An ecological environment monitoring method based on multi-source remote sensing collaborative interpretation

Abstract

The application relates to the technical field of image data processing and ecological environment remote sensing monitoring, and discloses an ecological environment monitoring method based on multi-source remote sensing collaborative interpretation, which has the technical scheme as follows: multi-source remote sensing data sets and basic mask layer sets are obtained through quality screening rules; a sensor effectiveness prior graph is constructed through feature extraction, and a feature vector and its consistency mark are obtained; a classification decision graph and its corresponding confidence are obtained through a double-threshold mode judgment and feature screening; and an environmental parameter set is obtained through a CASA model and a quantitative inversion model. Based on the multi-source remote sensing data sets and the basic mask layer sets, the feature vector is constructed, the classification decision is generated, and parameter inversion calculation is carried out, so that multi-source remote sensing collaborative interpretation of multi-source remote sensing data and comprehensive inversion of ecological environment parameters are realized, and the precision and operation efficiency of ecological environment monitoring are improved.

Description

Technical Field

[0001] This invention relates to the fields of image data processing and ecological environment remote sensing monitoring technology, specifically to an ecological environment monitoring method based on multi-source remote sensing collaborative interpretation. Background Technology

[0002] With the increasing demand for refined and routine ecological and environmental protection, satellite remote sensing technology has become a core support for dynamic monitoring of regional ecological and environmental conditions. Multi-source remote sensing data, including optical, synthetic aperture radar, and thermal infrared data, leverages the complementary advantages of spectral, structural, and thermal characteristics to be widely used in the collaborative monitoring of multiple elements such as vegetation growth, water quality, and urban thermal environment. However, current multi-source remote sensing ecological monitoring often employs a fixed-weight fusion framework, which lacks sufficient consideration of the spatial validity of data sources and struggles to adapt to data quality fluctuations under complex scenarios such as cloud and rain cover and topographic relief. Furthermore, the physical coupling relationships between cross-modal features are not fully utilized for quality verification, making it difficult to effectively identify hidden data deviations. Simultaneously, the general pursuit of full data fusion lacks dynamic filtering of information redundancy and negative contributions, easily leading to increased computational complexity and limited fusion accuracy improvement. It also makes it difficult to effectively distinguish between real ecological mutation signals and data noise. Therefore, existing technologies have shortcomings. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention aims to provide an ecological environment monitoring method based on multi-source remote sensing collaborative interpretation. The method processes raw image data through quality screening rules, extracts features from multi-source remote sensing datasets, and constructs a priori sensor validity map. Based on feature vectors and their consistency markers, a classification decision map and its corresponding confidence level are obtained through dual-threshold mode judgment and feature screening. Based on the classification decision map and its corresponding confidence level, an environmental parameter set is obtained through a CASA model and a quantitative inversion model. This achieves dynamic fusion of multi-source remote sensing data and comprehensive inversion of ecological environment parameters. It solves the technical problems of single remote sensing data sources being constrained by cloud and rain weather, lacking cross-modal mutual verification, the assumption that more data is always better in existing multi-source fusion methods, ignoring redundancy and negative contributions, and the easy misjudgment of real ecological change signals as noise. This improves the continuity and robustness of ecological environment monitoring, the accuracy and efficiency of multi-source fusion, the accuracy of ecological change signal identification, and the operational applicability of monitoring results.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides an ecological environment monitoring system based on multi-source remote sensing collaborative interpretation, comprising:

[0006] Acquire raw image data and obtain a multi-source remote sensing dataset and a basic mask layer set by using quality screening rules;

[0007] Based on the multi-source remote sensing dataset, feature vectors and their consistency labels are obtained by feature extraction and construction of a sensor effectiveness prior map.

[0008] Based on the feature vector and its consistency marker, a classification decision map and its corresponding confidence level are obtained through dual threshold pattern determination and feature filtering.

[0009] Based on the classification decision map and its corresponding confidence level, the environmental parameter set is obtained through the CASA model and the quantitative inversion model.

[0010] Furthermore, obtaining the multi-source remote sensing dataset and the basic mask layer set through quality screening rules includes:

[0011] Based on the original image data, a multi-source registered image set is obtained through intensity cross-correlation registration algorithm and multi-view processing;

[0012] Based on the multi-source registered image set, the basic data layer is obtained through radiometric calibration conversion and atmospheric correction of the radiative transfer model.

[0013] Based on the aforementioned basic data layer, a multi-source remote sensing dataset and a basic mask layer set are obtained by clipping the monitoring area boundaries and classifying ground features using prior knowledge.

[0014] Furthermore, the step of obtaining feature vectors and their consistency labels based on the multi-source remote sensing dataset by feature extraction and constructing a sensor effectiveness prior map includes:

[0015] Based on the multi-source remote sensing dataset, an initial feature layer is obtained through normalized difference operation, gray-level co-occurrence matrix texture analysis, and split window algorithm;

[0016] Based on the multi-source remote sensing dataset and digital elevation model, thermal infrared effectiveness score is obtained through attenuation calculation, local incident angle geometric discrimination, and cloud cover and emissivity correction.

[0017] Based on the initial feature layer and thermal infrared effectiveness score, the first resolution surface temperature is obtained by establishing a second-order polynomial regression relationship and performing inverse distance weighted interpolation on the regression residuals.

[0018] Based on the first resolution surface temperature and the initial feature layer, a feature vector is obtained through synchronous window standardization, and a consistency label is obtained according to the coupling relationship.

[0019] Furthermore, based on the multi-source remote sensing dataset, the initial feature layer is obtained through normalized difference operation, gray-level co-occurrence matrix texture analysis, and split-window algorithm, including:

[0020] Based on the multi-source remote sensing dataset, the first resolution vegetation index, the first resolution water index, and the land use cover classification map are obtained by normalized difference operation and random forest classifier.

[0021] Based on the multi-source remote sensing dataset, the first resolution structural feature layer is obtained by extracting information from dual polarization channels and analyzing gray-level co-occurrence matrix texture.

[0022] Based on the multi-source remote sensing dataset, the second-resolution land surface temperature and urban heat island intensity are obtained through the split-window algorithm and temperature difference calculation.

[0023] Furthermore, based on the multi-source remote sensing dataset and digital elevation model, the thermal infrared effectiveness score is obtained through attenuation calculation, local incident angle geometric discrimination, and cloud cover and emissivity correction, including:

[0024] The local incident angle and the radar system incident angle are calculated based on the digital elevation model to obtain the radar shadow binary identifier;

[0025] Based on the aforementioned binary radar shadow identifier, the overlapping intensity is obtained by calculating the ratio of the incident angle difference to the incident angle of the radar system and limiting it to the interval between zero and one.

[0026] Based on the mean and variance estimates of radar strength data or the equivalent number of views read from metadata, the speckle noise coefficient is obtained by calculating the reciprocal of the square root of the equivalent number of views;

[0027] Based on the overlay strength and speckle noise coefficient, the thermal infrared effectiveness score is obtained by directly setting the shadow area to zero and multiplying the overlay strength and speckle noise coefficient by subtracting the corresponding coefficient from each other.

[0028] Further, the step of obtaining the first resolution surface temperature based on the initial feature layer and the thermal infrared effectiveness score by establishing a second-order polynomial regression relationship and performing inverse distance weighted interpolation on the regression residuals includes:

[0029] Based on the initial feature layer, the regression coefficients and the regression residual values ​​of each coarse pixel are obtained through second-order polynomial fitting and regression residual calculation.

[0030] Based on the regression coefficients, the preliminary first-resolution surface temperature is calculated pixel-by-pixel using polynomials.

[0031] Based on the regression residual values ​​of each coarse pixel, the thermal infrared effectiveness score, and the preliminary first-resolution land surface temperature, areas with a preset number of effective pixels are selected, and the regression residual values ​​are transferred through inverse distance weighted interpolation to obtain the first-resolution land surface temperature.

[0032] Further, the step of obtaining a feature vector through synchronous window normalization based on the first resolution surface temperature and the initial feature layer, and obtaining a consistency label according to the coupling relationship, includes:

[0033] Vegetation cover is calculated based on the first-resolution vegetation index, and effective object masks are obtained by filtering pixels in combination with the land use cover classification map.

[0034] Based on the effective object mask and the initial feature layer, a consistency index is obtained, and pixels with a consistency index less than a preset consistency index threshold are physically suspected and the suspected data source is determined.

[0035] Based on the physical suspicious markers and suspected data sources, the validity scores of the corresponding data sources are checked back to obtain the feature status and weight decay factor.

[0036] The set of physically questionable tags is obtained by calculating the historical consistency integral through first-order exponential smoothing recursion.

[0037] Further, the step of obtaining the classification decision map and its corresponding confidence score based on the feature vector and its consistency marker through dual-threshold pattern determination and feature selection includes:

[0038] Based on the feature vector, the interpretation pattern label map is obtained by using a dual threshold pattern determination of local ecological heterogeneity index and temporal change intensity, and a transition zone smoothing strategy.

[0039] Based on the interpretation pattern label map, representative pixels are obtained by dividing the region and extracting them. The marginal information gain of each data source within each window is obtained by calculating the confidence difference for each data source and taking the average value of the window.

[0040] Based on the marginal information gain combined with the consistency label, the gating mask and feature subset of each data source are obtained by window-by-window decision-making combined with first-order exponential smoothing recursion and long-term failure threshold determination.

[0041] Based on the aforementioned feature subset and thermal infrared validity score, a classification decision map and its corresponding confidence level are obtained through the basic probability allocation function and Dempster synthesis rule.

[0042] Furthermore, based on the interpreted pattern label map, representative pixels are obtained through region division and extraction. The marginal information gain of each data source within each window is obtained by calculating the confidence difference for each data source and taking the average value within a window, including:

[0043] Based on the interpretation pattern label map, a monitoring area grid is obtained through area division;

[0044] Based on the monitoring area grid and land use cover classification map, the representative pixel coordinate set in each window is obtained by sliding the meta window at a preset step size and performing stratified random sampling within each window.

[0045] Based on the representative pixel coordinate set and feature vector, a full feature set random forest classifier and a single-source ablation classifier are constructed. By calculating the difference in output confidence before and after removal and taking the average value within the window, the marginal information gain of each data source is obtained.

[0046] Furthermore, based on the marginal information gain combined with the consistency label, and through window-by-window decision-making combined with first-order exponential smoothing recursion and long-term failure threshold determination, the gating mask and feature subset of each data source are obtained, including:

[0047] Based on the marginal information gain and the preset judgment threshold, the physically questionable features are reduced in weight by window-by-window decision-making and combined with the weight decay factor to obtain the gating mask of each data source.

[0048] Based on the gating masks of each data source, the historical contribution integral is calculated by first-order exponential smoothing recursion.

[0049] Based on the historical contribution integral, a feature subset is obtained by determining the long-term failure threshold.

[0050] Furthermore, the environmental parameter set obtained based on the classification decision map and its corresponding confidence level through the CASA model and the quantitative inversion model includes:

[0051] Based on the classification decision graph and its corresponding confidence level, the initial inversion parameters are obtained through the CASA model;

[0052] Based on the initial inversion parameters and verification data, an accuracy evaluation report is obtained by calculating quantitative indicators.

[0053] Based on the accuracy evaluation report, an environmental parameter set is obtained by hierarchical labeling and generating a confidence layer.

[0054] Further, the initial inversion parameters obtained through the CASA model based on the classification decision graph and its corresponding confidence level include:

[0055] Based on the initial feature layer and meteorological data, the net primary productivity parameters are calculated pixel by pixel using the CASA model.

[0056] A feature space is constructed based on the first-resolution land surface temperature and the first-resolution vegetation index. The first-resolution soil moisture content distribution map is obtained by calculating the temperature-vegetation-drought index.

[0057] Chlorophyll concentration is obtained by inversion based on the first-resolution water body index and through empirical algorithms.

[0058] Based on the classification decision map and its corresponding confidence level, the water ecological parameters and thermal environment parameters are obtained by calculating the mean difference of the first-resolution surface temperature within the built-up area mask and the suburban area mask.

[0059] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0060] This invention is based on a multi-source remote sensing dataset and a basic mask layer set obtained by processing raw image data through quality screening rules. It obtains feature vectors and their consistency markers through feature extraction and sensor validity prior map construction. According to the dual threshold mode determination, feature screening and the sequential processing of CASA model and quantitative inversion model, an environmental parameter set is obtained. It realizes the minimum sufficient fusion and accurate collaborative interpretation of multi-source remote sensing data, solves the problems of insufficient monitoring stability in complex scenarios with a single data source, lack of physical consistency verification of cross-modal data, information redundancy increasing the computational burden, and easy misjudgment and elimination of real ecological change signals. It improves the overall accuracy, operational robustness and operational adaptability of ecological environment monitoring. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the steps of an ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to the present invention.

[0062] Figure 2 Flowchart of steps to obtain a multi-source remote sensing dataset and a base mask layer set;

[0063] Figure 3 Flowchart of the steps to obtain the feature vector and its consistency label;

[0064] Figure 4 A flowchart outlining the steps to obtain the classification decision map and its corresponding confidence level;

[0065] Figure 5 A flowchart outlining the steps to obtain the set of environmental parameters. Detailed Implementation

[0066] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0067] Identical parts are indicated by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, while the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific part, respectively.

[0068] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0069] like Figure 1 As shown, this embodiment provides an ecological environment monitoring method based on multi-source remote sensing collaborative interpretation, including:

[0070] Obtain raw image data and use quality screening rules to obtain a multi-source remote sensing dataset and a basic mask layer set;

[0071] Based on a multi-source remote sensing dataset, feature vectors and their consistency labels are obtained by extracting features and constructing a priori map of sensor effectiveness.

[0072] Based on feature vectors and their consistency markers, a classification decision map and its corresponding confidence level are obtained through dual-threshold pattern determination and feature selection.

[0073] Based on the classification decision map and its corresponding confidence level, the environmental parameter set is obtained through the CASA model and the quantitative inversion model.

[0074] The original image data consists of multi-source Earth observation images acquired through satellite remote sensing, including optical multispectral remote sensing data, synthetic aperture radar data, and thermal infrared remote sensing data. Typical examples include Sentinel-2 satellite 10-meter resolution multispectral data, Sentinel-1 satellite C-band dual-polarization synthetic aperture radar data, and Landsat series satellite 100-meter resolution thermal infrared data. Quality screening rules are image quality admission criteria set for each data source. For optical data, the screening criterion is a cloud cover rate of less than 20%; for radar data, the screening criterion is the absence of significant geometric distortion or missing data; and for thermal infrared data, the screening criterion is the absence of stripe noise or missing data. These rules are used to remove substandard original images, ensuring the reliability of the data foundation for subsequent interpretation. The multi-source remote sensing dataset is a collection of optical, radar, and thermal infrared remote sensing data after geometric registration, radiometric calibration, atmospheric correction, and image cropping preprocessing. All data are unified to the WGS-84 / UTM geographic reference system. Optical and radar data are resampled to a 10-meter spatial resolution, while thermal infrared data retains its original 100-meter spatial resolution for preprocessing. The sensor effectiveness prior map consists of pixel-by-pixel effectiveness scoring layers constructed for three types of data sources: optical, radar, and thermal infrared. Values ​​range from 0 to 1 and are calculated using the quality impact factor of the corresponding data source. This quantifies the observation reliability of each data source at each spatial location, providing a basis for fusion weight settings and feature selection. The aerosol correction factor is determined based on aerosol optical thickness, which is obtained through inversion from satellite remote sensing data such as MODIS, or through simulation using atmospheric parameters such as the 6S radiative transfer model combined with NCEP reanalysis data. After normalization, the aerosol optical thickness is obtained as an aerosol correction factor ranging from 0 to 1. The larger the aerosol optical thickness, the closer the correction factor value is to 0, indicating a more severe attenuation of surface signals by aerosols and lower effectiveness of the pixel's optical data.

[0075] Feature selection is a mechanism that dynamically retains or removes features from various data sources based on marginal information gain assessment and consistency labeling. This achieves minimal and sufficient fusion of multi-source features, eliminating redundant features and those with negative contributions. The corresponding confidence level represents the credibility of the classification result for each pixel in the classification decision map, given by the maximum confidence level in the basic probability assignment quality function after fusion using Dempster-Shafer evidence theory, with a value ranging from 0 to 1. The CASA model is a classic mechanistic model for estimating net primary productivity of vegetation based on the principle of light energy utilization, used to calculate net primary productivity parameters. The quantitative inversion model is a collection of models including temperature and vegetation drought index inversion, empirical inversion of water chlorophyll concentration, and calculation of urban heat island intensity, used to invert ecological and environmental parameters such as soil moisture content, water chlorophyll concentration, and urban heat island intensity.

[0076] This embodiment constructs a unified spatial data benchmark through quality screening and standardized preprocessing of multi-source remote sensing data. It establishes a two-layer quality verification mechanism through cross-modal physical consistency verification and sensor validity prior maps. It achieves dynamic minimum full fusion through dual-threshold adaptive mode determination and feature screening driven by marginal information gain. It outputs full-dimensional ecological environment monitoring results through CASA model and multi-parameter quantitative inversion. This enables continuous and reliable ecological environment monitoring results even under cloudy and rainy weather conditions, effectively solving the technical problem of continuous monitoring under complex weather conditions caused by a single remote sensing data source. At the same time, physical consistency verification avoids the problem of misjudging real ecological change signals as data noise and incorrectly eliminating them. Feature screening driven by information bottlenecks significantly reduces the computational complexity of multi-source data fusion, improving the accuracy, robustness and operational efficiency of ecological environment monitoring.

[0077] Furthermore, this embodiment provides a step for obtaining a multi-source remote sensing dataset and a basic mask layer set through quality screening rules, including:

[0078] Based on the original image data, a multi-source registered image set is obtained through intensity cross-correlation registration algorithm and multi-view processing;

[0079] Based on a multi-source registered image set, the basic data layer is obtained through radiometric calibration transformation and atmospheric correction of the radiative transfer model.

[0080] Based on the basic data layer, a multi-source remote sensing dataset and a basic mask layer set are obtained by clipping the boundaries of the monitoring area and classifying ground features using prior knowledge.

[0081] Among them, the intensity cross-correlation registration algorithm is a geometric registration method based on the correlation of gray intensity in images to extract corresponding feature points and solve spatial transformation parameters. It is used to achieve spatial alignment of remote sensing images from different sources. The intensity cross-correlation registration algorithm uses the first-resolution optical image as the benchmark. The first-resolution optical image is an optical multispectral remote sensing image with a native spatial resolution of up to ten meters, with Sentinel-2 satellite multispectral image as a typical example, serving as the unified spatial resolution benchmark for this method. Multi-view processing is an image processing operation that performs multi-view averaging on synthetic aperture radar images in the azimuth direction. It is used to balance the spatial resolution and speckle noise level of radar images. The azimuth multi-view factor is set according to the native azimuth resolution of synthetic aperture radar images, the target spatial resolution requirements, and the speckle noise suppression requirements, satisfying the mapping constraint: the azimuth multi-view factor is positively correlated with the speckle noise suppression effect and negatively correlated with the azimuth spatial resolution performance; The better the azimuth resolution and the lower the target spatial resolution requirement, the larger the selectable multi-view factor can be. Conversely, a smaller multi-view factor needs to be selected to preserve resolution performance. Its value ranges from 1 to 4, determined based on the matching relationship between the native resolution of mainstream spaceborne synthetic aperture radar and the target resolution of ten meters. A multi-view factor exceeding 4 will result in excessive loss of azimuth resolution, making it impossible to maintain the geometric accuracy of the target spatial resolution through subsequent resampling. A multi-view factor of 1 corresponds to the native resolution imaging mode without multi-view processing. For example, in this embodiment, the value is 2. The native azimuth resolution of the synthetic aperture radar data used in this embodiment is about 20 meters. When the multi-view factor is set to 2, a balance can be achieved between speckle noise suppression and resolution preservation. It can effectively reduce multiplicative speckle noise through multi-view averaging and match the target spatial resolution of ten meters through subsequent resampling processing, maintaining consistency with the spatial reference of the first resolution optical image. A multi-source registered image set is a collection of various remote sensing images that have been geometrically registered and have a unified geographic reference system, including first-resolution optical registered images, first-resolution radar registered images, and second-resolution thermal infrared registered images.

[0082] Radiometric calibration is the process of converting the raw digital quantization values ​​of remote sensing images into physically meaningful radiance values. This process eliminates differences in sensor response and ensures the comparability of radiance data from different time phases and sensors. Atmospheric correction of the radiative transfer model is a method that simulates and corrects the influence of the atmosphere on electromagnetic wave propagation based on the radiative transfer mechanism. This method eliminates radiative distortion caused by atmospheric scattering and absorption, obtaining the true surface radiance parameters. The base data layer consists of a multi-source remote sensing physical quantity dataset after radiometric calibration and atmospheric correction, including surface reflectance, topographically corrected backscattering coefficients, and surface temperature datasets. Surface reflectance is the true surface radiance parameter after atmospheric correction. The real spectral reflectance, obtained from optical images through radiometric calibration and atmospheric correction, ranges from 0 to 1. The topographically corrected backscattering coefficient is the radar backscattering coefficient after topographic normalization correction, obtained from radar images through radiometric calibration and topographic correction, and is measured in decibels. It is used to characterize the ability of the land surface to backscatter radar beams. The land surface temperature dataset is a collection of real land surface temperature data after atmospheric correction, including a first land surface temperature and a second land surface temperature. The first land surface temperature is the land surface temperature data that has been downscaled to achieve the first resolution, with a spatial resolution of 10 meters. The second land surface temperature is the land surface temperature data at the original resolution of the thermal infrared data, with a spatial resolution of 100 meters.

[0083] Monitoring area boundary clipping is a processing operation that crops images based on preset monitoring range boundaries to remove redundant data outside the monitoring range and unify the spatial coverage of various data sources. Land feature prior knowledge classification is a method for classifying land features based on prior rules of spectral and structural characteristics, used to quickly generate binary mask layers with clear physical meaning. Structural feature prior rules establish judgment conditions based on the spatial geometric shape, texture pattern, and surface roughness of different land features. For example, built-up areas have regular rectangular boundaries and dense edge features; forest land has high surface roughness and uniform texture; water bodies have smooth surfaces and weak echo characteristics in radar images; and cultivated land has regular striped textures. Multi-source remote sensing datasets serve as the input basis for subsequent feature extraction and collaborative interpretation. The basic mask layer set is a collection of binary layers generated based on land feature prior knowledge, including three categories: water body masks, vegetation masks, and built-up area masks, used to limit the applicable spatial range for subsequent specialized calculations.

[0084] Specifically, such as Figure 2As shown, geometric registration and multi-view processing are first performed on the acquired raw image data. The first-resolution optical image is selected as the reference image, with its geographic reference system being the UTM projection under the WGS-84 coordinate system, which has high-precision geometric positioning capabilities. For the raw synthetic aperture radar (SAR) image, an intensity cross-correlation registration algorithm is used for registration. First, corresponding feature points are automatically extracted from the reference image and the radar image to be registered. Feature points are selected as spatially stable ground features such as road intersections and building corners. Then, affine transformation parameters are solved based on the coordinates of the corresponding points to achieve spatial alignment of the radar image with respect to the reference image. Considering the characteristics of SAR's native range resolution of approximately five meters and azimuth resolution of approximately twenty meters, multi-view processing is first performed on the azimuth direction, with the azimuth multi-view factor set to 2. Speckle noise is suppressed by multi-view averaging. Then, the multi-view processed radar image is resampled to a spatial resolution of ten meters to maintain consistency with the reference image resolution. For the raw thermal infrared image, geometric registration is performed using corresponding feature point matching. After registration, its native spatial resolution of one hundred meters is maintained, and no upsampling processing is performed. The registration accuracy of all data sources is controlled within one pixel, resulting in a multi-source registered image set with a unified geographic reference system.

[0085] Then, radiometric calibration transformation and atmospheric correction using the radiative transfer model were performed on the multi-source registered image set. For optically registered images, radiometric calibration transformation was performed first, converting the digital quantization values ​​of the images into the top-of-atmosphere radiance based on the sensor calibration parameters. Then, the apparent reflectance of the top of the atmosphere was calculated by combining the Earth-Sun distance and the solar zenith angle. The formula for calculating the apparent reflectance is based on the fundamental principles of remote sensing radiometric calibration, and its calculation formula is as follows: ;in The apparent reflectance at the top of the atmosphere. This refers to the radiance at the top of the atmosphere. The distance between the Earth and the Sun. This refers to the solar spectral irradiance outside the atmosphere. The zenith angle of the sun. The center wavelength corresponds to the optical band, used to characterize the spectral band to which the above radiation parameter values ​​belong. The unit is usually micrometers. Different multispectral bands correspond to different center wavelength values; for example, the center wavelength of the visible red band is approximately 0.66 micrometers, and the center wavelength of the near-infrared band is approximately 0.84 micrometers. Then, the 6S radiative transfer model is used to perform atmospheric correction. Input parameters include the sensor band response function, atmospheric aerosol optical thickness, water vapor content, and ozone content. Aerosol optical thickness can be obtained from MODIS products or ground-based measured data. By simulating atmospheric scattering and absorption effects along the sun-surface-sensor path, the true surface reflectance, i.e., the surface reflectance data at the first resolution, is output. For radar-registered imagery, radiometric calibration conversion is first performed. Based on the calibration constants and system correction factors provided by the satellite data provider, the digital quantization values ​​of the image are converted into backscattering coefficients, measured in decibels. Radar data is less affected by the atmosphere and does not require atmospheric correction. Only terrain normalization correction is needed based on the digital elevation model and the radar system's incident angle information to eliminate the influence of terrain undulations on the backscattering coefficient, yielding the terrain-corrected backscattering coefficient. Radiometric conversion is performed on the thermal infrared registered image. Based on Planck's radiation law, the digital quantization values ​​of the image are converted into the atmospheric top brightness temperature using the calibration constant of the thermal infrared band. Then, atmospheric correction is performed using the MODTRAN atmospheric radiative transfer model, with atmospheric profile data including temperature, humidity, and air pressure as input. The effects of upward and downward atmospheric radiation are eliminated, and the true surface temperature at the second resolution is output and incorporated into the surface temperature dataset.

[0086] The monitoring area boundary clipping and land feature prior knowledge classification are performed on the basic data layer. Based on the administrative or natural geographical boundary vector files of the monitoring area, the surface reflectance, topographically corrected backscattering coefficient, and surface temperature datasets are clipped to ensure that the spatial coverage of all data is completely consistent with the monitoring area. A base mask layer containing water body masks, vegetation masks, and built-up area masks is generated based on prior knowledge of land features. A normalized water index is calculated based on the first-resolution surface reflectance, and the water body range is extracted by segmentation using a preset water index segmentation threshold to obtain the water body mask. A normalized vegetation index is calculated based on the first-resolution surface reflectance, and the vegetation coverage area is extracted by segmentation using a preset vegetation index segmentation threshold to obtain the vegetation mask. A built-up area mask is generated by extracting the built-up area range based on the preliminary land use classification results or the urban impermeable surface index. The preliminary land use classification results are the initial land feature classification results obtained based on optical multispectral imagery and a random forest classifier, predicted according to the spectral characteristics of surface reflectance and a pre-trained random forest classification model. The urban impermeable surface index is a spectral index characterizing the degree of urban impermeable surface coverage, calculated based on the combination of visible light and near-infrared reflectance of buildings, roads, and other land features.

[0087] The preset water body index segmentation threshold is set based on the spectral characteristics of the water body in the monitoring area, the image time phase, and the water quality background conditions. The closer the water body's spectral characteristics are to the high reflectance green light of clean, deep water, the stronger the absorption of near-infrared features, the higher the image time phase (during the high-water season with low aquatic vegetation coverage), and the clearer the water with less suspended matter, the higher the overall water body index value, and the segmentation threshold can be adjusted accordingly. Conversely, the lower the water body index value is, the lower the overall water body index value is, and the segmentation threshold needs to be adjusted accordingly, if the water body is turbid with high sediment content, has many shallow areas exposed during the dry season, or has a high degree of eutrophication. Its value ranges from 0 to 0.2, based on normalized water body data. The physical characteristics of the index and the need to distinguish ground features determine the value range. The normalized water index has a theoretical range of -1 to 1. The index value of natural water bodies is generally higher than that of bare soil and vegetation. A threshold higher than 0.2 will miss turbid shallow water bodies, while a threshold lower than 0 will easily misidentify moist soil and riverbanks as water bodies. For example, in this embodiment, the value is 0. The water bodies in the monitoring area of ​​this embodiment are mainly clean and open water areas. The image phase is the high water period, and the spectral boundaries of the water bodies are clear. Using 0 as the threshold can completely extract the entire water body range, while effectively avoiding the misidentification of surrounding moist surfaces and ensuring the integrity of the water body mask. The preset vegetation index segmentation threshold is set based on the distribution of vegetation types in the monitoring area, the phenological period of the image, and the characteristics of the underlying soil. If the vegetation type in the monitoring area is mainly dense forest, the image is taken during the peak growth period, and the underlying soil is light-colored and dry, then the greater the difference in the vegetation index between the vegetation and the background, the higher the segmentation threshold can be. Conversely, if the vegetation is mainly sparse herbaceous, the image is taken during the dormant period, and the underlying soil is dark and moist, then the overall vegetation index value is low, and the segmentation threshold needs to be lowered accordingly. Its value range is 0.1 to 0.3, determined based on the land cover differentiation law of the normalized vegetation index. The theoretical range of the normalized vegetation index is -1 to 1, and bare soil areas refer to... The values ​​mostly fluctuate around 0. 0.1 is the typical lower limit for distinguishing between bare soil and sparse vegetation, and 0.3 is the commonly used upper limit for distinguishing between low-coverage vegetation and medium-to-high-coverage vegetation. If the threshold is below 0.1, bare soil and rocks are easily mistakenly classified as vegetation, and if the threshold is above 0.3, sparse grassland, fallow land and other low-coverage vegetation will be missed. For example, in this embodiment, the value is 0.2. The monitoring area in this embodiment includes multiple vegetation types such as woodland, grassland and cultivated land. The image phase is in the middle of the growing season, and the underlying surface is mainly zonal yellow-brown soil. Using 0.2 as the threshold can take into account the completeness of the extraction of vegetation with different coverage, while effectively removing non-vegetation areas such as bare soil and water bodies, and ensuring the classification accuracy of vegetation mask.

[0088] The pre-trained random forest classification model is a multi-decision tree voting classification model using an ensemble learning mechanism. Its pre-training method is as follows: Within the monitoring area, training samples are collected for six land cover categories—forest, grassland, cultivated land, water bodies, built-up areas, and bare land—based on high-resolution reference images or historical land use data as ground truth. At least 500 sample pixels are selected for each category. The input feature set is the band values ​​of surface reflectance, normalized vegetation index, and normalized water body index. The total number of decision trees is set to 100, the number of candidate features randomly selected during each node split is the square root of the total number of features, and the minimum number of samples per leaf node is 1. The classification model is obtained through supervised training. The input of this random forest classification model is a feature vector composed of multi-band surface reflectance, normalized vegetation index, and normalized water body index corresponding to a single pixel. The output is the posterior probability value of each land cover category corresponding to each pixel and the classification result with the highest confidence. The random forest classification model is a parallel integrated tree structure, consisting of one hundred independent decision trees. Each decision tree generates hierarchical nodes by splitting from the root node to the leaf node through feature thresholds. Finally, the classification results of all decision trees are used to obtain the final classification result of each pixel through a majority voting mechanism.

[0089] The cropped base data layer and the base mask layer together constitute the multi-source remote sensing dataset and the base mask layer set. In this embodiment, the azimuth multi-view factor, various mask segmentation thresholds, and registration accuracy control values ​​are merely examples. Those skilled in the art can set these values ​​according to the actual monitoring area and data source conditions; this embodiment does not impose any limitations on them.

[0090] This embodiment achieves a high-precision, high-reliability basic data layer by performing quality screening on the original image data to remove unqualified data, using an intensity cross-correlation registration algorithm based on the first-resolution optical image to unify multi-source images under the same geographic reference frame, balancing the azimuth and range resolution differences of SAR images and suppressing speckle noise through multi-view processing, converting digital quantization values ​​into physical quantities through radiometric calibration transformation, and performing atmospheric correction on optical and thermal infrared data using the 6S radiative transfer model and the MODTRAN radiative transfer model to eliminate atmospheric effects. This realizes spatial alignment, radiometric normalization, and range unification of multi-source heterogeneous remote sensing data, improving the spatial consistency and physical reliability of subsequent feature extraction and collaborative interpretation, and laying a standardized data foundation for subsequent multi-source remote sensing collaborative interpretation and ecological environment monitoring.

[0091] Furthermore, this embodiment provides a step for obtaining feature vectors and their consistency labels based on a multi-source remote sensing dataset by extracting features and constructing a sensor effectiveness prior map, including:

[0092] Based on a multi-source remote sensing dataset, an initial feature layer was obtained through normalized difference operation, gray-level co-occurrence matrix texture analysis, and split-window algorithm.

[0093] Based on multi-source remote sensing datasets and digital elevation models, thermal infrared effectiveness scores are obtained through attenuation calculation, local incident angle geometric discrimination, and cloud cover and emissivity correction.

[0094] Based on the initial feature layer and thermal infrared effectiveness score, the first resolution surface temperature is obtained by establishing a second-order polynomial regression relationship and performing inverse distance weighted interpolation on the regression residuals.

[0095] Based on the first-resolution surface temperature and the initial feature layer, feature vectors are obtained through synchronous window standardization, and consistency labels are obtained according to the coupling relationship.

[0096] The initial feature layer is a collection of various remote sensing feature parameters directly extracted from multi-source remote sensing datasets or obtained through preliminary calculations, including the first-resolution vegetation index, the first-resolution water index and land use cover classification map, the first-resolution structural feature layer, and the second-resolution land surface temperature and urban heat island intensity. Normalized difference calculation is a method for constructing spectral indices using the ratio of the reflectance difference between two different bands to their sum, used to highlight the spectral characteristics of specific land features and enhance and extract corresponding land feature information. The random forest classifier is a multi-decision tree voting classification model based on ensemble learning mechanisms, used to classify land features... Other automatic classifications are obtained through supervised training based on the spectral characteristics of surface reflectance and sample labels; the first-resolution vegetation index is a spectral index characterizing the growth status and coverage of vegetation at the first-resolution scale, calculated based on the normalized difference between the near-infrared band and the red band; the first-resolution water body index is a spectral index characterizing the distribution range of water bodies at the first-resolution scale, calculated based on the normalized difference between the green band and the near-infrared band; the land use cover classification map is a spatial distribution map obtained after classifying each pixel in the monitoring area into land cover categories, covering six land cover categories: forest, grassland, cultivated land, water bodies, built-up areas, and bare land.

[0097] Dual-polarization channel information extraction is a processing operation that separates backscattered signals of different polarization modes from synthetic aperture radar data, used to obtain the scattering characteristics of different polarization dimensions of the land surface; gray-level co-occurrence matrix texture analysis is an image processing method that extracts texture features based on the spatial joint distribution of pixel gray levels, used to characterize the spatial structure and roughness of the land surface; the first resolution structural feature layer is a feature set characterizing the land surface structure and texture attributes at the first resolution scale, composed of a combination of dual-polarization backscattering coefficients and texture features; the split-window algorithm is a land surface temperature inversion algorithm that uses the difference in absorption of atmospheric water vapor by two adjacent thermal infrared bands to correct for atmospheric influence, used to invert the true land surface temperature from thermal infrared brightness temperature data; temperature difference calculation is a method for calculating the thermal environment difference index by calculating the difference between the average land surface temperatures of different regions, used to quantify the degree of urban heat island effect relative to suburbs; the second resolution land surface temperature is the true land surface temperature data at the original resolution scale of thermal infrared data, with a spatial resolution of 100 meters; urban heat island intensity is a thermal environment index characterizing the degree to which the temperature in the built-up area is higher than that in the suburbs, obtained by subtracting the average surface temperature in the suburbs from the average surface temperature in the built-up area.

[0098] Digital elevation models (DEMs) are raster datasets describing the undulations of the Earth's surface elevation. They are used to calculate terrain parameters and support radar terrain correction and geometric distortion identification, and are generated through methods such as spaceborne photogrammetry or radar interferometry. Attenuation calculation is a process for calculating the attenuation degree of various data quality influencing factors, used to quantify the weakening effect of different factors on data reliability. Local incident angle geometric discrimination is a discrimination method based on the geometric relationship between the terrain slope and the radar beam to determine the radar illumination state, used to identify geometric distortion areas such as radar shadows and overlaps. The radar system incident angle is the deflection angle of the radar beam relative to the vertical direction of the nadir point, determined by satellite orbit parameters and imaging mode, and belongs to the radar... The inherent imaging parameters of the data; the radar shadow binary identifier is a binary raster layer that marks the radar shadow area, with a value of 1 for the shadow area and a value of 0 for the non-shadow area; the ratio of the incident angle difference to the radar system incident angle is the proportion of the difference between the radar system incident angle and the local incident angle to the system incident angle, used to characterize the severity of the overlay phenomenon; the restriction to the zero-to-one interval is a numerical constraint operation that truncates the calculation results to the range of 0 to 1, used to ensure that the value of the overlay intensity has a normalized physical meaning; cloud cover and emissivity correction are effectiveness correction processes for the cloud pollution effect and surface emissivity differences of optical and thermal infrared data, used to improve the accuracy of multi-source data effectiveness scoring.

[0099] Superimposed intensity is a normalized index characterizing the severity of radar superimposed phenomena, ranging from 0 to 1; a larger value indicates more significant superimposed distortion. Radar intensity data refers to the grayscale data characterizing backscattering intensity in synthetic aperture radar (SAR) images, obtained from the intensity information of the radar images before radiometric calibration. Metadata is supplementary descriptive data recording imaging parameters, sensor information, and processing history of remote sensing images, acquired synchronously from satellite data. Equivalent apparent number is a quantitative parameter measuring the speckle noise level of radar images, which can be estimated from the mean and variance of radar intensity data or directly read from image metadata. The speckle noise coefficient is a normalized index characterizing the multiplicative noise level of radar images, derived from... The value is calculated by the reciprocal of the square root of the equivalent number of views, ranging from 0 to 1. The larger the value, the more severe the noise. The shadow area is the back slope area of ​​the ground that cannot be illuminated by the radar beam. It is determined by the geometric condition that the local incident angle is greater than or equal to the preset incident angle threshold for shadow determination. The incident angle threshold for shadow determination is set according to the geometric illumination mechanism between the radar beam and the ground slope. It corresponds to the critical angle at which the radar beam cannot reach the slope at all. For example, in this embodiment, the value is 90 degrees. The thermal infrared effectiveness score is a pixel-by-pixel effectiveness score system constructed for multi-source remote sensing data sources. The value ranges from 0 to 1 and covers the effectiveness components of three types of data sources: optical, radar, and thermal infrared.

[0100] The second-order polynomial regression relationship is a polynomial fitting relationship constructed using the first-order, second-order, and constant terms of the independent variable. It is used to establish a statistical correlation between coarse-resolution land surface temperature and vegetation indices, achieving spatial downscaling of land surface temperature. The regression residual is the difference between the measured value of the coarse-resolution land surface temperature and the polynomial fitting value, used to characterize the degree of deviation of temperature in a local area. The regression coefficients are the weight parameters corresponding to the constant, first-order, and second-order terms in the second-order polynomial, obtained through least squares fitting. The coarse pixel is the land surface temperature pixel at the second-resolution scale, with one hundred first-resolution pixels corresponding to the coverage area of ​​a single coarse pixel. The regression residual value is the specific numerical value of the land surface temperature fitting residual corresponding to each coarse pixel. The preliminary first-resolution land surface temperature is a preliminary estimate of the first-resolution land surface temperature obtained only through second-order polynomial regression calculation, and has not yet been incorporated into the calculation. Local residual information; the effective pixel count is the number of coarse pixels in the neighborhood of a coarse pixel whose thermal infrared validity score is greater than a preset thermal infrared validity score threshold, used to determine the data support conditions for residual interpolation. The thermal infrared validity score threshold is set as the minimum reliable quality requirement for thermal infrared data to participate in downscaling calculation. For example, in this embodiment, the value is 0.5; the preset quantity is the threshold for determining whether to perform residual interpolation, set according to the quality requirements of thermal infrared data and downscaling accuracy requirements, with a value range of 3 to 9. For example, in this embodiment, the value is 5; inverse distance-weighted interpolation is a spatial interpolation method that uses the power of the reciprocal of the distance as the weight, used to transfer coarse-resolution residuals to high-resolution grids; the first resolution surface temperature is the surface temperature data that has reached a spatial resolution of ten meters after downscaling, which can achieve spatial scale matching with optical and radar features.

[0101] The synchronous window standardization is a dimensionless normalization method based on the statistical characteristics of historical synchronous data to eliminate the influence of seasonal periodic fluctuations and extract true ecological anomaly signals. The feature vector is a multi-dimensional vector formed by dimensionlessly standardizing each component of the initial feature layer, containing six components: normalized vegetation index, normalized water index, 10-meter resolution surface temperature, radar cross-polarization backscattering coefficient, altitude, and slope. This is used to characterize the multi-dimensional ecological attributes of pixels under a unified dimension. The coupling relationship refers to the inherent correlation between different modal remote sensing features based on physical mechanisms, used for cross-modal physical consistency verification. The consistency label is the labeling of each component of the feature vector through cross-modal physical consistency verification. The "Physical Doubt Status" is used to identify the physical credibility of features and provide a basis for subsequent fusion weight adjustments. Vegetation coverage is a parameter characterizing the proportion of the vertical projection area of ​​the vegetation canopy to the total surface area, ranging from zero to one, calculated based on the first-resolution vegetation index. Filtered pixels are the operation of selecting pixels that meet the verification range based on vegetation coverage and land cover category conditions, used to limit the applicable area for physical consistency verification. The "Valid Object Mask" is a binary layer marking pixels that meet the physical consistency verification conditions; only pixels that meet the vegetation coverage requirements and whose land cover category is vegetation are marked as valid. The consistency index is a normalized index that quantifies the degree of physical coupling consistency between multimodal features, ranging from zero to one; the closer the value is to one, the better the physical consistency between features.

[0102] The preset consistency index threshold is the threshold for triggering the physical doubt marker. It is set according to the verification accuracy requirements and historical data statistics, and its value ranges from 0.6 to 0.8. For example, in this embodiment, the value is 0.7. The physical doubt marker is a status indicator that marks the physical consistency of pixel features as questionable, used to indicate that the observation data of that pixel may have quality problems or abnormal ecological changes. The suspected data source is the corresponding data source that caused the physical consistency abnormality, which is determined by comparing the deviation of each feature from the predicted value. The validity score is the pixel-by-pixel validity score result of the corresponding data source, which comes from the thermal infrared validity score system. The feature status is the status of each data source feature after physical consistency verification. The physical consistency features are categorized into three states: normal passage, reduced-weight passage, and complete blockage. The weight decay factor is a proportional coefficient that reduces the weight of physically suspicious features, ranging from 0 to 1. For example, in this embodiment, the value is 0.5. The first-order exponential smoothing recursion is a recursive algorithm that achieves time-dimensional smoothing by weighted summation of historical and current values, used to distinguish between single-period accidental anomalies and continuous real changes. The historical consistency integral is a cumulative index that characterizes the long-term trend of physical consistency of pixels, calculated through first-order exponential smoothing recursion. The physical suspicion label set is a set of labels containing complete physical suspicion information, including the physical suspicion status of each pixel, the suspected data source, the weight decay factor, and the status awaiting ground verification warning.

[0103] Specifically, such as Figure 3 As shown, firstly, for the first-resolution surface reflectance data in the multi-source remote sensing dataset, the first-resolution vegetation index and the first-resolution water index are calculated by normalized difference operation. The first-resolution vegetation index is obtained by dividing the difference in reflectance between the near-infrared band and the red band by the sum of the two, which can intuitively reflect the growth status and coverage of vegetation. The first-resolution water index is obtained by dividing the difference in reflectance between the green band and the near-infrared band by the sum of the two, which can highlight the spectral characteristics of water bodies. Based on this, a random forest classifier is used for land use cover classification. The random forest classifier is pre-trained in a supervised manner: sample ground truth values ​​are selected in the monitoring area based on high-resolution reference images or historical land use data. No less than 500 sample pixels are selected for each of the six categories: forest land, grassland, cultivated land, water bodies, built-up areas, and bare land. The surface reflectance values ​​of each band, the first-resolution vegetation index, and the first-resolution water body index are used as input features. The total number of decision trees is set to 100, the number of features randomly selected when splitting nodes is the square root of the total number of features, and the minimum number of samples for leaf nodes is 1. The classification model is trained. During classification, the surface reflectance data is input into the model pixel by pixel, and the first-resolution land use cover classification map is output.

[0104] For the terrain-corrected backscattering coefficients in the multi-source remote sensing dataset, dual-polarization channel information extraction was performed to obtain the backscattering coefficient values ​​of the VV and VH polarization channels respectively. Subsequently, texture analysis was carried out using the gray-level co-occurrence matrix. The gray-level co-occurrence matrix represents texture features by statistically analyzing the gray-level joint distribution of pixel pairs at specific directions and distances. Three texture parameters, contrast, correlation, and entropy, were extracted from it. During the calculation, the average values ​​of four directions, 0°, 45°, 90°, and 135°, were used to eliminate directional bias. The analysis window was set to seven by seven pixels. The dual-polarization backscattering coefficients were combined with the three types of texture features to obtain the first resolution structural feature layer.

[0105] For the second-resolution land surface temperature data in the multi-source remote sensing dataset, the split-window algorithm is used to retrieve the second-resolution land surface temperature, which is the true land surface temperature at the native resolution of 100 meters. The split-window algorithm is based on the difference in the absorption characteristics of atmospheric water vapor by two adjacent thermal infrared bands. It estimates the atmospheric transmittance and surface emissivity by using the brightness temperature difference between the two channels, and then corrects for atmospheric influences to obtain the second-resolution land surface temperature. The input of this algorithm is the brightness temperature of the top of the atmosphere and the surface emissivity of the two thermal infrared bands. The surface emissivity is determined according to the land use type. For example, water bodies are taken as 0.99, vegetation as 0.98, bare soil as 0.95, and built-up areas as 0.92 to 0.96. On this basis, the temperature difference is calculated to obtain the urban heat island intensity. Based on the land use cover classification map, pixels of built-up areas and suburban pixels within a buffer zone of 5 to 10 kilometers outside the built-up areas are extracted respectively. The mean of the second-resolution land surface temperature of the two types of areas is calculated. The urban heat island intensity is obtained by subtracting the mean of the suburban area from the mean of the built-up area. Then, based on the digital elevation model, the slope and aspect of each pixel are calculated. The local incident angle is calculated by combining the radar system's incident angle and azimuth angle. The local incident angle is the angle between the radar incident wave vector and the surface slope normal vector. When the local incident angle is greater than or equal to the preset shadow determination incident angle threshold, the pixel is determined to be a radar shadow area. The corresponding radar shadow binary identifier is set to 1, and the non-shadow area is set to 0, thus obtaining the radar shadow binary identifier.

[0106] For pixels in non-shaded areas, the overlay intensity is further calculated. For pixels with a local incident angle smaller than the radar system incident angle and a slope facing the satellite direction, the difference between the radar system incident angle and the local incident angle is calculated. The ratio of this difference to the radar system incident angle is used as the base value of the overlay intensity, and the result is limited to the range of 0 to 1. The closer the overlay intensity is to 1, the more severe the overlay phenomenon, and the lower the reliability of the corresponding radar data. Then, the speckle noise coefficient is calculated. The equivalent number of views is estimated from the mean and variance of the radar intensity data, or the equivalent number of views parameter is directly read from the image metadata. The reciprocal of the square root of the equivalent number of views is used as the speckle noise coefficient. The larger the equivalent number of views, the smaller the speckle noise coefficient and the higher the image quality.

[0107] Then, multi-dimensional correction and effectiveness synthesis are completed. The optical effectiveness component is calculated by combining cloud coverage probability, cloud shadow probability, and aerosol correction factor. The cloud coverage probability is obtained based on cloud detection results from optical multispectral imagery. Cloud pixel distribution is extracted using the Fmask cloud detection algorithm or the quality assessment band provided by the satellite product. The cloud pixel proportion is obtained by statistically analyzing the neighborhood window of each pixel. The value ranges from 0 to 1, with a larger value indicating a higher degree of cloud coverage. The cloud shadow probability is calculated based on cloud area distribution and imaging geometric parameters. Combining the solar zenith angle, solar azimuth angle, and cloud top height parameters at the imaging time, the spatial range of cloud shadow is determined by geometric projection. The cloud shadow probability is obtained by neighborhood smoothing and statistical analysis of each pixel. The value ranges from 0 to 1, with a larger value indicating a more significant cloud shadow impact. The aerosol correction factor is calculated based on contemporaneous aerosol optical thickness data combined with the 6S radiative transfer model. It is normalized to the 0-1 range based on the observation quality without aerosol influence. The closer the value is to 1, the smaller the aerosol attenuation impact. The thermal infrared effectiveness component is calculated by combining cloud cover probability and surface emissivity correction factor. The cloud cover probability and optical component are obtained from the same source cloud detection results, which are then resampled and matched to the thermal infrared spatial resolution. The surface emissivity correction factor is calculated based on land use cover classification and land cover emissivity characteristics. A nominal surface emissivity value is assigned to each pixel according to the classification results, and the correction factor is obtained by normalizing the influence of emissivity uncertainty on temperature inversion accuracy. The value ranges from 0 to 1, with a value closer to 1 indicating a smaller emissivity deviation and higher reliability of thermal infrared observation. The radar effectiveness component is calculated according to the shadow, overlay, and speckle noise coefficients mentioned above. For radar shadow areas, the effectiveness score is directly set to 0. For non-shadow areas, the radar effectiveness score is the product of 1 minus the overlay intensity, 1 minus the speckle noise coefficient, and 1 minus the shadow binary identifier. The three types of effectiveness components together constitute a complete sensor effectiveness prior map. The value range of each type of score is from 0 to 1, with a value closer to 1 indicating a higher reliability of the corresponding data source observation in that pixel.

[0108] Then, a second-order polynomial regression relationship between land surface temperature and normalized vegetation index (NVI) is established at the second resolution scale. The NVI at the first resolution is aggregated to a 100-meter resolution by averaging. Using the aggregated NVI as the independent variable and the land surface temperature at the second resolution as the dependent variable, the regression coefficients of the second-order polynomial are obtained by least squares fitting, including constant, linear, and quadratic coefficients. The difference between the measured and fitted land surface temperature values ​​is calculated for each coarse pixel to obtain the regression residuals for each coarse pixel. The fitted regression coefficients are then applied to the NVI at the first resolution, and the second-order polynomial is substituted into each 10-meter pixel to obtain the preliminary first-resolution land surface temperature. This result is a preliminary estimate based on the statistical relationship between vegetation and temperature and does not yet incorporate local residual information.

[0109] Then, the validity determination and interpolation propagation of residual interpolation are performed. For each 100-meter coarse pixel, the number of valid coarse pixels with a thermal infrared validity score greater than the preset thermal infrared validity score threshold within its three-by-three neighborhood is counted. If the number of valid pixels reaches the preset number or above, it indicates that the area has sufficient high-quality thermal infrared observation data to support residual interpolation. In this case, the inverse distance-weighted interpolation method is used to propagate the regression residual value of the coarse pixel to the 10-meter pixels it covers. The interpolation weight is inversely proportional to the square of the distance. The closer the coarse pixel residual is, the greater its contribution to the 10-meter pixel. The 10-meter resolution residual obtained by interpolation is superimposed with the preliminary first-resolution surface temperature to obtain the final first-resolution surface temperature. If the number of valid pixels is less than the preset number, it is determined that the thermal infrared observation data in the area is insufficient and marked as having no valid thermal infrared inversion state. Downscaling is not performed.

[0110] Then, an effective object mask is generated. Vegetation coverage is calculated based on the first-resolution vegetation index. The value at 5% of the cumulative frequency of the first-resolution vegetation index in the entire image is used as the value of bare soil pixels, and the value at 95% of the cumulative frequency is used as the value of pure vegetation pixels. The vegetation index is subtracted from the bare soil value, and then divided by the difference between the pure vegetation value and the bare soil value to obtain a vegetation coverage value ranging from 0 to 1. Pixels with vegetation coverage greater than a preset effective verification vegetation coverage threshold and whose land use type is forest, grassland, or cultivated land are selected and marked as effective objects, forming an effective object mask. Only pixels within the mask area participate in subsequent physical consistency verification. The effective verification vegetation coverage threshold is set according to the applicable vegetation coverage conditions of cross-modal physical coupling relationships; for example, in this embodiment, the value is 0.3.

[0111] Furthermore, a consistency index is calculated and a physical suspicion marker is executed. Based on high-quality historical data from the same period, second-order polynomial regression functions are fitted between the uniform vegetation index and surface temperature, and between the vegetation index and radar cross-polarization backscattering coefficient. For each pixel within the effective mask of the current time phase, the deviation of the actual value of the three features from the corresponding predicted value is calculated. The negative exponent of the sum of the squares of the three deviations is used to obtain the consistency index. When the consistency index is less than the preset consistency index threshold, a physical suspicion marker is triggered. The magnitude of the deviation of each of the three features is compared, and the data source corresponding to the feature with the largest deviation is marked as a suspected data source.

[0112] Next, the validity score of the suspected data source is checked to determine the feature status and weight decay factor. If the validity score of the suspected data source is less than the preset data source blocking validity threshold, it indicates that the conventional quality assessment has confirmed that the data source has a problem, and the gating status of its corresponding feature is set to complete blocking, and it will not participate in subsequent fusion. If the validity score of the suspected data source is greater than or equal to the preset data source blocking validity threshold, it indicates that the conventional quality is qualified but the physical consistency is abnormal. It is forbidden to replace the actual observation value with the predicted value. Instead, a weight decay factor is generated and the feature is downweighted, while the physical doubt mark is retained. The data source blocking validity threshold is set according to the minimum judgment standard for the conventional quality of the data source. For example, in this embodiment, the value is 0.5.

[0113] Finally, the historical consistency integral is calculated recursively using first-order exponential smoothing. The recursive method is that the historical consistency integral of the current period equals the difference between the historical consistency integral of the previous period multiplied by one minus the forgetting factor, plus the product of the current period's consistency index multiplied by the forgetting factor. The forgetting factor is the historical consistency integral forgetting factor, which is set according to the weight requirements of time-dimensional smoothing and the sensitivity requirements of abnormal signals. For example, in this embodiment, the value is 0.3. If the historical consistency integral of three consecutive periods is less than the preset consistency index threshold, it indicates that the physical inconsistency is persistent, and the pixel is upgraded to a state awaiting ground verification warning. The physical suspicion state, suspected data source, weight decay factor, and state awaiting ground verification warning are integrated to obtain a complete set of physical suspicion markers. In this embodiment, the number of random forest decision trees, texture analysis window size, consistency index threshold, forgetting factor, and preset effective pixel number are only examples. Those skilled in the art can set them according to the actual monitoring area and data source situation. This embodiment does not limit them.

[0114] This embodiment constructs an initial feature layer covering spectral, structural, and thermal infrared dimensions through multi-source feature extraction. It establishes a priori system for the validity of multi-source data through geometric distortion discrimination, noise assessment, and multi-factor correction. Resolution matching of thermal infrared data is achieved through second-order polynomial downscaling with an effective threshold. Seasonal fluctuations are eliminated through synchronous window standardization, and consistency verification is conducted by combining cross-modal physical coupling relationships. This achieves dimensional unification and quality stratification of multi-source features, enabling continuous and reliable ecological environment monitoring results even under cloudy and rainy weather conditions. It effectively distinguishes between data anomalies and real ecological mutations, solving the technical challenge of continuous monitoring under complex weather conditions with a single remote sensing data source. It avoids the problem of erroneous correction of high-quality observation data. Feature filtering driven by information bottlenecks significantly reduces the computational complexity of multi-source data fusion, improving the accuracy, robustness, and operational efficiency of ecological environment monitoring.

[0115] Furthermore, this embodiment provides a step for obtaining a classification decision map and its corresponding confidence score based on feature vectors and their consistency markers through dual-threshold pattern determination and feature selection, including:

[0116] Based on feature vectors, the interpretation pattern label map is obtained by using a dual-threshold pattern determination of local ecological heterogeneity index and temporal change intensity, as well as a transition zone smoothing strategy.

[0117] Based on the interpreted pattern label map, representative pixels are obtained through region division and extraction. The marginal information gain of each data source within each window is obtained by calculating the confidence difference for each data source and taking the average value of the window.

[0118] Based on marginal information gain and consistency labeling, the gating mask and feature subset of each data source are obtained by window-by-window decision-making combined with first-order exponential smoothing recursion and long-term failure threshold determination.

[0119] Based on feature subsets and thermal infrared validity scores, classification decision maps and their corresponding confidence levels are obtained through basic probability assignment functions and Dempster synthesis rules.

[0120] Among them, the local ecological heterogeneity index is a quantitative indicator reflecting the degree of dispersion of the pixel neighborhood in the multidimensional feature space. It is used to characterize the fragmentation of land cover types and the strength of spatial heterogeneity within the region. It is obtained by averaging the standard deviations of each component of the standardized feature vector within the neighborhood window. The temporal change intensity is a quantitative indicator reflecting the rate of feature change of a pixel in the time dimension. It is used to characterize the magnitude of change of ecological features per unit time. It is obtained by dividing the Euclidean distance of the standardized feature vectors of adjacent time phases by the time interval. The dual-threshold mode determination is an adaptive determination mechanism for dividing the interpretation mode based on the threshold of the local ecological heterogeneity index and the temporal change intensity. It is used to adaptively switch different interpretation strategies according to the regional spatial heterogeneity and the degree of temporal change. The transition zone smoothing strategy is a processing method that uses linear weighted mixing of interpretation modes near the threshold boundary. It is used to avoid abrupt changes in results caused by mode switching and to ensure the spatial continuity of classification results. The interpretation mode label map is a spatial distribution map that marks the interpretation mode to which each pixel belongs, including three types of labels: temporal fusion mode, spatial hard classification mode, and abrupt emergency mode.

[0121] The region division involves segmenting the monitoring area into fixed grid units to support sampling evaluation and parallel computing of local windows. The monitoring area grid is a set of regular grids divided according to a fixed cell size, serving as the basic spatial unit for marginal information gain evaluation. The fixed cell size is set according to the monitoring area scale, data resolution, and computational efficiency requirements; for example, in this embodiment, it is set to 50×50 pixels. Representative pixels are typical pixels sampled from each window to characterize the overall feature distribution within the window, reducing the computational load of pixel-by-pixel calculations. The meta-window is a basic unit on the monitoring area grid with a fixed cell size. The sliding window for unit division is used for local sampling to assess marginal information gain. Its size is set according to the scale of the monitoring area and the computational efficiency requirements, ranging from 30×30 to 100×100 pixels, and usually 50×50 pixels. The preset step size is the interval distance of the sliding window, set according to the window overlap rate requirements, ranging from 20% to 80% of the window size, and usually 50% of the window size, corresponding to 50% overlap between windows. Stratified random sampling is a sampling method that randomly selects samples in each layer after stratifying by land cover type, to ensure that the sampled samples cover all land cover types, thereby improving the representativeness of the assessment results.

[0122] The representative pixel coordinate set is the set of spatial coordinates of all representative pixels extracted within each pixel window, covering sampled pixels of various land cover categories. The total number of representative pixels in a single window is set according to the window size and accuracy requirements, ranging from 30 to 100, typically 50. The full feature set random forest classifier is a random forest classification model that uses all standardized feature components as input, used to output the land cover classification results and corresponding confidence scores under full feature conditions. It is obtained through supervised training based on sample data from the monitoring area. The single-source ablation classifier is a random forest classification model that is retrained after removing features from a specific data source one by one from the full feature set. It is used to evaluate the contribution of a single data source to the classification results, obtained by removing feature components from the corresponding data source and training on the same samples. The marginal information gain is a quantitative indicator characterizing the degree to which a single data source improves the confidence score. It is calculated by the difference between the confidence score of the full feature set classifier and the confidence score of the single-source ablation classifier. A positive value indicates that the data source has a positive contribution, a negative value indicates a negative contribution, and a value close to 0 indicates that the data source is redundant.

[0123] The preset judgment threshold is the critical value for determining whether the contribution of the data source is significant. It is set according to the minimum information gain requirement of feature fusion, and the value ranges from 0.01 to 0.05, usually 0.02. The gating mask of each data source is a set of binary identifiers that mark whether the features of each data source participate in the subsequent fusion. It is 1 when allowed to pass and 0 when blocked. The first-order exponential smoothing recursion is a recursive algorithm that achieves time-dimensional smoothing by weighted accumulation of historical values ​​and current values. It is used to distinguish between single-period random fluctuations and long-term contribution trends. The historical contribution integral is a cumulative index that characterizes the long-term contribution trend of the data source at the pixel. It is obtained by accumulating the sign value of the marginal information gain over time through the first-order exponential smoothing recursion, and the value ranges from -1 to 1. The long-term failure threshold judgment is a judgment mechanism that judges whether the data source is long-term invalid based on the historical contribution integral. It is used to automatically block data sources that have no contribution or even negative contribution for a long time to improve the efficiency of subsequent calculations. The feature subset is the set of feature components that finally participate in the fusion after gating. Only the features corresponding to the data sources with significant positive contributions are retained, and weight decay processing is applied to physically questionable features.

[0124] The basic probability assignment function is a quality assignment function in Dempster-Shafer evidence theory used to describe the degree of support of evidence for a proposition. It is used to transform the posterior probabilities of classification from various data sources into a fusionable body of evidence. The Dempster synthesis rule is the core rule in Dempster-Shafer evidence theory for orthogonal synthesis of multi-source evidence. It is used to fuse classification evidence from multiple data sources to obtain a comprehensive classification confidence level. The classification decision map is a spatial distribution map obtained after classifying each pixel in the monitoring area. The categories include woodland, grassland, cultivated land, water bodies, built-up areas, and bare land. The corresponding confidence level is given by the maximum basic probability assignment quality after synthesis.

[0125] Specifically, such as Figure 4As shown, for the standardized feature vector, the local ecological heterogeneity index is calculated pixel by pixel. Within a 5x5 pixel neighborhood window of each pixel, the standard deviation of each feature component of the feature vector is calculated. The arithmetic mean of the six standard deviations is taken as the local ecological heterogeneity index of that pixel. The larger the index value, the more fragmented the vegetation types in the neighborhood and the stronger the spatial heterogeneity. Simultaneously, the temporal change intensity is calculated pixel by pixel. The Euclidean distance between the standardized feature vectors of the current time phase and the previous time phase is calculated. This distance value is divided by the time interval between the two observations to obtain the feature change amplitude per unit time, i.e., the temporal change intensity. The larger the value, the more drastic the ecological changes that have occurred in that pixel per unit time. Then, the thresholds for local ecological heterogeneity and temporal variation intensity are determined. The local ecological heterogeneity threshold is determined using the cumulative frequency method. The cumulative frequency distribution of the local ecological heterogeneity index of all pixels in the entire region is taken, and the 80th percentile is used as the baseline threshold. At the same time, the 80th percentiles of the four types of land cover—forest, grassland, cultivated land, and built-up area—are calculated separately. The weighted average of the unified threshold for the entire region and the thresholds for each category is taken as the final threshold to adapt to the inherent heterogeneity differences of different land cover. The threshold for temporal variation intensity is determined based on historical time series data. Historical observation data covering at least two complete years are used to calculate the distribution of the temporal variation intensity of each pixel in the same seasonal window of the same historical period. The 95th percentile is used as the threshold for that pixel. This threshold is automatically updated with the seasons to adapt to phenological fluctuations.

[0126] Then, based on the above thresholds, the interpretation mode is determined. When the local ecological heterogeneity index is less than the local ecological heterogeneity threshold and the temporal change intensity is less than the temporal change intensity threshold, it is determined to be a temporal fusion mode, and the standard multi-source temporal fusion interpretation is performed. When the local ecological heterogeneity index is greater than or equal to the local ecological heterogeneity threshold and the temporal change intensity is less than the temporal change intensity threshold, it is determined to be a spatial hard classification mode, and the temporal fusion logic is abandoned, and the current independent hard classification is used instead. When the temporal change intensity is greater than or equal to the temporal change intensity threshold, regardless of the value of the local ecological heterogeneity index, it is determined to be a sudden emergency mode, triggering the ground or UAV emergency observation process. Then, transition zone smoothing is performed. When the local ecological heterogeneity index is within the transition zone range of ±10% of the local ecological heterogeneity threshold, both the temporal fusion mode and the spatial hard classification mode are enabled simultaneously. A linear weighting method is used to achieve a smooth mixture of the two modes. Within the transition zone, the temporal fusion mode weight is dominant near the lower limit of the local ecological heterogeneity threshold, while the spatial hard classification mode weight is dominant near the upper limit of the local ecological heterogeneity threshold. When the classification inconsistency rate of the two modes exceeds 5%, weighted smoothing is officially enabled. The weights change linearly with the local ecological heterogeneity index of the pixel within the transition zone, ultimately obtaining a continuous interpretation mode label map for the entire domain.

[0127] Then, the monitoring area is divided into a grid of 50×50 pixels, which are then slid across the grid at a preset step size to form a grid system that fully covers the monitoring area. Next, stratified random sampling is performed to obtain a set of representative pixel coordinates. Within each pixel window, based on the land use cover classification map, random sampling is performed according to six land cover categories: forest, grassland, cultivated land, water bodies, built-up areas, and bare land. The number of samples for each land cover category is allocated according to the proportion of each category within the window, ensuring that a total of 50 representative pixels are extracted within a single window, covering all land cover types. The coordinates of the representative pixels within each window are then aggregated to obtain the set of coordinates. Then, classifiers are constructed and marginal information gain is calculated. For each representative pixel, a full-feature-set random forest classifier and a single-source ablation classifier are constructed. The full-feature-set random forest classifier uses all components of the standardized feature vector as input features and outputs the classification result and the corresponding maximum class probability as the confidence score. The single-source ablation classifier is a classification model obtained by removing all feature components corresponding to a certain data source one by one and retraining it with the same training parameters and training samples as the full-feature-set classifier. It is used to simulate the classification effect when the data source is missing. The difference between the confidence score of the full-feature-set classifier and the confidence score of the single-source ablation classifier is calculated for each data source. This difference is the marginal information gain of the data source at a single representative pixel. The arithmetic mean of the marginal information gain of each representative pixel in the window is taken to obtain the marginal information gain of the corresponding data source in the window. This window-level gain value is applied to the feature removal decision of all pixels in the window.

[0128] Then, gating decisions are performed based on marginal information gain combined with consistency marking. Gating masks and feature subsets for each data source are obtained through time-dimensional accumulation and long-term failure determination. First, gating decisions are performed window-by-window to obtain the gating mask for each data source. The marginal information gain of each data source within a window is compared with a preset judgment threshold. When the marginal information gain is greater than the judgment threshold, the data source is determined to have a significant positive contribution, and its features are retained for fusion, with the corresponding gating mask set to 1. When the absolute value of the marginal information gain is less than or equal to the judgment threshold, the data source is determined to be redundant and has no significant contribution, and its features are removed, with the corresponding gating mask set to 0. When the marginal information gain is less than a negative judgment threshold, the data source is determined to have a significant negative contribution, and its features are forcibly blocked, with the corresponding gating mask set to 0. For feature channels previously marked as physically questionable by physical consistency verification and with a weight decay factor, if their marginal information gain is greater than the judgment threshold, the feature is retained for fusion, but its weight is halved using the weight decay factor; if their marginal information gain is less than or equal to the judgment threshold, the feature is directly removed.

[0129] The sign attribute of marginal information gain is accumulated over time. The historical contribution integral of each data source at each cell is defined and calculated using a first-order exponential smoothing recursive method. The current historical contribution integral is equal to the difference between the historical contribution integral of the previous period and the historical contribution forgetting factor, plus the product of the sign value of the current marginal information gain and the historical contribution forgetting factor. The sign value of the marginal information gain is 1 when it is greater than the judgment threshold, 0 when the absolute value is less than or equal to the judgment threshold, and -1 when it is less than the negative judgment threshold. The historical contribution forgetting factor is set according to the requirements of time smoothing intensity and anomaly sensitivity, and the value range is 0.2 to 0.5. For example, in this embodiment, the value is 0.3. The historical contribution integral ranges from -1 to 1. The higher the value, the more stable the long-term positive contribution of the data source. When the historical contribution integral is less than the preset long-term failure threshold, the pixel is marked as having a long-term invalid state for the corresponding data source. In subsequent time phases, the gating mask of this data source is always set to 0, and it no longer participates in gain evaluation. The long-term failure threshold is set according to the strictness of the long-term invalidity determination, and its value ranges from -0.6 to -0.4. For example, in this embodiment, the value is -0.5. Combining the current time phase gating mask and the long-term blocking state, a subset of features that will ultimately participate in the fusion is selected from the standardized feature vector. Physically questionable features in the feature subset are weighted down by a weight attenuation factor, while the remaining normal features retain their original weights.

[0130] For the three types of data sources—optical, radar, and thermal infrared—the feature subsets ultimately used for fusion are input into the corresponding random forest classifiers of the data sources to obtain the posterior probability of a pixel belonging to each land cover category. A basic probability assignment quality function is constructed based on the classification posterior probability and the effectiveness score of the corresponding data source. For each land cover category, the difference between the posterior probability output by the data source classifier and the uncertainty coefficient is used as the basic probability assignment quality for that category. The uncertainty coefficient itself is used as the global uncertainty quality assigned to the entire category recognition framework. The uncertainty coefficient is equal to one minus the effectiveness score of the data source; a higher effectiveness score results in lower uncertainty, higher quality assigned to a specific category, and lower uncertainty quality assigned to the global framework.

[0131] The Dempster synthesis rule is used to perform an orthogonal sum operation on the basic probability assignment quality functions of the three data sources. For any land cover category in the recognition framework, the synthesized basic probability assignment quality is equal to the sum of the quality products of the combinations whose intersection with the basic probability assignments of all data sources is that category, divided by the difference of the conflict coefficient for normalization. The conflict coefficient is equal to the sum of the quality products of the combinations whose intersection with the basic probability assignments of all data sources is an empty set, and is used to reflect the degree of conflict between the evidence from each data source. If the conflict coefficient is greater than a preset conflict threshold, it is judged as a serious conflict of evidence, and a low confidence mark is output for that pixel. The conflict threshold is set according to the multi-source evidence conflict tolerance and classification reliability requirements. For example, in this embodiment, the value is 0.9. After synthesis, the land cover category with the highest basic probability assignment quality is selected as the final classification result for that pixel. This maximum quality value is the corresponding confidence of the classification decision map. After traversing the pixel by pixel, the classification decision map of the whole area and its corresponding confidence layer for each pixel are obtained. In this embodiment, the element window size, sliding step size, number of representative pixels, preset judgment threshold, historical contribution forgetting factor, and long-term failure threshold are merely examples. Those skilled in the art can set these parameters according to the actual monitoring area scale and computing performance requirements. This embodiment does not impose any restrictions on these parameters.

[0132] This embodiment achieves spatial differentiation adaptation and smooth transition of the interpretation strategy through dual-threshold adaptive mode determination, realizes low-computation marginal information gain quantification through local window hierarchical sampling and single-source ablation evaluation, realizes dynamic minimum sufficient feature screening through window-by-window gating decision combined with historical contribution accumulation in the time dimension, and realizes uncertainty synthesis of multi-source classification evidence through Dempster-Shafer evidence theory fusion. It significantly reduces the computational complexity of multi-source fusion while ensuring interpretation accuracy, effectively eliminates redundant and negative contribution data source features, improves the robustness and confidence of classification results, and takes into account the fusion accuracy of homogeneous regions and the interpretation adaptability of fragmented and abrupt regions.

[0133] Furthermore, this embodiment provides a step for obtaining an environmental parameter set based on a classification decision graph and its corresponding confidence level, using a CASA model and a quantitative inversion model, including:

[0134] Based on the classification decision graph and its corresponding confidence level, the initial inversion parameters are obtained through the CASA model.

[0135] Based on the initial inversion parameters and validation data, an accuracy evaluation report is obtained by calculating quantitative indicators.

[0136] Based on the accuracy evaluation report, an environmental parameter set is obtained by hierarchical labeling and generating a confidence layer.

[0137] The meteorological data consists of ground observation and reanalysis datasets describing the meteorological conditions of the monitoring area. These datasets provide driving data for the inversion of ecological parameters and are obtained through observations at ground meteorological stations or publicly available meteorological reanalysis products. They include elements such as total solar radiation, average monthly temperature, monthly precipitation, and potential evapotranspiration. The CASA model is a classic mechanistic model for estimating net primary productivity of vegetation based on the principle of light energy utilization. It is used to calculate the net primary productivity parameter of vegetation and is constructed based on the product relationship between the photosynthetically active radiation absorption ratio and the actual light energy utilization rate. The net primary productivity parameter is an ecological parameter characterizing the amount of net organic matter accumulated by vegetation per unit time and per unit area. It includes the spatial distribution value of net primary productivity per pixel, usually expressed in grams of carbon per square meter per year. The feature space is a two-dimensional scatter distribution space constructed with the first-resolution vegetation index as the horizontal axis and the first-resolution land surface temperature as the vertical axis. It is used to characterize the coupling relationship between vegetation cover and land surface temperature and is obtained based on the statistical analysis of the corresponding feature values ​​of all pixels in the region.

[0138] The Temperature-Vegetation Drought Index (TVI) is a quantitative indicator of soil drought severity derived from the inversion of vegetation indices and surface temperature feature space. It characterizes the relative level of soil moisture content and is calculated through the linear relationship between dry and wet edges in the feature space. Its value ranges from 0 to 1, with higher values ​​indicating more severe soil drought. The first-resolution soil moisture distribution map is a raster layer representing the spatial distribution of relative soil moisture content at a 10-meter spatial resolution, calculated pixel-by-pixel from the TVI. The empirical algorithm inversion involves establishing empirical regression relationships based on measured samples to quantitatively estimate parameters. The method is used to invert chlorophyll concentration in water bodies. An estimation model is constructed by regression fitting of measured chlorophyll concentration data and remote sensing reflectance during the same period. Chlorophyll concentration is a core water quality parameter characterizing the degree of eutrophication of water bodies. The unit is usually micrograms per liter. It is obtained by inverting the ratio of red light to near-infrared surface reflectance of water body pixels through an empirical model. The built-up area mask is a binary layer that marks the distribution range of impermeable surfaces such as urban buildings and roads. It is extracted from the built-up area category in the land use cover classification map. The built-up area pixel is set to 1, and the non-built-up area pixel is set to 0.

[0139] The suburban mask is a binary layer that marks the distribution range of the reference area outside the built-up area. It is obtained by extending the built-up area mask outward by a preset suburban buffer zone width and then removing the built-up area portion. It is used as a reference background for calculating the intensity of the urban heat island. The preset suburban buffer zone width is set according to the spatial influence range of the urban heat island and the representativeness requirements of the suburban reference. For example, in this embodiment, the value is 5 to 10 kilometers. The water ecological parameters are a set of parameters characterizing the ecological environment quality of the water body, including the statistical value of water body area, the spatial distribution of chlorophyll concentration, and the eutrophication level evaluation results. The thermal environment parameters are a set of parameters characterizing the regional thermal environment status, including the spatial distribution of surface temperature at the first resolution and the numerical value of the urban heat island intensity. The initial inversion parameters are a set of preliminary inversion results of various ecological environment parameters that have not been verified for accuracy, including net primary productivity parameters, the first resolution soil moisture content distribution map, chlorophyll concentration, water ecological parameters, and thermal environment parameters.

[0140] The validation data consists of ground and low-altitude observation ground-based data used to verify the accuracy of the remote sensing inversion results. These data were acquired through ground monitoring station observations and high-resolution aerial surveys by UAVs, including measured values ​​of land cover categories, vegetation parameters, surface temperature, and chlorophyll concentration in water bodies at ground sampling points. The quantitative indicators are a set of statistical indicators used to quantify the accuracy of the inversion results, including classification accuracy indicators and quantitative inversion accuracy indicators, corresponding to the accuracy evaluation of land cover classification results and parameter inversion results, respectively. The accuracy evaluation report is an analytical document summarizing the calculation results and compliance status of various accuracy indicators, including classification accuracy evaluation results, inversion accuracy evaluation results for each quantitative parameter, and a comparative analysis of the inversion accuracy of multi-source collaborative schemes and single-source data.

[0141] The grading and labeling process is used to classify the quality level of the output product based on the accuracy evaluation results. It is used to distinguish the inversion results with different levels of confidence, including two quality levels: standard level and reference level. The confidence level layer is a raster layer that marks the confidence level of each pixel's inversion results. It is used to visually display low confidence areas and physically questionable areas, providing a clear confidence reference for the use of the results. The environmental parameter set is the final set of ecological and environmental monitoring results after accuracy verification and grading and labeling. It includes spatial distribution maps of various ecological and environmental parameters, corresponding accuracy evaluation reports, and time series analysis results.

[0142] Specifically, such as Figure 5 As shown, firstly, based on the classification decision map and its corresponding confidence level, combined with the initial feature layer and meteorological data, the initial inversion parameters are calculated through a combination of mechanistic model and empirical inversion. The CASA model is used for pixel-by-pixel calculation. The calculation logic of this model is that net primary productivity equals the product of absorbed photosynthetically active radiation and actual light energy utilization rate, and the corresponding calculation formula is: ,in for Phase Net primary productivity of a pixel In order to absorb photosynthetically active radiation, The actual light energy utilization rate is calculated as follows: Based on the first resolution normalized vegetation index (NDI), the photosynthetically active radiation (PANR) absorption ratio is estimated using the empirical relationship between the NDI and the PANR. A higher NDI corresponds to a higher proportion of PANR absorbed by the vegetation. The upper limit of this value is constrained by a preset upper limit for PANR absorption, which is set based on the physiological characteristics of the maximum light energy absorption of the vegetation canopy. For example, in this embodiment, the value is 0.95. The PANR is obtained by multiplying the observed total solar radiation from meteorological data by a preset PANR ratio coefficient. This coefficient is set based on the energy proportion range of the photosynthetically active band in the solar spectrum. For example, in this embodiment, the value is 0.45. The absorbed PANR is obtained by multiplying the PANR by the PANR absorption ratio. The fourth step determines the preset maximum light energy utilization rate, based on classification decisions. The land use cover type in the map determines the preset maximum light energy utilization rate corresponding to different vegetation types, with forests, grasslands, and cultivated land corresponding to different physiological parameter values. Then, the environmental stress coefficient is calculated, which includes the temperature stress coefficient and the water stress coefficient. The temperature stress coefficient reflects the degree of inhibition of photosynthesis by temperature deviation from the vegetation's optimum temperature, and is calculated based on the relationship between monthly average temperature and the vegetation's optimum temperature. The water stress coefficient reflects the degree of inhibition of photosynthesis by water deficit, and is calculated based on the relationship between monthly precipitation and potential evapotranspiration or based on the characteristic space of surface temperature and normalized vegetation index. The environmental stress coefficient ranges from 0 to 1, with values ​​closer to 1 indicating a weaker stress level. The actual light energy utilization rate is calculated by multiplying the maximum light energy utilization rate by the temperature stress coefficient and then by the water stress coefficient. The first-resolution net primary productivity spatial distribution map is then calculated pixel by pixel.

[0143] A two-dimensional feature space is constructed based on the first-resolution land surface temperature and the first-resolution normalized vegetation index (NDVI). All pixels in the feature space are distributed in a triangular or trapezoidal shape. For each NDVI value range, the maximum value of the land surface temperature within that range is determined as the dry edge corresponding value, and the minimum value as the wet edge corresponding value. Dry edge equations and wet edge equations are then fitted to each range. For any pixel, its Temperature-Vegetation Aridity Index (TVDI) is equal to the current land surface temperature minus the wet edge temperature corresponding to the NDVI, then divided by the difference between the dry edge temperature and the wet edge temperature corresponding to the NDVI. The calculation formula is as follows:

[0144]

[0145] in The first-resolution surface temperature. and These represent the dry edge temperature and wet edge temperature under the corresponding normalized vegetation index; the temperature vegetation drought index ranges from 0 to 1, with the closer the value is to 1 indicating drier soil and the closer the value is to 0 indicating wetter soil. The first resolution soil moisture distribution map is obtained after pixel-by-pixel calculation.

[0146] All water body pixels were extracted based on the first-resolution normalized water body index. Chlorophyll concentration was retrieved from the water body pixels using an empirical algorithm. The empirical relationship between the ratio of surface reflectance in the red band and the near-infrared band and chlorophyll concentration was used for calculation. The coefficients of the empirical relationship were obtained by regression fitting of measured chlorophyll concentration data from typical water body sampling points in the monitoring area with the remote sensing reflectance of the same period. The fitted empirical model was applied to all water body pixels to obtain a spatial distribution map of chlorophyll concentration.

[0147] Based on the classification decision map, built-up area pixels are extracted. After extending the built-up area mask outward to the width of the pre-defined suburban buffer zone, the built-up area portion is removed to obtain the suburban mask. The mean values ​​of all first-resolution surface temperature pixels within both the built-up area mask and the suburban mask are calculated. The urban heat island intensity is obtained by subtracting the suburban mean value from the built-up area mean value. Combined with the water body area and chlorophyll concentration distribution results obtained from water body pixel statistics, aquatic ecological parameters are constructed. Combined with surface temperature distribution and urban heat island intensity, thermal environment parameters are constructed. By summarizing all the above parameter results, complete initial inversion parameters are obtained.

[0148] Then, based on the initial inversion parameters and verification data, an accuracy evaluation report is calculated using multi-dimensional quantitative indicators. First, verification sample points and areas are established, employing a combination of random and stratified sampling to cover different land cover types such as forest, grassland, water bodies, and built-up areas. The number of verification sample areas for each land cover type is no less than the preset minimum number of verification sample areas for a single land cover type. This number is set according to the spatial representativeness and statistical significance requirements of the land cover type; for example, in this embodiment, it is 3. The area of ​​a single verification sample area is no less than the preset minimum area of ​​a single verification sample area. This area is set according to the spatial scale matching and result representativeness requirements; for example, in this embodiment, it is 1 square kilometer. The total number of verification sample points within the monitoring area is no less than the preset minimum total number of verification sample points. This total number is set according to the statistical reliability requirements of the accuracy evaluation; for example, in this embodiment, it is 100. Ground monitoring station measured data and UAV high-resolution image data are used as the verification true values.

[0149] For the land use cover classification results, a confusion matrix was constructed using the measured land cover categories of the validation samples as the true values. Two indicators were calculated: overall accuracy and Kappa coefficient. Overall accuracy is the percentage of correctly classified pixels out of the total number of validation pixels, while the Kappa coefficient measures the difference between the observed consistency rate and the expected consistency rate, comprehensively reflecting the consistency level of the classification results. Subsequently, a quantitative inversion accuracy evaluation was conducted, calculating corresponding accuracy indicators for vegetation parameters, temperature parameters, and productivity parameters. The quantitative indicators used included root mean square error (RMSE), coefficient of determination (CCD), and mean absolute percentage error (MAE). RMSE reflects the average deviation between the inverted and measured values, CCD reflects the linear correlation between the inverted and measured values, and MAE reflects the relative error of the inverted values.

[0150] The pass / fail criteria for each parameter correspond to preset accuracy thresholds. Vegetation parameters include the Normalized Difference Vegetation Index (NDVI) and vegetation cover. The preset root mean square error (RMSE) threshold for vegetation parameters is set according to the industry-standard accuracy inversion of vegetation parameters; for example, in this embodiment, it is set to 0.1. The preset coefficient of determination (COD) threshold for vegetation parameters is also set according to the industry-standard accuracy inversion of vegetation parameters; for example, in this embodiment, it is set to 0.8. Temperature parameters, i.e., surface temperature, have a preset RMSE threshold set according to the industry-standard accuracy inversion of surface temperature; for example, in this embodiment, it is set to 2.5℃. The COD threshold for temperature parameters is also set according to the industry-standard accuracy inversion of surface temperature; for example, in this embodiment, it is set to 0.8. Productivity parameters, i.e., net primary productivity, have a preset relative error threshold set according to the industry-standard accuracy inversion of net primary productivity of vegetation; for example, in this embodiment, it is set to 0.2. The COD threshold for productivity parameters is also set according to the industry-standard accuracy inversion of net primary productivity of vegetation; for example, in this embodiment, it is set to 0.75. Then, the synergistic effect evaluation is carried out. The inversion accuracy is calculated using only optical data, only radar data, and the multi-source synergistic scheme of this method. The accuracy improvement of this method compared with single-source data is obtained by comparison. The calculation results and compliance status of all the above indicators are summarized to form a complete accuracy evaluation report.

[0151] Based on the accuracy evaluation report, a tiered labeling strategy and a corresponding confidence level layer are used to obtain the final set of environmental parameters. Products that meet all preset qualification thresholds for all accuracy indicators are labeled as standard-level products and can be directly used as formal monitoring results. Products that fail to meet some indicators due to the complete failure of data sources in a specific area, such as areas where continuous rain results in optical and thermal infrared validity scores of 0, are labeled as reference-level products and require additional confidence explanations for reference. Then, a confidence level labeling layer is generated, overlaying low-confidence area labels onto the thematic map of the reference-level products. Simultaneously, areas with questionable physical consistency from previous physical consistency checks and warning points awaiting ground verification are overlaid, visually demonstrating the reliability of the results for each area. A complete set of environmental parameters is compiled, including a ten-meter-level thematic atlas of ecological and environmental parameters and a time-series analysis report. The thematic atlas includes spatial distribution maps of vegetation cover, net primary productivity, land use cover classification, water body distribution and chlorophyll concentration, surface temperature and urban heat island intensity, and soil moisture content. All thematic maps are accompanied by corresponding confidence level marker layers. The time-series analysis report includes monthly, seasonal, and annual temporal variation curves for each ecological and environmental parameter, spatiotemporal location results of key ecological events, and comprehensive assessment conclusions of ecological and environmental quality. The preset photosynthetically active radiation ratio coefficient, preset upper limit of photosynthetically active radiation absorption ratio, preset suburban buffer zone width, preset accuracy qualification thresholds, and preset number of verification sample points and sample areas in this embodiment are merely examples. Those skilled in the art can set these parameters according to the actual monitoring area scale and operational needs; this embodiment does not impose any limitations on these settings.

[0152] This embodiment achieves quantitative output of all-dimensional ecological and environmental parameters by combining the CASA mechanism model with multi-index empirical inversion. It constructs a multi-dimensional accuracy evaluation system using multi-source verification data combining ground monitoring and UAV aerial surveys. The hierarchical labeling strategy ensures the continuity of monitoring operations in extreme weather and data-missing scenarios. The overlay of confidence layers and warning points enhances the interpretability and practicality of the results. It achieves accurate inversion of ecological and environmental parameters at a high resolution of ten meters, realizing the visualization, quantification, and operational output of ecological and environmental monitoring results, and improving the overall accuracy, robustness, and operational adaptability of ecological and environmental monitoring results.

[0153] In summary, this invention constructs a highly reliable data foundation through standardized preprocessing of multi-source remote sensing data and a two-layer quality verification mechanism. It achieves dynamic minimum full fusion through dual-threshold adaptive interpretation mode switching and feature selection driven by marginal information gain. By combining mechanistic models with empirical inversion, it outputs multi-dimensional ecological and environmental parameters. This effectively solves the technical problem of continuous monitoring under complex weather conditions with a single remote sensing data source, avoids the defect of misjudging real ecological mutation signals as noise, reduces the computational complexity of multi-source fusion, and improves the accuracy, robustness, and operational efficiency of ecological and environmental monitoring. It provides stable and reliable technical support for routine regional ecological and environmental monitoring and early warning of sudden ecological events.

[0154] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0155] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0156] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0157] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An ecological environment monitoring method based on multi-source remote sensing collaborative interpretation, characterized in that, include: Acquire raw image data and obtain a multi-source remote sensing dataset and a basic mask layer set by using quality screening rules; Based on the multi-source remote sensing dataset, feature vectors and their consistency labels are obtained by feature extraction and construction of a sensor effectiveness prior map. Based on the feature vector and its consistency marker, a classification decision map and its corresponding confidence level are obtained through dual threshold pattern determination and feature filtering. Based on the classification decision map and its corresponding confidence level, the environmental parameter set is obtained through the CASA model and the quantitative inversion model.

2. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 1, characterized in that, The multi-source remote sensing dataset and basic mask layer set were obtained through quality screening rules, including: Based on the original image data, a multi-source registered image set is obtained through intensity cross-correlation registration algorithm and multi-view processing; Based on the multi-source registered image set, the basic data layer is obtained through radiometric calibration conversion and atmospheric correction of the radiative transfer model. Based on the aforementioned basic data layer, a multi-source remote sensing dataset and a basic mask layer set are obtained by clipping the monitoring area boundaries and classifying ground features using prior knowledge.

3. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 1, characterized in that, Based on the aforementioned multi-source remote sensing dataset, a sensor effectiveness prior map is constructed through feature extraction. The feature vectors and their consistency markers are obtained, including: Based on the multi-source remote sensing dataset, an initial feature layer is obtained through normalized difference operation, gray-level co-occurrence matrix texture analysis, and split window algorithm; Based on the multi-source remote sensing dataset and digital elevation model, thermal infrared effectiveness score is obtained through attenuation calculation, local incident angle geometric discrimination, and cloud cover and emissivity correction. Based on the initial feature layer and thermal infrared effectiveness score, the first resolution surface temperature is obtained by establishing a second-order polynomial regression relationship and performing inverse distance weighted interpolation on the regression residuals. Based on the first resolution surface temperature and the initial feature layer, a feature vector is obtained through synchronous window standardization, and a consistency label is obtained according to the coupling relationship.

4. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 3, characterized in that, Based on the aforementioned multi-source remote sensing dataset, an initial feature layer is obtained through normalized difference calculation, gray-level co-occurrence matrix texture analysis, and split-window algorithm, including: Based on the multi-source remote sensing dataset, the first resolution vegetation index, the first resolution water index, and the land use cover classification map are obtained by normalized difference operation and random forest classifier. Based on the multi-source remote sensing dataset, the first resolution structural feature layer is obtained by extracting information from dual polarization channels and analyzing gray-level co-occurrence matrix texture. Based on the multi-source remote sensing dataset, the second-resolution land surface temperature and urban heat island intensity are obtained through the split-window algorithm and temperature difference calculation.

5. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 4, characterized in that, Based on the aforementioned multi-source remote sensing dataset and digital elevation model, a thermal infrared effectiveness score is obtained through attenuation calculation, local incident angle geometric discrimination, and cloud cover and emissivity correction, including: The local incident angle and the radar system incident angle are calculated based on the digital elevation model to obtain the radar shadow binary identifier; Based on the aforementioned binary radar shadow identifier, the overlapping intensity is obtained by calculating the ratio of the incident angle difference to the incident angle of the radar system and limiting it to the interval between zero and one. Based on the mean and variance estimates of radar strength data or the equivalent number of views read from metadata, the speckle noise coefficient is obtained by calculating the reciprocal of the square root of the equivalent number of views; Based on the overlay strength and speckle noise coefficient, the thermal infrared effectiveness score is obtained by directly setting the shadow area to zero and multiplying the overlay strength and speckle noise coefficient by subtracting the corresponding coefficient from each other.

6. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 5, characterized in that, Based on the initial feature layer and thermal infrared effectiveness score, a first-resolution surface temperature is obtained by establishing a second-order polynomial regression relationship and performing inverse distance-weighted interpolation on the regression residuals, including: Based on the initial feature layer, the regression coefficients and the regression residual values ​​of each coarse pixel are obtained through second-order polynomial fitting and regression residual calculation. Based on the regression coefficients, the preliminary first-resolution surface temperature is calculated pixel-by-pixel using polynomials. Based on the regression residual values ​​of each coarse pixel, the thermal infrared effectiveness score, and the preliminary first-resolution land surface temperature, areas with a preset number of effective pixels are selected, and the regression residual values ​​are transferred through inverse distance weighted interpolation to obtain the first-resolution land surface temperature.

7. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 6, characterized in that, Based on the first-resolution surface temperature and the initial feature layer, a feature vector is obtained through synchronous window normalization, and a consistency label is obtained according to the coupling relationship, including: Vegetation cover is calculated based on the first-resolution vegetation index, and effective object masks are obtained by filtering pixels in combination with the land use cover classification map. Based on the effective object mask and the initial feature layer, a consistency index is obtained, and pixels with a consistency index less than a preset consistency index threshold are physically suspected and the suspected data source is determined. Based on the physical suspicious markers and suspected data sources, the validity scores of the corresponding data sources are checked back to obtain the feature status and weight decay factor. The set of physically questionable tags is obtained by calculating the historical consistency integral through first-order exponential smoothing recursion.

8. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 1, characterized in that, Based on the feature vector and its consistency marker, a classification decision map and its corresponding confidence score are obtained through dual-threshold pattern determination and feature selection, including: Based on the feature vector, the interpretation pattern label map is obtained by using a dual threshold pattern determination of local ecological heterogeneity index and temporal change intensity, and a transition zone smoothing strategy. Based on the interpretation pattern label map, representative pixels are obtained by dividing the region and extracting them. The marginal information gain of each data source within each window is obtained by calculating the confidence difference for each data source and taking the average value of the window. Based on the marginal information gain combined with the consistency label, the gating mask and feature subset of each data source are obtained by window-by-window decision-making combined with first-order exponential smoothing recursion and long-term failure threshold determination. Based on the aforementioned feature subset and thermal infrared validity score, a classification decision map and its corresponding confidence level are obtained through the basic probability allocation function and Dempster synthesis rule.

9. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 8, characterized in that, Based on the interpreted pattern label map, representative pixels are obtained through region division and extraction. The marginal information gain of each data source within each window is obtained by calculating the confidence difference for each data source and taking the average value within a window, including: Based on the interpretation pattern label map, a monitoring area grid is obtained through area division; Based on the monitoring area grid and land use cover classification map, the representative pixel coordinate set in each window is obtained by sliding the meta window at a preset step size and performing stratified random sampling within each window. Based on the representative pixel coordinate set and feature vector, a full feature set random forest classifier and a single-source ablation classifier are constructed. By calculating the difference in output confidence before and after removal and taking the average value within the window, the marginal information gain of each data source is obtained.

10. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 9, characterized in that, Based on the marginal information gain combined with the consistency label, and through window-by-window decision-making combined with first-order exponential smoothing recursion and long-term failure threshold determination, the gating mask and feature subset of each data source are obtained, including: Based on the marginal information gain and the preset judgment threshold, the physically questionable features are reduced in weight by window-by-window decision-making and combined with the weight decay factor to obtain the gating mask of each data source. Based on the gating masks of each data source, the historical contribution integral is calculated by first-order exponential smoothing recursion. Based on the historical contribution integral, a feature subset is obtained by determining the long-term failure threshold.

11. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 1, characterized in that, Based on the classification decision map and its corresponding confidence level, an environmental parameter set is obtained through the CASA model and the quantitative inversion model, including: Based on the classification decision graph and its corresponding confidence level, the initial inversion parameters are obtained through the CASA model; Based on the initial inversion parameters and verification data, an accuracy evaluation report is obtained by calculating quantitative indicators. Based on the accuracy evaluation report, an environmental parameter set is obtained by hierarchical labeling and generating a confidence layer.

12. The ecological environment monitoring method based on multi-source remote sensing collaborative interpretation according to claim 11, characterized in that, Based on the classification decision graph and its corresponding confidence level, the initial inversion parameters are obtained through the CASA model, including: Based on the initial feature layer and meteorological data, the net primary productivity parameters are calculated pixel by pixel using the CASA model. A feature space is constructed based on the first-resolution land surface temperature and the first-resolution vegetation index. The first-resolution soil moisture content distribution map is obtained by calculating the temperature-vegetation-drought index. Chlorophyll concentration is obtained by inversion based on the first-resolution water body index and through empirical algorithms. Based on the classification decision map and its corresponding confidence level, the water ecological parameters and thermal environment parameters are obtained by calculating the mean difference of the first-resolution surface temperature within the built-up area mask and the suburban area mask.

Images