Multi-source remote sensing fusion forest and grass wet resource dynamic identification classification method and system

Abstract

This invention discloses a method and system for dynamic identification and classification of forest, grassland, and wetland resources based on multi-source remote sensing fusion, belonging to the field of resource identification and classification technology. It includes: S1, collecting multi-source remote sensing observation data, performing data preprocessing, and conducting observation quality assessment; S2, performing object segmentation and feature aggregation to form an object-level three-dimensional feature set; S3, acquiring the cross-time object-level three-dimensional feature set, extracting pseudo-change driving features, performing pseudo-change assessment, and determining the consistency of true changes, combining the results of pseudo-change assessment and true change consistency determination to distinguish change types; S4, performing supervised training and hierarchical classification based on the object-level three-dimensional feature set, outputting the probability of forest, grassland, and wetland categories, and combining change features for rule recognition. This solves the problems in multi-source remote sensing cross-time comparisons where non-real fluctuations are difficult to distinguish, leading to confusion between pseudo-changes and real land cover transfers, high false alarm and false false alarm rates, and insufficient result stability.

Description

Technical Field

[0001] This invention relates to the field of resource identification and classification technology, specifically to a method and system for dynamic identification and classification of forest, grassland and wetland resources based on multi-source remote sensing fusion. Background Technology

[0002] With the rapid development of multi-source remote sensing technologies such as high-resolution optical remote sensing, synthetic aperture radar, and lidar, the dynamic monitoring of ecological resources such as forests, grasslands, and wetlands is gradually evolving from single-sensor to multi-source fusion. By fusing multispectral reflectance information, backscattering characteristics, and three-dimensional structural information, land cover status and structural attributes can be obtained at larger spatial scales and higher temporal frequencies. This enables resource type identification, change monitoring, and ecological status assessment, providing data support for natural resource management, ecological protection and restoration, and carbon sequestration.

[0003] For example, invention patent CN120655997A discloses a mineral resource image classification method based on VGG, including an image preprocessing stage: the remote sensing image is segmented into small blocks according to the coordinates of the mineral deposits and the classification results, and labeled; then image upsampling, sharpening, and filtering are performed. A model building stage is also included: a VGG-based classification model is designed, and the processed images are divided into training and testing sets. The training set is used to train the model, and the testing set is used to evaluate the model's performance. The model parameters are then adjusted to find the optimal parameter combination. Finally, a mineral deposit prediction stage is performed: the trained model is containerized and deployed, the predicted image is input, and the output result is obtained. This invention uses the VGG algorithm and has achieved good results in mineral resource image classification using real geological data, showing broad application value and promising prospects in the field of geology and mineral resources.

[0004] For example, invention patent CN120014498A discloses a method for rapid sampling and verification analysis of land use category classification, including the following steps: S1, acquiring a vertical aerial view of land use categories taken by a drone; S2, dividing the land use category outlines in the vertical aerial view of land use categories obtained in step S1; S3, labeling the land use categories in the vertical aerial view of land use categories obtained in step S2; S4, classifying the labeled land use categories in the vertical aerial view of land use categories obtained in step S3 for rapid sampling and verification. This invention can quickly verify and identify land use categories based on the sampling map, improving efficiency and accuracy.

[0005] However, in forest and grassland wetland dynamics monitoring, a common practice is to register and classify remote sensing images from different sensors or different time phases, and then directly perform cross-temporal difference comparisons to determine changes. However, when there are significant differences in phenological stages, short-term water level fluctuations, solar altitude angles, and observational geometry, the same land cover unit may exhibit non-changing fluctuations in spectral reflectance and texture or structural characterization. These fluctuations can be confused with actual land cover transfer or degradation signals during difference comparisons, making it difficult to distinguish "pseudo-changes." This leads to an increase in false alarms and false negatives in change detection and a decrease in the comparability of cross-temporal results.

[0006] Therefore, in order to address the above problems, there is an urgent need for a dynamic identification and classification method and system for forest, grassland and wetland resources based on multi-source remote sensing fusion. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a method and system for dynamic identification and classification of forest, grassland and wetland resources based on multi-source remote sensing fusion. This solves the problems of difficulty in distinguishing between non-real spectral and structural fluctuations caused by phenological differences, water level fluctuations and observed geometric changes in multi-source remote sensing over time, leading to confusion between pseudo-changes and real land cover transfers, high false alarm and false negative rates in change detection, and insufficient stability and comparability of results.

[0009] Technical solution

[0010] To achieve the above objectives, this invention provides the following technical solution: a dynamic identification and classification method for forest, grassland, and wetland resources based on multi-source remote sensing fusion, comprising the following steps: S1, collecting multi-source remote sensing observation data and preprocessing the data; performing observation quality assessment based on the preprocessed multi-source remote sensing observation data to construct standardized observation data under credibility constraints; S2, performing object segmentation and feature aggregation based on the standardized observation data under credibility constraints to form an object-level three-dimensional feature set that integrates spectrum, structure, and texture; S3, constructing cross-time object correspondences and extracting driving features of phenology, water level, geometry, and quality using the cross-time object-level three-dimensional feature set, and using the driving features to perform pseudo-change assessment; using the cross-time object-level three-dimensional feature set to determine the consistency of true changes; combining the pseudo-change assessment results and the true change consistency determination results to distinguish the stability of change types; S4, based on the object-level three-dimensional feature set, using the stability distinction results of change types as the basis for data screening, performing supervised training and hierarchical classification, outputting the category probabilities of forest land, grassland, and wetland, and combining the change features to perform rule recognition, forming land use update and change type labeling results.

[0011] Furthermore, the specific process of collecting and preprocessing multi-source remote sensing observation data is as follows: Multi-source remote sensing observation data is collected in real time, including: optical multispectral reflectance, SAR backscattering coefficient, point cloud height, solar elevation angle, observation zenith angle, cloud probability, and shadow probability; the timestamps and coordinate reference information of the multi-source remote sensing observation data are read, and the data is unified to the same geographic coordinate system and spatial grid through coordinate projection transformation and spatial resampling; based on the target resolution, interpolation resampling and raster alignment methods are used to resample and align the multi-source remote sensing observation data to pixels, and cross-source temporal registration is performed based on feature matching registration methods. The following steps were taken: obtaining the pixel-level registration residual field; performing radiometric consistency checks and outlier removal on the optical multispectral reflectance; performing radiometric calibration and speckle noise filtering on the SAR backscattering coefficient; separating ground points from non-ground points in the point cloud through point cloud classification and unifying the height reference using an elevation reference conversion method; reading the solar elevation angle and observation zenith angle parameters, and rasterizing them according to the spatial range and resolution of the corresponding image to ensure a one-to-one correspondence between the solar elevation angle and observation zenith angle and the image pixel space, generating an observation geometric auxiliary raster; performing dimensionless normalization on the multi-source remote sensing observation data; and establishing a resource identification and classification database to store the raw and preprocessed multi-source remote sensing observation data.

[0012] Furthermore, based on the preprocessed multi-source remote sensing observation data, the specific process of performing observation quality assessment and constructing standardized observation data under confidence constraints is as follows: Obtain the pixel-level registration residual field; extract the median and median absolute deviation of the pixel-level registration residuals; for each raster cell, calculate the absolute difference between each pixel-level registration residual and the median; multiply the median absolute deviation by the MAD normal consistency correction coefficient and add a very small positive number as the denominator; divide the absolute difference by the denominator to obtain the robust standardized residual value; calculate the negative exponent of the robust standardized residual value to obtain the geometric registration stability. The process involves defining the terms; subtracting the cloud probability and the shadow probability from each other, multiplying the two differences, and taking the square root to obtain the image quality stability term; multiplying the geometric registration stability term and the image quality stability term to obtain the multi-source availability confidence value; calculating the multi-source availability confidence value for each grid cell, and establishing spatial and temporal index associations with the multi-source remote sensing observation data; using the multi-source availability confidence value as a weighting factor to perform pixel-level weighting on the corresponding preprocessed optical multispectral reflectance, SAR backscattering coefficient, and point cloud height to obtain standardized observation data under confidence constraints.

[0013] Furthermore, the specific process of performing object segmentation and feature aggregation based on standardized observation data under confidence constraints to form an object-level three-dimensional feature set integrating spectrum, structure, and texture is as follows: Weighted optical multispectral reflectance, SAR backscattering coefficient, and point cloud height are used as the segmentation feature input layer. A multi-scale segmentation algorithm is used to perform superpixel segmentation on the weighted data to generate an object unit set. For each object unit in the set, weighted feature statistics are performed: the optical multispectral reflectance and SAR backscattering coefficient of all pixels within the object unit are weighted and averaged based on the multi-source availability confidence value to obtain the spectral vector and SAR backscattering statistics; the weighted point cloud heights within the object unit are sorted and quantile statistics are performed. The upper bound quantile and center quantile of point cloud height are obtained. Based on optical multispectral reflectance, the near-infrared band is selected as the single-channel grayscale input, and the corresponding reflectance is linearly mapped to a finite number of grayscale levels. A grayscale co-occurrence matrix is ​​constructed within the object unit region, and texture statistics are calculated. Based on the spatial contour of the segmented object unit, the area of ​​the object unit region is calculated using the raster counting method, and the perimeter of the object unit is calculated using the boundary cell recognition and boundary length accumulation method. Based on the area and perimeter, the shape compactness is calculated using the shape compactness calculation model. The spectral vector, SAR backscattering statistics, upper bound quantile of point cloud height, center quantile of point cloud height, texture statistics, and shape compactness are combined to construct an object-level three-dimensional feature set.

[0014] Furthermore, the specific process of constructing cross-temporal object correspondences and extracting driving features of phenology, water level, geometry, and mass using cross-temporal object-level 3D feature sets is as follows: For each object unit in the object unit set, obtain object-level 3D feature sets for two time phases, and establish cross-temporal object correspondences using the object space superposition matching method; extract the near-infrared band reflectance and red band reflectance from the spectral vector, divide the difference between the near-infrared band reflectance and the red band reflectance by the sum of the near-infrared band reflectance and the red band reflectance to obtain the vegetation index of the object unit, perform smooth spline fitting on the vegetation index of the object unit in the continuous time series, extract the time point corresponding to the peak from the fitted vegetation index curve, calculate the difference between the peak time points between the two time phases to obtain the phenological phase difference of the object unit; extract... The reflectance of the green band and the shortwave infrared band in the spectral vector are taken. The difference between the green band reflectance and the shortwave infrared band reflectance is divided by the sum of the green band reflectance and the shortwave infrared band reflectance to obtain the water index. The difference of the water index between the two time phases is calculated. The difference of the SAR backscatter statistics of the target unit in the two time phases is also calculated. The difference of the water index is multiplied by the sum of the difference of the SAR backscatter statistics and one to obtain the water level wetting difference of the target unit. The difference of the solar altitude angle and the difference of the observed zenith angle of the target unit in the two time phases are calculated. The two differences are substituted into the cosine function respectively. The product of the two cosine functions is subtracted from one to obtain the observation geometric difference of the target unit. The mean of the multi-source availability confidence value in the target unit is calculated, and the difference of the mean of the multi-source availability confidence value in the two time phases is calculated to obtain the quality difference.

[0015] Furthermore, the specific process of using driving features for pseudo-change assessment is as follows: The phenological phase difference, water level wetting difference, observation geometric difference, and quality difference are combined to form a pseudo-change driving vector for the object unit; the object units in the object unit set whose absolute difference between the upper bound quantile values ​​of point cloud height in two time phases is less than the change threshold are selected to obtain a reference object set; based on the pseudo-change driving vectors of each object unit in the reference object set, the covariance between each component is calculated to form a covariance matrix; the pseudo-change driving vector is transposed and multiplied by the inverse of the covariance matrix to obtain an intermediate result vector; the intermediate result vector is multiplied by the pseudo-change driving vector to obtain a scalar value; the square root of the scalar value is taken to obtain the pseudo-change driving intensity value of the object unit.

[0016] Further, the specific process of using the cross - temporal object - level three - dimensional feature set for true change consistency determination is as follows: For each object unit, calculate three types of change amplitudes respectively: Calculate the Euclidean distance between the spectral vectors in two time phases to obtain the spectral change amplitude; calculate the absolute difference between the upper quantile values of the point cloud height in two time phases to obtain the structural height change amplitude; calculate the absolute difference between the SAR backscattering statistics in two time phases to obtain the scattering change amplitude; Based on the spectral change amplitude, structural height change amplitude, and scattering change amplitude of all object units, calculate the median and median absolute deviation of each type of change amplitude; Calculate the difference between each type of change amplitude and its corresponding median respectively, and use the median absolute deviation multiplied by the MAD normal consistency correction coefficient plus a very small positive number as the denominator, and divide the difference by the denominator to obtain the standardized change values corresponding to the three types of change amplitudes; Perform natural exponential operations on the three standardized change values respectively and then add them together, take the natural logarithm of the added result to obtain the comprehensive change intensity value; Divide the three natural exponential operation results by the added result respectively to obtain the spectral change proportion, structural change proportion, and scattering change proportion. Multiply each change proportion by its corresponding natural logarithm value and then sum them, and take the negative value to obtain the change evidence consistency entropy; Subtract the ratio of the change evidence consistency entropy to the natural logarithm of three from one, and multiply it by the comprehensive change intensity value to obtain the true change evidence consistency value of the object unit.

[0017] Further, the specific process of distinguishing the stability of change types by combining the false change evaluation result and the true change consistency determination result is as follows: Compare the false change driving intensity value of each object unit with the false change threshold t1 respectively, and compare the true change evidence consistency value with the consistency threshold t2: When < t1 and ≥ t2, determine that the object unit is a real change object; When ≥ t1 and < t2, determine that the object unit is a false change object; When ≥ t1 and ≥ t2, determine that the object unit is an uncertain change object; When < t1 and < t2, determine that the object unit is a stable object.

[0018] Furthermore, based on the object-level 3D feature set, and using the stability of change types as the data selection criterion, supervised training and hierarchical classification are performed to output the category probabilities of forest, grassland, and wetland. Combined with change characteristics, rule recognition is conducted to form the land use update and change type labeling results. The specific process is as follows: using the object-level 3D feature set as input features, combined with the already labeled land use tags, object units are spatially matched with their corresponding land use types to form supervised training samples; after normalizing the input features of the supervised training samples, a random forest supervised learning algorithm is used for model training, and the model parameters are optimized through cross-validation to construct a hierarchical classification model. For each object unit, the object-level 3D feature set is input into the hierarchical classification model. Specifically, the post-temporal object-level 3D feature set is input for real-change objects, while the pre-temporal object-level 3D feature set is input for stable objects and pseudo-change objects. The classification probabilities of forest, grassland, and wetland for each object unit are output. The category with the highest classification probability is selected as the main classification label for the object unit. For real-change objects, the spectral change amplitude, structural height change amplitude, scattering change amplitude, phenological phase difference, and water level wetting difference are read and compared with the corresponding judgment thresholds to perform rule recognition and obtain the different change types of real-change objects.

[0019] The second aspect of this invention provides a dynamic identification and classification system for forest, grassland, and wetland resources based on multi-source remote sensing fusion, comprising: a data acquisition and processing module for acquiring multi-source remote sensing observation data and preprocessing the multi-source remote sensing observation data; performing observation quality assessment based on the preprocessed multi-source remote sensing observation data to construct standardized observation data under credibility constraints; an object feature construction module for performing object segmentation and feature aggregation based on the standardized observation data under credibility constraints to form an object-level three-dimensional feature set that integrates spectrum, structure, and texture; a pseudo-change determination module for constructing cross-time object correspondences and extracting driving features of phenology, water level, geometry, and quality using the cross-time object-level three-dimensional feature set, and using the driving features to perform pseudo-change assessment; using the cross-time object-level three-dimensional feature set to perform true change consistency determination; combining the pseudo-change assessment results and the true change consistency determination results to distinguish the stability of change types; and a classification change identification module for performing supervised training and hierarchical classification based on the object-level three-dimensional feature set, using the stability distinction results of change types as the basis for data screening, outputting the category probabilities of forest land, grassland, and wetland, and combining change features to perform rule recognition to form land use update and change type labeling results.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) In this invention, a multi-source usable confidence value is constructed by using pixel-level registration residuals, cloud probability and shadow probability to perform pixel-level weighted constraints on optical, SAR and point cloud data, which effectively weakens the impact of observation geometric differences, cloud and shadow interference and registration errors on the results and improves the consistency and stability of cross-time data.

[0023] (2) In this invention, an object-oriented segmentation and feature aggregation method is used to fuse spectral vectors, SAR backscatter statistics, point cloud height quantile features, texture statistics and shape compactness into a model, thereby realizing the transformation from pixel-level analysis to object-level comprehensive expression and enhancing the discrimination ability under complex terrain and heterogeneous material conditions.

[0024] (3) In this invention, a pseudo-change driving vector is constructed by phenology, water level, observation geometry and quality difference, and intensity assessment is carried out by combining covariance constraints. At the same time, the consistency of true change is determined by the multi-evidence consistency entropy mechanism, which effectively distinguishes between non-real fluctuations and real land cover transfers, and reduces the risk of false alarms and false alarms.

[0025] (4) This invention, based on the differentiation of change stability, inputs the features of the previous or subsequent time phase for different object categories to perform hierarchical classification, and combines the change amplitude and driving features to perform rule recognition, thereby realizing the integrated output of land category update and change type labeling, and improving the refinement and practicality of resource dynamic identification.

[0026] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0027] Figure 1 Flowchart of a dynamic identification and classification method for forest, grassland and wetland resources based on multi-source remote sensing fusion;

[0028] Figure 2 A structural diagram of a dynamic identification and classification system for forest, grassland and wetland resources based on multi-source remote sensing fusion.

[0029] Figure 3 A comparison chart of the distribution of pseudo-change driving intensity values;

[0030] Figure 4 A comparison chart of consistency values ​​for evidence of true change. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. As those skilled in the art will understand, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figures 1-4 This invention provides a technical solution: a dynamic identification and classification method for forest, grassland, and wetland resources based on multi-source remote sensing fusion, such as... Figure 1 As shown, the process includes the following steps: S1, collecting multi-source remote sensing observation data and preprocessing the data; performing observation quality assessment based on the preprocessed data to construct standardized observation data under credibility constraints; S2, performing object segmentation and feature aggregation based on the standardized data to form an object-level three-dimensional feature set integrating spectral, structural, and textural data; S3, constructing cross-temporal object correspondences and extracting driving features of phenology, water level, geometry, and quality using the cross-temporal object-level three-dimensional feature set, and using these features to evaluate pseudo-changes; using the cross-temporal object-level three-dimensional feature set to determine the consistency of true changes; and combining the pseudo-change evaluation results with the true change consistency determination results to distinguish the stability of change types; S4, based on the object-level three-dimensional feature set, using the stability distinction results of change types as the basis for data selection, performing supervised training and hierarchical classification, outputting the category probabilities of forest, grassland, and wetland, and combining change features to perform rule recognition, forming land use update and change type labeling results.

[0033] Specifically, the process of collecting and preprocessing multi-source remote sensing data is as follows: Multi-source remote sensing data is collected in real time, including: optical multispectral reflectance, SAR backscattering coefficient, point cloud height, solar elevation angle, observation zenith angle, cloud probability, and shadow probability. Optical multispectral reflectance includes surface reflectance data in the blue, green, red, near-infrared, and short-wave infrared bands. The surface reflectance data for each band are converted to apparent surface reflectance values ​​through atmospheric correction and radiometric calibration, and anomalous pixels are masked. The red and near-infrared bands are used to construct vegetation indices, the green and short-wave infrared bands are used to construct water body indices, and the remaining bands participate in the construction of object-level spectral features. The data includes: optical multispectral reflectance data obtained from surface reflectance data acquired by satellite optical remote sensing sensors; SAR backscattering coefficient data obtained from radiometrically calibrated data acquired by synthetic aperture radar sensors; point cloud height data obtained from point cloud data acquired by an airborne LiDAR laser scanning system and elevation information extracted; solar altitude angle and observed zenith angle parameters acquired by reading the metadata files of the corresponding remote sensing images; and cloud probability and shadow probability data acquired by generating quality control bands after applying a cloud detection algorithm to the remote sensing images. All of the above data are input in a standard digital raster format. The timestamps and coordinate reference information of the multi-source remote sensing observation data are read, and the multi-source remote sensing observation data are unified to the same geographic coordinate system and the same spatial grid through coordinate projection transformation and spatial resampling. Coordinate projection transformation is preferably implemented using GDAL or equivalent geographic information processing tools to uniformly transform different data sources to a preset projection coordinate system. The spatial grid is preferably set as a regular raster with a uniform spatial resolution, and the resolution is adaptively set according to the application scenario. The timestamp is used to identify the observation time and establish a cross-time data index. Based on the target resolution, interpolation resampling and raster alignment methods are used to resample and align pixels of multi-source remote sensing observation data. Cross-source temporal registration is then performed based on feature matching registration to obtain a pixel-level registration residual field. Specifically, the target resolution is automatically determined during the initialization phase by reading and comparing the spatial resolution metadata of the input data. It is preferably set to the coarsest spatial resolution among the multi-source remote sensing observation data to ensure that all data sources are fused at a unified spatial scale, avoiding spurious detail errors introduced by excessive interpolation from high-resolution data to low-resolution data. When multiple data sources have inconsistent resolutions, the coarsest resolution is used as the unified resampling scale. Bilinear interpolation is preferred for interpolation resampling. Raster alignment is achieved by unifying the origin and row / column number rules. Cross-source temporal registration preferably uses a method based on feature point matching and affine mapping. The registration residual is calculated using the spatial distance difference between corresponding control points, forming a rasterized pixel-level registration residual field.The radiometric consistency of optical multispectral reflectance is checked and outliers are removed. Radiometric calibration and speckle noise filtering are performed on the SAR backscattering coefficient. Point cloud height is separated into ground and non-ground points through point cloud classification, and a height benchmark conversion method is used to unify the height benchmark. Specifically, the radiometric consistency check preferably uses statistical analysis of the mean and standard deviation of different bands to determine stability; outlier removal is preferably achieved based on the median absolute deviation method; SAR radiometric calibration preferably converts sensor calibration coefficients to physical backscattering coefficients; speckle noise filtering preferably uses the Lee filtering method; point cloud classification preferably uses a progressively denser triangulation algorithm to distinguish ground and non-ground points; and height benchmark conversion is preferably performed based on a unified geodetic height. The solar elevation angle and observed zenith angle parameters are read and rasterized according to the spatial range and resolution of the corresponding image to ensure a one-to-one correspondence between the solar elevation angle and observed zenith angle and the image pixel space, generating an observation geometric auxiliary raster. If the solar elevation angle and observed zenith angle are image-level constants, a raster of the same size is generated by pixel-by-pixel replication. If they are pixel-by-pixel parameters, they are directly read and resampled to a unified grid. The observation geometric auxiliary raster is used for subsequent geometric difference analysis and feature extraction. The "image pixel space" refers to the two-dimensional raster spatial structure formed by image data in a unified geographic coordinate system and a unified spatial grid. Each raster unit corresponds to an image pixel location and has a unique row and column index and spatial coordinates, representing the discrete spatial representation of the image data. By expanding the solar elevation angle and observed zenith angle to a size consistent with this raster structure, their spatial positions are made to correspond one-to-one with each image pixel. Dimensionless normalization was performed on multi-source remote sensing data. The preferred normalization method was Robus standardization based on statistics, mapping different physical quantities to a unified numerical range to eliminate the impact of dimensional differences on subsequent fusion analysis and enhance robustness against outliers. A resource identification and classification database was established to store raw and preprocessed multi-source remote sensing data. The database supports spatial and temporal indexes, enabling unified storage and rapid retrieval of pixel-level, object-level, and time-series data, providing data support for subsequent object segmentation, feature construction, and change determination.

[0034] This implementation plan achieves standardized fusion of multi-source optical, SAR, and LiDAR data at the spatial, temporal, and physical levels by constructing a data framework with unified projection, unified grid, and unified temporal index. Combined with radiometric correction, registration residual control, geometric parameter rasterization expansion, and dimensionless normalization processing, it effectively improves the consistency and comparability of cross-source and cross-temporal data. Simultaneously, relying on spatial and temporal index databases, it enables structured management and efficient retrieval, providing a stable, reliable, and feasible data foundation for subsequent object-level feature construction and change determination, thereby enhancing overall accuracy, robustness, and engineering feasibility.

[0035] Specifically, the process of constructing standardized observation data under credibility constraints by performing observation quality assessment based on preprocessed multi-source remote sensing observation data is as follows: First, obtain the pixel-level registration residual field. The pixel-level registration residual field is continuous raster data generated by interpolation of control point spatial offsets during cross-source temporal registration, used to characterize the geometric registration error distribution at each pixel location. Extract the median and median absolute deviation of the pixel-level registration residuals. The median is used to characterize the central level of the overall registration error, and the median absolute deviation is used to characterize the dispersion of the residuals to enhance robustness to outliers. For each raster cell, calculate the absolute difference between each pixel-level registration residual and the median. Multiply the median absolute deviation by the MAD normal consistency correction coefficient and add a very small positive number as the denominator. The MAD normal consistency correction coefficient is used to map the median absolute deviation to a scale consistent with the standard deviation; a recommended value is 1.4826. A very small positive number is used to avoid numerical instability caused by a zero denominator; a recommended value is [missing value]. The robust standardized residual value is obtained by dividing the absolute difference by the denominator. The negative exponent of the robust standardized residual value is then calculated to obtain the geometric registration stability term. Smaller residuals correspond to a geometric registration stability term closer to 1, while larger residuals correspond to a geometric registration stability term closer to 0, thus forming a monotonically decreasing stability mapping relationship. The cloud probability and shadow probability are subtracted by 1, where the cloud and shadow probabilities are pixel-by-pixel probability values ​​generated by the quality control algorithm, ranging from 0 to 1. The square root of the two difference results is then used to obtain the image quality stability term. This square root form achieves a balance constraint on the two types of quality factors, avoiding the dominance of a single factor. The geometric registration stability term is then multiplied by the image quality stability term to obtain... A multi-source availability confidence value is constructed to achieve a joint constraint expression of geometric consistency and imaging quality. The multi-source availability confidence value is constructed using a multiplicative coupling structure, with the geometric registration stability term as the spatial consistency constraint factor and the imaging quality stability term as the observation reliability constraint factor. The joint expression mechanism of "simultaneous satisfaction of dual constraints" is achieved through a product form. When either factor is low, the product result decreases synchronously, thereby suppressing the interference of geometric mismatch or poor imaging quality on subsequent feature construction and change judgment. When both factors remain stable, the multi-source availability confidence value remains at a high level, which is used to enhance the weight contribution of high-quality observation data in fusion analysis and realize a confidence-driven data weighting control mechanism. The multi-source availability confidence value of each grid cell is calculated, and spatial and temporal indices are established and associated with the multi-source remote sensing observation data. The spatial index is established based on the row and column numbers of the unified grid, and the temporal index is established based on the observation timestamp, which is used to realize rapid retrieval and access of data across time. The multi-source availability confidence value is used as a weighting factor to perform pixel-level weighting on the corresponding preprocessed optical multispectral reflectance, SAR backscattering coefficient, and point cloud height, so that pixels with higher observation quality have higher weights in subsequent feature construction and change analysis, thereby improving the stability and reliability of the fusion results and obtaining standardized observation data under confidence constraints.

[0036] The specific formula for the multi-source available credibility value is as follows:

[0037] ;

[0038] In the formula, This represents the multi-source availability confidence value, used to measure the grid cell. The overall reliability of multi-source observation data is indicated by a higher value, which means that the pixels are more stable in terms of geometric registration and imaging quality. It represents the coordinate position of a raster unit, i.e., a pixel; This represents the robust standardized residual value, which measures the degree of geometric registration anomaly of the current pixel. The larger the value, the more significant the pixel shift during cross-time registration. The value indicates the probability that a pixel is covered by clouds. The larger the value, the more likely the optical reflectivity may be obstructed. The value represents the probability of a pixel being in the shadow of terrain or clouds. The larger the value, the more likely the spectral reflectance is affected by darkening interference.

[0039] In this implementation scheme, by introducing a geometric registration stability assessment based on robust statistics and an imaging quality assessment based on cloud shadow probability, a multi-source availability confidence value under the dual constraints of geometric consistency and observation reliability is constructed, realizing quantitative expression and pixel-level weighted control of the quality of multi-source remote sensing data. The multiplicative coupling mechanism is used to suppress the interference of registration errors and imaging quality anomalies on subsequent analysis, and spatial and temporal indexes are combined to achieve accurate correlation and fast retrieval, thereby improving the stability, consistency and reliability of multi-source data fusion results.

[0040] Specifically, the process of performing object segmentation and feature aggregation based on standardized observation data under credibility constraints to form an object-level 3D feature set integrating spectrum, structure, and texture is as follows: Weighted optical multispectral reflectance, SAR backscattering coefficient, and point cloud height are used as input layers for segmentation features. The weighted data is standardized raster data processed by introducing multi-source available credibility values ​​as weighting factors at the pixel level. Each feature layer maintains complete consistency in spatial resolution, projection coordinate system, and row and column numbers. A multi-scale segmentation algorithm is used to perform superpixel segmentation on the weighted data to generate a set of object units. The preferred multi-scale segmentation algorithm is an object-oriented segmentation method based on region growing and region merging mechanisms. Specifically, firstly, using pixels as initial units, a similarity measurement function is constructed based on the spectral differences, structural differences, and spatial adjacency relationships between adjacent pixels. Then, under the condition of satisfying a preset heterogeneity threshold, adjacent regions are gradually merged until the heterogeneity increment within the region exceeds the scale parameter control threshold, thereby forming a stable set of object units. The similarity metric function employs a weighted combination of spectral heterogeneity and shape heterogeneity. Spectral heterogeneity is calculated from the weighted reflectance differences across bands, while shape heterogeneity is expressed by a shape compactness index. Spectral and shape weight parameters control the relative contribution of the two types of heterogeneity in the merging decision. The algorithm can be implemented based on existing object-oriented image analysis frameworks, using region adjacency graph construction and iterative merging strategies. The algorithm flow includes five steps: initialization, adjacency relationship establishment, heterogeneity calculation, candidate region selection, and iterative merging.For each object unit in the object unit set, weighted feature statistics are performed: the optical multispectral reflectance and SAR backscattering coefficient of all pixels within the object unit are weighted and averaged based on the multi-source availability confidence value to obtain the spectral vector and SAR backscattering statistics; the weighted average is achieved by multiplying the optical multispectral reflectance or SAR backscattering coefficient of each pixel within the object by its corresponding weight, summing the results, and then dividing by the total weighted average to ensure that pixels with higher observation quality contribute more to the object features; the weighted point cloud heights within the object unit are sorted and quantile statistics are performed to obtain the upper bound quantile value of the point cloud height and the point cloud height. The point cloud height center quantile is preferably P50, used to characterize the median structural height level of the object region; the point cloud height upper bound quantile is preferably P95, used to characterize the upper bound feature of the high-value structure of the object region, to enhance the ability to depict the canopy or tall vegetation structure; based on optical multispectral reflectance, the near-infrared band is selected as the single-channel grayscale input, and the corresponding reflectance is linearly mapped to a finite number of grayscale levels. The number of grayscale levels is preferably set to 32 to balance computational efficiency and texture expression ability, and a grayscale co-occurrence matrix is ​​constructed within the object unit region to calculate texture statistics; wherein the grayscale co-occurrence matrix... Based on preset pixel spacing and orientation, the system is preferably generated at 0° with a distance of one pixel, and contrast, homogeneity, and entropy statistics are extracted as texture features. Based on the spatial contour of the segmented object units, the area of ​​the object unit region is calculated using a raster counting method. The area is obtained by multiplying the number of pixels inside the object by the area of ​​a single pixel. The perimeter of the object unit is calculated using boundary pixel recognition and boundary length accumulation methods. The perimeter is achieved by traversing the boundary pixels of the object and accumulating their boundary lengths. Based on the area and perimeter, a morphological compactness calculation model is used, preferably employing a standard of four times pi multiplied by the area and then divided by the square of the perimeter. The compactness model calculates shape compactness to ensure that the index values ​​are stable and have clear physical meaning. The spectral vector, SAR backscatter statistics, upper bound quantile of point cloud height, center quantile of point cloud height, texture statistics, and shape compactness are combined to construct an object-level 3D feature set. The object-level 3D feature set includes a feature set that integrates spectral features, structural height features, and texture features. The object-level 3D feature set is stored with the unique index of the object unit as the primary key and is associated with the spatial index and temporal index. It is used for subsequent cross-time feature comparison, change judgment, and classification recognition processes to ensure that the feature data source is clear, the structure is complete, and it is traceable.

[0041] In this implementation scheme, by performing object segmentation and multi-source feature aggregation under credibility constraints, spectral information, SAR scattering features, structural height features, and texture and morphological features are uniformly expressed at the object scale, achieving a stable mapping from pixel-level data to object-level three-dimensional features. At the same time, by combining weighted statistics and quantile extraction mechanisms, the interference of low-quality observations on feature construction is effectively suppressed, enhancing the robustness and consistency of object representation. This provides a highly reliable feature foundation with a clear structure, complementary information, and traceability for subsequent cross-time change determination and hierarchical classification.

[0042] Specifically, the process of constructing cross-temporal object correspondences and extracting driving features of climatology, water level, geometry, and mass using cross-temporal object-level 3D feature sets is as follows: For each object unit in the object unit set, two temporal-level 3D feature sets are obtained, with the two temporal phases corresponding to two different observation timestamps. The object units of the two temporal phases are constructed based on a unified spatial grid and a consistent projected coordinate system, and the cross-temporal object correspondences are established using an object space overlay matching method. Specifically, the object space overlay matching method calculates the overlap area ratio of the spatial contours of the two temporal object units, selects the object unit with the largest overlap area ratio as the matching object, and confirms that the object unit is a cross-temporal corresponding object of the same spatial entity when the overlap ratio exceeds the matching threshold, so as to ensure the uniqueness and stability of the correspondence.Near-infrared and red reflectance are extracted from the spectral vector. Both the near-infrared and red reflectances originate from the weighted aggregated spectral vector components in the object-level 3D feature set. The vegetation index of the object unit is obtained by dividing the difference between the near-infrared and red reflectances by the sum of their values. Smoothing spline fitting is then applied to the vegetation index of the object unit over a continuous time series. The continuous time series consists of vegetation indices of the same object unit at multiple timestamps arranged chronologically. Smoothing spline fitting is used to reduce the impact of short-term fluctuations on peak identification. To assess the impact of phenological factors, the peak time points were extracted from the fitted vegetation index curves, and the difference between the peak time points of the two time phases was calculated to obtain the phenological phase difference of the target unit. The green band reflectance and shortwave infrared band reflectance were extracted from the spectral vector. Since the green band and shortwave infrared band also originate from the corresponding components of the target-level spectral vector, the difference between the green band reflectance and shortwave infrared band reflectance was divided by the sum of the green band reflectance and shortwave infrared band reflectance to obtain the water index. The difference between the water indexes of the two time phases was calculated. Furthermore, the SAR values ​​of the target unit in the two time phases were calculated respectively. The difference in backscattering statistics is calculated, where the SAR backscattering statistics are the scattering intensity characteristics after object-level weighted averaging. The difference in water index is multiplied by the sum of the difference in SAR backscattering statistics and one to obtain the water level wetting difference of the object unit. This modulates the water index change by introducing scattering variations to enhance sensitivity to changes in wetting levels and reduce the risk of misjudgment caused by single spectral changes. The difference in solar altitude angle and observed zenith angle of the object unit under two time phases is calculated. The solar altitude angle and observed zenith angle are derived from the aforementioned observation geometric auxiliary grid and are averaged within the object unit area. After calculation, the two differences are substituted into the cosine function, and the product of the two cosine functions is subtracted from one to obtain the observation geometric difference of the object unit. This is used to characterize the potential impact of observation geometric differences on spectral and scattering changes. The mean of the multi-source available confidence value within the object unit is calculated, and the difference between the mean of the multi-source available confidence values ​​of the two time phases is calculated to obtain the quality difference. The mean of the multi-source available confidence value is the arithmetic mean of the confidence of all pixels within the object unit, which is used to characterize the overall observation quality level of the object. The quality difference is used to reflect the degree of influence of changes in observation conditions at two time points on the reliability of the results.

[0043] In this implementation plan, by establishing a stable cross-time object correspondence, multi-dimensional driving features such as phenology, water level, observation geometry, and data quality are extracted together at the object scale to achieve a systematic decoupled expression of spectral changes, structural changes, and changes in observation conditions. This can not only reduce the interference of non-real fluctuations caused by phenological differences, water level fluctuations, and changes in observation geometry, but also enhance the sensitivity to real land cover transfer and structural change signals, thereby improving the stability, accuracy, and comparability of cross-time comparative analysis.

[0044] Specifically, the process of pseudo-change assessment using driving characteristics is as follows: The phenological phase difference, water level wetting difference, observation geometric difference, and quality difference are combined to form a pseudo-change driving vector for the object unit. This pseudo-change driving vector is a four-dimensional column vector, with each of its four components representing the differences in phenological, water level, geometric, and quality factors between two time points, used to uniformly characterize the comprehensive impact of non-land cover transfer factors on observed changes. Object units in the object unit set whose absolute difference between the upper bound quantile values ​​of point cloud height in two time phases is less than a change threshold are selected to obtain a reference object set. This change threshold is a structural stability threshold set based on the overall statistical distribution of the upper bound quantile values ​​of point cloud height, used to screen object units whose structural height remains basically stable, ensuring that the reference object set mainly reflects pseudo-change characteristics under non-structural change backgrounds. Based on the pseudo-change driving vectors of each object unit in the reference object set, the covariance between each component is calculated to form a covariance matrix. The covariance matrix is ​​used to characterize the correlation and joint fluctuation characteristics between different driving components, thereby establishing the statistical distribution structure of pseudo-change factors. The pseudo-change driving vector is transposed and multiplied by the inverse of the covariance matrix to obtain an intermediate result vector. This intermediate result vector is then multiplied by the pseudo-change driving vector to obtain a scalar value. To avoid singular or near-singular covariance matrix issues due to a small sample size in the reference set or high correlation between driving components, the covariance matrix is ​​stabilized before calculating its inverse. Specifically, regularization terms are added to the diagonal of the covariance matrix. ,in It is the identity matrix. The positive regularization coefficient is non-negative, and the preferred value is [value missing]. to A fixed number of small positive numbers within a certain range is used to enhance the numerical stability of the matrix; then, the covariance matrix after adding the regularization term is inverted. When the number of samples in the reference object set is insufficient to meet the full-rank condition of the covariance matrix, the Moore-Penrose pseudo-inverse matrix is ​​preferred to replace the direct inversion method for matrix operations, so as to ensure the computability and engineering feasibility of the statistical distance measurement process. The scalar value is equivalent to a multidimensional distance metric based on the covariance structure, used to characterize the degree to which the driving characteristics of the current object unit deviate from the reference background distribution. The square root of the scalar value is used to obtain the pseudo-change driving intensity value of the object unit; whereby the square root is used to restore the distance scale consistent with the original dimension, so that the pseudo-change driving intensity value has comparability and interpretability, thereby realizing the quantitative evaluation of non-real change driving factors. The pseudo-change driving intensity value is based on a multidimensional covariance structure established by a set of reference objects. It performs statistical distance measurement on the driving vector of the current object unit, and eliminates the scale differences and correlation effects of different driving components by covariance weighting, quantifying the degree of deviation from the "structural stability background distribution". The greater the deviation, the stronger the possibility of being driven by non-land type factors such as phenology, water level, geometry or mass, thus achieving a unified characterization of pseudo-change risk.

[0045] The specific formula for the pseudo-change driving intensity value is as follows:

[0046] ;

[0047] In the formula, The pseudo-change driving intensity value represents the pseudo-change intensity of the object unit in the current cross-time comparison, and measures whether the current object change is mainly caused by external factors such as phenological differences, water level fluctuations, observed geometric changes or mass fluctuations. Represents the object cell index, and indicates the object-level cell number generated after segmentation; This represents a pseudo-change driving vector, which describes the non-real change driving characteristics of an object unit in cross-time comparison; This represents the transpose of the pseudo-change driving vector, used to perform matrix multiplication. It represents the inverse of the covariance matrix, used to eliminate the differences in the dimensions of different driving components, consider the correlation between the driving factors, and achieve standardization in the form of Mahalanobis distance.

[0048] In this embodiment, Table 1 is a data table of pseudo-change driving intensity values. The table details the phenological phase difference, water level wetting difference, observation geometric difference, mass difference, and pseudo-change driving intensity values ​​for five object units; specifically, object unit 1 has a phenological phase difference of 0.05, a water level wetting difference of 0.07, an observation geometric difference of 0.08, a mass difference of 0.05, and a pseudo-change driving intensity value of 0.14; object unit 2 has a phenological phase difference of 0.10, a water level wetting difference of 0.12, an observation geometric difference of 0.18, a mass difference of 0.12, and a pseudo-change driving intensity value of 0.27; object unit 3 has a phenological phase difference of 0.05, a water level wetting difference of 0.12, an observation geometric difference of 0.18, a mass difference of 0.12, and a pseudo-change driving intensity value of 0.27; object unit 3 has a phenological phase difference of 0.05, a water level wetting difference of 0.07, an observation geometric difference of 0.08, a mass difference of 0.05 ...7, a mass difference of 0.08, a mass difference of 0.05, and a pseudo-change driving intensity value of 0.27. The phenological phase difference is 0.80, the water level wetting difference is 0.95, the observation geometric difference is 1.10, the mass difference is 0.70, and the pseudo-change driving intensity value is 2.02; for object unit 4, the phenological phase difference is 0.30, the water level wetting difference is 0.45, the observation geometric difference is 0.60, the mass difference is 0.40, and the pseudo-change driving intensity value is 0.93; for object unit 5, the phenological phase difference is 1.20, the water level wetting difference is 1.50, the observation geometric difference is 1.60, the mass difference is 1.10, and the pseudo-change driving intensity value is 3.08.

[0049] Table 1 Data on the driving intensity of pseudo-changes

[0050]

[0051] like Figure 3 The figure shows a comparison of the distribution of pseudo-change driving intensity values. It illustrates the distribution of pseudo-change driving intensity values ​​for five object units. The horizontal axis represents the object unit number, the vertical axis represents the corresponding pseudo-change driving intensity value, and the dashed line represents the pseudo-change threshold. (This is in conjunction with Table 1 and...) Figure 3 It can be seen that the values ​​of object units 1 and 2 are significantly lower than the threshold line, indicating that the overall phenological phase difference, water level and humidity difference, observational geometric difference, and mass difference are relatively small, and the influence of external disturbances is weak, belonging to a low pseudo-change driven state. Object units 3 and 5 are significantly higher than the threshold line, especially object unit 5, which has the highest value, indicating that its driving factors fluctuate significantly, and the change is more likely to be dominated by external factors such as phenological differences, water level fluctuations, or observational geometric changes, belonging to a strong pseudo-change driven state. Object unit 4 is close to the threshold line and is in the critical range, requiring further comprehensive judgment based on the consistency value of true change evidence. The overall distribution is consistent with the numerical change trend of the original driving factors, including phenology, water level, geometric difference, and mass difference, indicating that the comprehensive assessment in the form of Mahalanobis distance can effectively amplify the superposition effect of multiple factors.

[0052] In this implementation plan, by constructing a set of reference objects under a structurally stable background, a multidimensional distance measurement mechanism with covariance weighting is introduced to comprehensively characterize the joint disturbance effects of non-land-type factors such as phenology, water level, observation geometry and quality on a unified scale, thereby achieving a quantitative assessment of the degree of pseudo-change driving force. This not only eliminates the interference of scale differences and correlations between different driving components, but also enhances the ability to identify non-real change signals, thus effectively reducing the risk of misjudgment in cross-time change analysis and improving the stability and reliability of change determination.

[0053] Specifically, the process of using the cross-time object-level three-dimensional feature set to determine the consistency of true changes is as follows: For each object unit, three types of change amplitudes are calculated respectively: the Euclidean distance between the spectral vectors under two time phases is calculated to obtain the spectral change amplitude; wherein, the spectral vector is composed of the weighted average reflectance of each band in the object unit, and the Euclidean distance is calculated based on the square root of the sum of the squares of the differences of each corresponding band, which is used to characterize the degree of difference in the overall spectral structure. The absolute difference between the upper bound quantile values ​​of point cloud height in two time phases is calculated to obtain the structural height variation amplitude. This amplitude characterizes significant changes in the canopy or surface structure of the target unit at high quantile heights, reflecting structural changes such as logging, degradation, or growth. The absolute difference between the SAR backscatter statistics in two time phases is calculated to obtain the scatter variation amplitude. This SAR backscatter statistics reflect changes in surface roughness and humidity conditions. Based on the spectral variation amplitude, structural height variation amplitude, and scatter variation amplitude of all target units, the median and median absolute deviation of each type of variation amplitude are calculated. The median characterizes the central level of overall variation, while the median absolute deviation enhances robustness to anomalous large variations. The difference between each type of variation amplitude and its corresponding median is calculated, and the corresponding median absolute deviation is multiplied by the MAD normality consistency correction coefficient, then a minimum positive number is added as the denominator. The difference is divided by the denominator to obtain the standardized variation values ​​corresponding to the three types of variation amplitudes. The recommended value for the MAD normality consistency correction coefficient is 1.4826, and the recommended minimum positive number is [value missing]. To ensure numerical stability, the three types of standardized change values ​​are calculated using natural exponents and then summed. The natural logarithm of the sum is then taken to obtain the comprehensive change intensity value. This process uses logarithmic exponent transformation to achieve smooth fusion and scale compression of multiple components, avoiding abnormal amplification of a single component. The results of the three natural exponent calculations are divided by the sum to obtain the proportions of spectral change, structural change, and scattering change. The sum of these three proportions is 1, used to characterize the relative contribution of the three types of change in the comprehensive change. Each change proportion is multiplied by its corresponding natural logarithm, summed, and the negative is taken to obtain the consistency entropy of change evidence. The consistency entropy measures the degree of balance in the distribution of the contributions of the three types of change; a lower value indicates that a certain change is more dominant, while a higher value indicates that the contributions of multiple types of change are similar. The true change evidence consistency value of an object unit is obtained by subtracting the ratio of the consistency entropy of change evidence to the natural logarithm 3 from 1, and then multiplying it by the comprehensive change intensity value. Here, the natural logarithm 3 is the theoretical maximum entropy value of the three-component probability distribution, used to normalize the consistency entropy of change evidence, limiting it to the interval between 0 and 1, thereby achieving a consistent and comparable expression between different objects. Forest and grassland wetland transfer is usually accompanied by a single dominant transition in structure or spectrum. Uniform small-amplitude fluctuations are mostly disturbances. Therefore, true change should be driven by a single dominant channel. The true change evidence consistency value is based on the joint expression of the amplitudes of three types of change: spectral, structural, and scattering. It characterizes the scale of change by comprehensively assessing the change intensity and introduces a normalized consistency entropy to measure the degree of dominance among the three types of evidence. While ensuring that the change amplitude is sufficiently significant, it strengthens the prominent change scenarios of a single dominant evidence and suppresses the disturbance scenarios of uniform fluctuations in multiple evidences, thereby achieving stable identification and consistent constraint expression of true land cover transfer signals.

[0054] The specific formula for the consistency value of true change evidence is as follows:

[0055] ;

[0056] In the formula, The consistency value represents the true change evidence, which is used to further assess whether the change evidence comes from multi-source consistent enhancement rather than a single channel anomalous amplification, based on the calculated overall change intensity. This represents the object cell index; the larger the value, the stronger the change and the more consistent the evidence from multiple sources. This represents the overall change intensity value, indicating the total change magnitude of the object unit in the multi-source feature space; Entropy represents the consistency of evidence of change, measuring the degree of equilibrium among sources of change.

[0057] In this embodiment, Table 2 is a data table of consistency values ​​for true change evidence. The table details the overall change intensity value, change evidence consistency entropy, and true change evidence consistency value for five object units; specifically, object unit 1 has an overall change intensity value of 0.08, a change evidence consistency entropy of 0.90, and a true change evidence consistency value of 0.01; object unit 2 has an overall change intensity value of 1.80, a change evidence consistency entropy of 0.20, and a true change evidence consistency value of 1.47; object unit 3 has an overall change intensity value of 0.15, a change evidence consistency entropy of 1.00, and a true change evidence consistency value of 0.01; object unit 4 has an overall change intensity value of 0.90, a change evidence consistency entropy of 0.45, and a true change evidence consistency value of 0.53; and object unit 5 has an overall change intensity value of 0.20, a change evidence consistency entropy of 0.95, and a true change evidence consistency value of 0.02.

[0058] Table 2. Consistency Values ​​of Evidence of True Change

[0059]

[0060] like Figure 4 The figure shown is a comparison chart of the distribution of consistency values ​​for evidence of true change. It displays the distribution of consistency values ​​for evidence of true change across five object units. The horizontal axis represents the object unit number, the vertical axis represents the consistency value for evidence of true change, and the dashed line represents the consistency threshold. (This is in conjunction with Table 2 and...) Figure 4 It can be seen that the consistency value of true change evidence for object unit 2 is significantly higher than the threshold, making it the only object that clearly exceeds the judgment threshold, indicating high change intensity and good consistency of multi-source evidence. The consistency value of true change evidence for object unit 4 is close to but below the threshold, classifying it as a borderline object, indicating a certain degree of change intensity, but insufficient evidence consistency or overall strength to support a definitive judgment. The consistency values ​​of true change evidence for objects units 1, 3, and 5 are far below the threshold, indicating that although individual components may fluctuate, the overall change intensity is low or multi-source consistency is poor. Combined with... Figure 3 and Figure 4It can be seen that the pseudo-change driving intensity value of object unit 2 is lower than the threshold, while the true-change evidence consistency value is significantly higher than the threshold, indicating that its change is not driven by external factors such as phenology, water level, or observation geometry, but rather by the consistent enhancement of multi-source spectral, structural, and scattering evidence, belonging to a true-change object; for object units 3 and 5, the pseudo-change driving intensity values are significantly higher than the threshold while the true-change evidence consistency values are extremely low, indicating that their changes are mainly affected by external disturbances and the multi-source evidence is inconsistent, being typical pseudo-change objects; object units 1 and 4 are below the threshold in both indicators, being stable objects. Thus, it can be seen that the pseudo-change driving intensity value is used to characterize the intensity of external driving interference, and the true-change evidence consistency value is used to measure the consistency of multi-source change evidence. The two jointly constitute a two-dimensional discrimination space, which can effectively distinguish true changes, pseudo-changes, and stable objects, reflecting the effectiveness of the determination mechanism in terms of numerical separability and engineering feasibility.

[0061] In this implementation plan, by constructing a unified expression system for the change amplitudes of spectral, structural, and scattering types, and using a robust normalization and logarithmic exponential fusion mechanism to achieve the smooth integration of multi-component changes, on this basis, a normalized consistency entropy is introduced to constrain the degree of dominance concentration of the three types of changes, enabling the change scale and evidence concentration to participate in the determination simultaneously, thereby effectively distinguishing the true land-cover transfer dominated by a single factor from the fluctuations caused by multi-factor balanced disturbances, and enhancing the stability, anti-interference ability, and determination reliability of cross-temporal change recognition.

[0062] Specifically, the specific process of distinguishing the stability of change types by combining the pseudo-change evaluation results and the true-change consistency determination results is as follows: The pseudo-change driving intensity value of each object unit is compared with the pseudo-change threshold t1, and the true-change evidence consistency value is compared with the consistency threshold t2. Among them, the pseudo-change threshold is preferably adaptively set based on the statistical distribution of the pseudo-change driving intensity values in the reference object set to characterize the typical fluctuation upper limit of non-real changes; the consistency threshold is preferably set based on the distribution range of the true-change evidence consistency values in the sample data to represent the lower limit requirements for true changes in terms of comprehensive intensity and evidence concentration: When < t1 and ≥ t2, the object unit is determined to be a true-change object; at this time, it indicates that the object unit is weakly affected by driving factors such as phenology, water level, observation geometry, and quality factors, while the comprehensive change intensity is relatively high and the change evidence has obvious dominance, conforming to the statistical characteristics of true land-cover transfer. When ≥ t1 and < t2, the object unit is determined to be a pseudo-change object; at this time, it indicates that the object unit is mainly driven by non-land-cover factors and the change evidence lacks a concentrated dominant feature, and the change is manifested as the superposition result of multi-factor disturbances and should be suppressed. When ≥ t1 and When it is ≥ t2, the object unit to be determined is an object with uncertain changes; such objects have both a relatively high false change driving intensity and a relatively high true change consistency value, indicating a complex situation where driving factors and structural change signals are intertwined, and further analysis is required in subsequent classification or manual verification stages. When < t1 and < t2, the object unit to be determined is a stable object; such objects have neither significant false change driving nor obvious evidence of true changes, and belong to land type units that remain stable over time. The above four types of determinations use a double-threshold two-dimensional decision-making mechanism to jointly constrain and express the false change risk and true change consistency, avoiding misjudgment problems caused by single-index determination, and thus achieving stable differentiation and structured output of change types in a statistical sense.

[0063] In this implementation plan, by constructing a joint determination mechanism for two indicators, namely false change driving intensity and true change evidence consistency, the coordinated constraint and hierarchical differentiation of non-real disturbances and real land type transfers are achieved, avoiding false alarms or missed reports caused by single-index determination, thereby improving the stability, anti-interference ability, and overall determination accuracy of change type recognition.

[0064] Specifically, based on object-level 3D feature sets, the stability of change types is used as the basis for data selection. Supervised training and hierarchical classification are performed to output the category probabilities of forest land, grassland, and wetland. Rule recognition is combined with change characteristics to form land use update and change type labeling results. The specific process is as follows: Using object-level 3D feature sets as input features, combined with the labeled land use tags, object units are spatially matched with corresponding land use types to form supervised training samples. Among them, the labeled land use tags are preferably derived from field survey data, existing thematic resource maps, or historical classification results after manual interpretation and verification. Spatial matching is achieved by performing spatial overlay analysis of object units and labeled patches. When the overlap area ratio between an object unit and a certain labeled category exceeds a preset ratio, the category is assigned to the object unit as a supervised label, thereby ensuring the spatial consistency and label reliability of the training samples. After normalizing the input features of the supervised training samples, a random forest supervised learning algorithm is used for model training, and the model parameters are optimized through cross-validation to construct a hierarchical classification model. Among these, the normalization process preferably adopts a standardization method based on training sample statistics; the random forest model is validated by constructing multiple decision trees and using an out-of-bag error evaluation mechanism; the cross-validation preferably adopts K-fold cross-validation, and the parameters of the number of trees, maximum depth, and number of feature subsets are optimized to improve the model's generalization ability and classification stability. For each object unit, the object-level 3D feature set is input into the hierarchical classification model. For truly changed objects, the post-temporal 3D feature set is input; for stable objects and pseudo-changed objects, the pre-temporal 3D feature set is input. Since truly changed objects have been determined to have undergone land cover shifts or structural changes, the post-change characteristic state should be used as the basis for current land cover determination to ensure consistency between the classification results and the actual resource status. For stable objects and pseudo-changed objects, no true land cover shift has occurred, and their post-temporal features may be affected by phenological, water level, or observational condition disturbances. Therefore, the pre-temporal 3D feature set is continued as the classification input to avoid interference from non-real fluctuations, thus maintaining the continuity and stability of the land cover update process. The classification probabilities of forest land, grassland land, and wetland for each object unit are output. These classification probabilities are the posterior probabilities of the categories output by the random forest model, used to characterize the confidence level of the object unit belonging to each land cover. The category with the highest classification probability is selected as the primary classification label for the object unit; the classification probability value is also retained as an auxiliary indicator for subsequent quality control and uncertainty analysis.For real-world changes, the amplitudes of spectral changes, structural height changes, scattering changes, phenological phase differences, and water level-humidity differences are read and compared with their corresponding thresholds. Rule-based identification is then performed to determine different change types for the real-world changes. Specifically, the rules are as follows: when both structural height and spectral changes exceed the threshold and the height shows a decreasing trend, it is identified as forest degradation or deforestation; when both exceed the threshold and show an increasing trend, it is identified as vegetation growth or recovery; when the water level-humidity difference and scattering changes exceed the threshold and the structural height change is below the threshold, it is identified as wetland water level fluctuation; when both exceed the threshold and phenological phase differences and are below the threshold, it is identified as phenologically driven change; when the water level-humidity difference, structural height change, and spectral change all exceed the threshold, it is identified as wetland-to-land transformation; and other situations where multiple indicators change significantly simultaneously are identified as composite-driven changes. This achieves a systematic classification and labeling of real-world changes.

[0065] In this implementation plan, a supervised learning classification model based on object-level three-dimensional feature sets is constructed to achieve probabilistic hierarchical identification of forest land, grassland and wetland. A rule-based judgment mechanism is established by combining the multidimensional change characteristics of real changed objects, and the statistical learning results are organically integrated with the change mechanism analysis. While ensuring classification accuracy, the mechanism labeling of change types is realized, thereby improving the reliability, consistency and interpretability of land use update results.

[0066] Reference Figure 2 As shown, the second aspect of the present invention provides a dynamic identification and classification system for forest, grassland, and wetland resources based on multi-source remote sensing fusion, applied to the aforementioned dynamic identification and classification method for forest, grassland, and wetland resources based on multi-source remote sensing fusion. The system includes: a data acquisition and processing module for acquiring multi-source remote sensing observation data and performing data preprocessing on the multi-source remote sensing observation data; performing observation quality assessment based on the preprocessed multi-source remote sensing observation data to construct standardized observation data under confidence constraints; an object feature construction module for performing object segmentation and feature aggregation based on the standardized observation data under confidence constraints to form an object-level three-dimensional feature set integrating spectrum, structure, and texture; and a pseudo-change determination module for constructing... A cross-temporal object correspondence is established, and driving features of phenology, water level, geometry, and quality are extracted using a cross-temporal object-level 3D feature set. These driving features are then used for pseudo-change assessment. The cross-temporal object-level 3D feature set is used for true change consistency determination. The stability of change types is distinguished by combining the pseudo-change assessment results and the true change consistency determination results. A change classification module is used to perform supervised training and hierarchical classification based on the object-level 3D feature set, using the stability distinction results of change types as the basis for data filtering. It outputs the category probabilities of forest land, grassland, and wetland, and combines change features for rule recognition, resulting in land use update and change type labeling results.

[0067] This implementation plan establishes a collaborative mechanism that integrates multi-source observation data quality constraints, object-level three-dimensional feature representation, pseudo-change-driven assessment, and consistency determination of true changes. This mechanism enables stable classification and cross-temporal change identification of forest, grassland, and wetland resources. While suppressing the influence of non-real disturbances such as phenology, water level, and observation geometry, it strengthens the identification capability of real land cover transfer signals, thereby improving the accuracy, consistency, and reliability of dynamic monitoring results for engineering applications.

[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0069] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. As those skilled in the art will understand, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for dynamic identification and classification of forest, grassland and wetland resources based on multi-source remote sensing fusion, characterized in that, Includes the following steps: S1. Collect multi-source remote sensing observation data and perform data preprocessing on the multi-source remote sensing observation data; perform observation quality assessment based on the preprocessed multi-source remote sensing observation data, and construct standardized observation data under credibility constraints. S2, based on standardized observation data under credibility constraints, performs object segmentation and feature aggregation to form an object-level three-dimensional feature set that integrates spectrum, structure and texture; The specific process of performing object segmentation and feature aggregation based on standardized observation data under credibility constraints to form an object-level three-dimensional feature set that integrates spectrum, structure, and texture is as follows: The weighted optical multispectral reflectance, SAR backscattering coefficient and point cloud height are used as the segmentation feature input layer. A multi-scale segmentation algorithm is used to perform superpixel segmentation on the weighted data to generate a set of object units. For each object unit in the set of object units, weighted feature statistics are performed. The optical multispectral reflectance and SAR backscattering coefficient of all pixels in the object unit are weighted and averaged based on the multi-source availability confidence value to obtain the spectral vector and SAR backscattering statistics. Sort all weighted point cloud heights within the object cell and perform quantile statistics to obtain the upper bound quantile value and the center quantile value of the point cloud height. Based on optical multispectral reflectance, the near-infrared band is selected as the single-channel gray-level input, the corresponding reflectance is linearly mapped to a finite number of gray levels, and a gray-level co-occurrence matrix is ​​constructed within the object unit region to calculate texture statistics. Based on the spatial contour of the object unit generated by segmentation, the area of ​​the object unit region is calculated by the grid counting method, the perimeter of the object unit is calculated by the boundary cell recognition and boundary length accumulation method, and the shape compactness is calculated based on the area and perimeter using the morphological compactness calculation model. An object-level 3D feature set is constructed by combining spectral vectors, SAR backscatter statistics, point cloud height upper bound quantile, point cloud height center quantile, texture statistics, and shape compactness. S3: Construct cross-time object correspondences and extract driving features of climate, water level, geometry and mass using cross-time object-level 3D feature sets. Use driving features to evaluate pseudo-changes; use cross-time object-level 3D feature sets to determine the consistency of true changes; combine pseudo-change evaluation results with true change consistency determination results to distinguish the stability of change types. The specific process of constructing cross-temporal object correspondences and extracting driving features of climate, water level, geometry, and mass using cross-temporal object-level 3D feature sets is as follows: For each object unit in the object unit set, obtain the object-level three-dimensional feature sets of two time phases, and establish cross-time phase object correspondence using the object space superposition matching method; The reflectance of the near-infrared band and the reflectance of the red band are extracted from the spectral vector. The difference between the reflectance of the near-infrared band and the reflectance of the red band is divided by the sum of the reflectance of the near-infrared band and the reflectance of the red band to obtain the vegetation index of the target unit. The vegetation index of the target unit in the continuous time series is fitted with a smooth spline. The time point corresponding to the peak value is extracted from the fitted vegetation index curve. The difference between the peak time points of the two time phases is calculated to obtain the phenological phase difference of the target unit. The reflectance of the green band and the shortwave infrared band are extracted from the spectral vector. The difference between the green band reflectance and the shortwave infrared band reflectance is divided by the sum of the green band reflectance and the shortwave infrared band reflectance to obtain the water index. The difference between the water indexes of the two time phases is calculated. The difference between the SAR backscatter statistics of the target unit in the two time phases is also calculated. The difference between the water indexes is multiplied by the sum of the difference between the SAR backscatter statistics and one to obtain the water level wetting difference of the target unit. Calculate the difference between the solar altitude angle and the observed zenith angle of the object unit in two time phases, substitute the two differences into the cosine function respectively, and subtract the product of the two cosine functions from one to obtain the observed geometric difference of the object unit. The mean of the multi-source availability confidence values ​​within the target unit is calculated, and the difference between the mean of the multi-source availability confidence values ​​of two time phases is calculated to obtain the quality difference; S4, based on object-level 3D feature sets, uses the stability of change types as the basis for data selection, performs supervised training and hierarchical classification, outputs the category probabilities of forest land, grassland and wetland, and combines change features to perform rule recognition, forming land use update and change type labeling results.

2. The method for dynamic identification and classification of forest, grassland and wetland resources based on multi-source remote sensing fusion according to claim 1, characterized in that, The specific process of acquiring multi-source remote sensing observation data and performing data preprocessing on the multi-source remote sensing observation data is as follows: Real-time acquisition of multi-source remote sensing observation data, including: optical multispectral reflectance, SAR backscattering coefficient, point cloud height, solar elevation angle, observation zenith angle, cloud probability, and shadow probability; The process involves reading the timestamps and coordinate reference information of multi-source remote sensing observation data, unifying the data to the same geographic coordinate system and spatial grid through coordinate projection transformation and spatial resampling, and then resampling and aligning the multi-source remote sensing observation data according to the target resolution using interpolation resampling and raster alignment methods. Cross-source temporal registration is performed based on feature matching registration to obtain a pixel-level registration residual field. Radiometric consistency checks and outlier removal are performed on the optical multispectral reflectance, and radiometric calibration and speckle noise filtering are applied to the SAR backscattering coefficient. Point cloud height is separated into ground and non-ground points through point cloud classification, and height benchmark unification is achieved using elevation benchmark transformation. Solar altitude angle and observation zenith angle parameters are read and rasterized based on the spatial range and resolution of the corresponding image to ensure a one-to-one correspondence between the solar altitude angle, observation zenith angle, and image pixel space, generating an observation geometric auxiliary raster. Dimensionless normalization is performed on the multi-source remote sensing observation data. A resource identification and classification database is established to store the raw and preprocessed multi-source remote sensing observation data.

3. The method for dynamic identification and classification of forest, grassland and wetland resources based on multi-source remote sensing fusion according to claim 2, characterized in that, The specific process of performing observation quality assessment based on preprocessed multi-source remote sensing observation data and constructing standardized observation data under confidence constraints is as follows: Obtain the pixel-level registration residual field, extract the median and median absolute deviation of the pixel-level registration residual, calculate the absolute difference between each pixel-level registration residual and the median for each raster cell, multiply the median absolute deviation by the MAD normal consistency correction coefficient and add a very small positive number as the denominator, divide the absolute difference by the denominator to obtain the robust normalized residual value, and calculate the negative exponent of the robust normalized residual value to obtain the geometric registration stability term. Subtract the cloud probability from one and the shadow probability from one, multiply the two differences and take the square root to obtain the image quality stability term; multiply the geometric registration stability term and the image quality stability term to obtain the multi-source availability confidence value. The multi-source availability confidence value of each grid cell is calculated, and spatial and temporal indices are established with the multi-source remote sensing observation data. The multi-source availability confidence value is used as a weighting factor to perform pixel-level weighting on the corresponding preprocessed optical multispectral reflectance, SAR backscattering coefficient, and point cloud height, so as to obtain standardized observation data under confidence constraints.

4. The method for dynamic identification and classification of forest, grassland and wetland resources based on multi-source remote sensing fusion according to claim 1, characterized in that, The specific process of using driving features to evaluate pseudo-changes is as follows: The phenological phase difference, water level wetting difference, observation geometric difference, and quality difference are combined to form a pseudo-change driving vector for the object unit; The reference object set is obtained by selecting object units in the object unit set whose absolute difference between the upper bound quantile values ​​of point cloud height in two time phases is less than the change threshold; Based on the pseudo-change driving vector of each object unit in the reference object set, the covariance between each component is calculated to form a covariance matrix; The pseudo-change driving vector is transposed and multiplied by the inverse of the covariance matrix to obtain an intermediate result vector. The intermediate result vector is then multiplied by the pseudo-change driving vector to obtain a scalar value. The square root of the scalar value is then taken to obtain the pseudo-change driving intensity value of the object element.

5. The method for dynamic identification and classification of forest, grassland and wetland resources based on multi-source remote sensing fusion according to claim 4, characterized in that, The specific process for determining true change consistency using a cross-time object-level three-dimensional feature set is as follows: For each object unit, three types of variation amplitudes are calculated: the Euclidean distance between the spectral vectors in two time phases is calculated to obtain the spectral variation amplitude; the absolute difference between the upper bound quantile values ​​of the point cloud height in two time phases is calculated to obtain the structural height variation amplitude; and the absolute difference between the SAR backscattering statistics in two time phases is calculated to obtain the scattering variation amplitude. Based on the spectral variation amplitude, structural height variation amplitude, and scattering variation amplitude of all object units, the median and median absolute deviation of each type of variation amplitude are calculated. The difference between each type of variation amplitude and the corresponding median is calculated, and the corresponding median absolute deviation is multiplied by the MAD normality correction coefficient and a very small positive number is added as the denominator. The difference is divided by the denominator to obtain the standardized variation values ​​corresponding to the three types of variation amplitudes. The three types of standardized variation values ​​are then subjected to natural exponentiation and summed. The natural logarithm of the sum is taken to obtain the comprehensive variation intensity value. Divide the results of the three natural indices by the sum of the results to obtain the proportion of spectral change, the proportion of structural change, and the proportion of scattering change. Multiply each proportion of change by its corresponding natural logarithm, sum the results, and take the opposite number to obtain the entropy of consistency of change evidence. Subtract the ratio of the change evidence consistency entropy to the natural logarithm 3 from 1, and multiply it by the comprehensive change intensity value to obtain the true change evidence consistency value of the object unit.

6. The method for dynamic identification and classification of forest, grassland and wetland resources based on multi-source remote sensing fusion according to claim 5, characterized in that, The specific process for distinguishing the stability of change types by combining the results of pseudo-change assessment and the results of consistency determination of true change is as follows: The pseudo-change driving strength value of each object unit is respectively compared with the pseudo-change threshold t1, and the true-change evidence consistency value is compared with the consistency threshold t2: when < t1 and ≥ t2, it is determined that the object unit is a true-change object; when ≥ t1 and < t2, it is determined that the object unit is a pseudo-change object; when ≥ t1 and ≥ t2, it is determined that the object unit is an uncertain-change object; when < t1 and < t2, it is determined that the object unit is a stable object.

7. The method for dynamic identification and classification of forest, grassland and wetland resources based on multi-source remote sensing fusion according to claim 6, characterized in that, The specific process of using object-level 3D feature sets, distinguishing the stability of change types as the basis for data filtering, conducting supervised training and hierarchical classification, outputting the category probabilities of forest land, grassland, and wetland, and combining change features for rule recognition to form land use update and change type labeling results is as follows: Using the object-level 3D feature set as input features, combined with the labeled land class labels, the object unit is spatially matched with the corresponding land class to form supervised training samples; after normalizing the input features of the supervised training samples, the model is trained using the supervised learning algorithm of random forest, and the model parameters are optimized through cross-validation to build a hierarchical classification model. For each object unit, the object-level 3D feature set is input into the hierarchical classification model. Specifically, the post-temporal object-level 3D feature set is input for real-change objects, and the pre-temporal object-level 3D feature set is input for stable objects and pseudo-change objects. The model outputs the classification probabilities of forest, grassland, and wetland for each object unit. The category with the highest classification probability is selected as the main classification label of the object unit. For real objects undergoing changes, the ranges of spectral changes, structural height changes, scattering changes, phenological phase differences, and water level wetting differences are read and compared with the corresponding judgment thresholds to perform rule recognition and obtain different change types of real objects undergoing changes.

8. A multi-source remote sensing fusion-based dynamic identification and classification system for forest, grassland, and wetland resources, employing the multi-source remote sensing fusion-based dynamic identification and classification method for forest, grassland, and wetland resources as described in any one of claims 1-7, characterized in that... include: The data acquisition and processing module is used to acquire multi-source remote sensing observation data and perform data preprocessing on the multi-source remote sensing observation data; Observation quality assessment is performed based on preprocessed multi-source remote sensing observation data, and standardized observation data under credibility constraints are constructed. The object feature construction module is used to perform object segmentation and feature aggregation based on standardized observation data under credibility constraints, forming an object-level three-dimensional feature set that integrates spectrum, structure and texture. The pseudo-change determination module is used to construct the correspondence between objects across time periods, and extract the driving features of climate, water level, geometry and mass using the cross-time object level three-dimensional feature set. The driving features are used to evaluate pseudo-changes; the cross-time object level three-dimensional feature set is used to determine the consistency of true changes; and the stability of change types is distinguished by combining the pseudo-change evaluation results and the true change consistency determination results. The classification and change recognition module is used to perform supervised training and hierarchical classification based on the object-level three-dimensional feature set and the stability of change type as the basis for data screening. It outputs the category probabilities of forest land, grassland and wetland, and combines change features to perform rule recognition, forming land use update and change type labeling results.

Images