components and background within the scene are calculated to generate a scene with arbitrary surface radiance at any angle. A high spatial resolution three-dimensional atmospheric field is established, and atmospheric radiative transfer is calculated based on the high spatial resolution three-dimensional atmospheric field to obtain remote sensing entrance pupil radiance data. Based on the payload imaging system and parameters, spatial, spectral and radiative response simulations are performed on the entrance pupil radiance data to achieve high spatial resolution hyperspectral imaging simulation and generate digital simulation data. S120: Acquire space-based and air-based hyperspectral remote sensing data and corresponding variable condition parameters, and combine them with the digital simulation data generated in S110 to construct space-based and air-based hyperspectral remote sensing datasets under variable conditions. S130, the space-based and space-based hyperspectral remote sensing datasets under the aforementioned variable conditions are calibrated, reflectance inverted, quantitatively traced, and uncertainty assessed to form a space-based and space-based hyperspectral remote sensing surface reflectance dataset containing data uncertainty.

3. The hyperspectral remote sensing intelligent identification method based on uncertainty constraints according to claim 1, characterized in that: Step S200 includes: S210, establish a multi-angle hyperspectral remote sensing indoor physical simulation system with continuous spatial illumination and controllable illuminance, including: a sunlight simulation subsystem, a skylight simulation subsystem, an observation simulation subsystem, and a ground scene simulation subsystem; By coordinating the adjustment of irradiance intensity, illumination geometry, and observation geometry, the observation simulation subsystem is equipped with imaging spectral detection equipment and non-imaging spectral detection equipment respectively to collect ground scene reflectance spectral data under varying conditions; When the sunlight simulation subsystem is turned on and the sky light simulation subsystem is turned off, the biaxial reflectivity factor and biaxial reflectance distribution function of the ground scene are obtained. When the sunlight simulation subsystem and the sky light simulation subsystem are turned on at the same time, the hemispherical conical reflectivity factor of the ground scene is obtained, and physical simulation data under varying conditions are acquired. S220: Obtain ground-measured hyperspectral surface reflectance data and corresponding variable condition parameters, and combine them with the physical simulation data and corresponding variable condition parameters obtained in S210 to construct a ground-measured hyperspectral dataset under variable conditions. S230, perform metrological traceability and uncertainty assessment on the ground-measured hyperspectral dataset to form a ground-measured hyperspectral surface reflectance dataset containing data uncertainty.

4. The hyperspectral remote sensing intelligent identification method based on uncertainty constraints according to claim 1, characterized in that: Step S300 includes: S310, extract the space-based and space-based hyperspectral remote sensing surface reflectance data and the ground-measured hyperspectral surface reflectance data from the space-based and space-based hyperspectral remote sensing surface reflectance dataset and the ground-measured hyperspectral surface reflectance dataset, as well as the lighting conditions, observation conditions, ground cover conditions and data quality conditions expressed in terms of data uncertainty for each data. S320, the lighting conditions, observation conditions, ground object status, and data quality conditions expressed as data uncertainty are encoded into multimodal conditional features through linear mapping, and embedded in the spectral feature encoding process to obtain the multimodal conditional spectral features corresponding to the space-based and air-based hyperspectral remote sensing surface reflectance data and the ground-measured hyperspectral surface reflectance data, respectively, forming a multimodal correlation characterization; S330, Based on the multimodal correlation characterization, the similarity of the multimodal conditional spectral characteristics of space-based and airborne hyperspectral remote sensing surface reflectance data and ground-measured hyperspectral surface reflectance data is measured to obtain the similarity measurement results; S340, Based on the similarity measurement results, determine the relationship between space-based and airborne hyperspectral remote sensing surface reflectance data and ground-measured hyperspectral surface reflectance data corresponding to the same type of land cover, establish a cross-scale correlation mapping between the scale of space-based and airborne hyperspectral remote sensing surface reflectance data and the scale of ground-measured hyperspectral surface reflectance data, and form a comprehensive dataset that coordinates space, air and ground.

5. The hyperspectral remote sensing intelligent identification method based on uncertainty constraints according to claim 1, characterized in that: Step S400 includes: The S410 employs a dual-branch distillation architecture for both teachers and students. S420: Extract ground-measured hyperspectral surface reflectance data containing data uncertainty from the integrated dataset of the sky-ground collaboration, input it into the teacher branch, and generate ground feature category sample features; S430, extract space-based and air-based hyperspectral remote sensing surface reflectance data containing data uncertainty from the integrated dataset of the space-air-ground collaboration, input it into the student branch, and use its data uncertainty as a constraint to extract pixel-level spatial-spectral features, so that the higher the data uncertainty, the lower the attention weight of the corresponding feature, and the lower the data uncertainty, the higher the attention weight of the corresponding feature. S440, construct the comparative distillation constraint of the teacher branch and the student branch, so that the pixel-level spatial spectral features are aligned with the features of the land cover category samples, and the aligned pixel-level spatial spectral features are obtained. S450, based on the aligned pixel-level spatial spectral features, performs pixel-by-pixel ground object identification, while simultaneously assessing the uncertainty of the intelligent sensing network, and fusing it with the data uncertainty of space-based and air-based hyperspectral remote sensing surface reflectance data, finally outputting the pixel-by-pixel identification result and the comprehensive uncertainty.

6. The hyperspectral remote sensing intelligent identification method based on uncertainty constraints according to claim 5, characterized in that: The S440 includes: S441, map the features of the local land cover category samples generated by the teacher branch and the pixel-level spatial spectral features extracted by the student branch to the same feature space, and perform feature normalization; S442, for each pixel-level spatial spectral feature, the land cover category sample feature that is the same as its land cover category is determined as the positive sample feature; S443, calculate the feature similarity between the pixel-level spatial spectral features and the positive sample features respectively; S444, Contrastive distillation loss is constructed based on the feature similarity, and the pixel-level spatial spectral features are aligned with the corresponding land cover category sample features by increasing the feature similarity between the pixel-level spatial spectral features and the positive sample features.

7. The hyperspectral remote sensing intelligent identification method based on uncertainty constraints according to claim 5, characterized in that: The S450 includes: S451, based on Monte Carlo random deactivation and evidence-based deep learning methods, assesses the uncertainty of intelligent sensing networks during the forward propagation process in the inference phase; S452, Extract the data uncertainty of space-based and air-based hyperspectral remote sensing surface reflectance data from the integrated dataset of the space-air-ground collaboration; S453, Based on the weighted fusion method, the uncertainty of the intelligent sensing network and the data uncertainty of the space-based and air-based hyperspectral remote sensing surface reflectance data are fused to obtain the comprehensive uncertainty.

8. A hyperspectral remote sensing intelligent identification system based on uncertainty constraints, used to execute the hyperspectral remote sensing intelligent identification method based on uncertainty constraints as described in any one of claims 1 to 7, characterized in that, include: The digital simulation model and space-based and airborne dataset construction module are used to establish a high spatial resolution hyperspectral remote sensing end-to-end digital simulation model and form a space-based and airborne hyperspectral remote sensing surface reflectance dataset containing data uncertainty; wherein, the end-to-end digital simulation model is used to generate digital simulation data; The physical simulation system and ground measurement dataset construction module are used to establish a multi-angle hyperspectral remote sensing indoor physical simulation system with continuous spatial illumination and controllable illuminance, and to form a ground measurement hyperspectral surface reflectance dataset containing data uncertainty; wherein, the multi-angle hyperspectral remote sensing indoor physical simulation system is used to acquire physical simulation data; The integrated dataset construction module for the sky-ground collaboration is used to form an integrated dataset for the sky-ground collaboration based on the space-based and air-based hyperspectral remote sensing surface reflectance dataset containing data uncertainty and the ground-based measurement hyperspectral surface reflectance dataset containing data uncertainty. The intelligent recognition module is used to construct a hyperspectral remote sensing intelligent perception network based on uncertainty constraints based on the comprehensive dataset of the sky-ground collaboration, and output pixel-by-pixel recognition results and comprehensive uncertainty based on the hyperspectral remote sensing intelligent perception network.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the hyperspectral remote sensing intelligent identification method based on uncertainty constraints as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1-7.

Images