An image processing method and system for dynamic monitoring of a water body

Abstract

The application discloses an image processing method for water body dynamic monitoring, and relates to the technical field of remote sensing image processing, and comprises the following steps: obtaining a reference map sequence by acquiring original observation data flow; taking a hydrological periodic law as a soft time sequence constraint anchor point to code and map the space-time dynamics of the reference map sequence; performing an adversarial scene drift suppression to extract scene-independent essential water body response features, and identifying a high-ambiguity area; performing multiple parallel inferences through a random activation masking mechanism, and differentiating and soft-suppressing each spatial position according to the prediction distribution and change type marking of the inference result; performing analysis at a primitive level, a target entity level and a scene semantic level, respectively calculating the global uncertainty indexes of each atlas, determining a contribution coefficient, and generating a water body dynamic monitoring result. The application can effectively improve the cross-domain generalization capability of multi-source remote sensing images, suppress uncertainty errors, and realize high-precision dynamic monitoring of water body changes in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to an image processing method for dynamic monitoring of water bodies. Background Technology

[0002] Remote sensing technology has become a primary means of monitoring water dynamics due to its advantages of large-scale and periodic observation. In actual monitoring, multi-source remote sensing data, such as optical and synthetic aperture radar, are usually integrated. However, differences in the imaging mechanisms and radiometric response characteristics of different sensors lead to radiometric inconsistencies among multi-source images. Existing radiometric normalization methods are insufficient in terms of automatic screening of pseudo-invariant feature spatial locations and model robustness, and are prone to introducing systematic errors during change detection, affecting the reliability of the detection results.

[0003] Existing water change detection methods typically rely on labeled samples from specific regions for training. However, in cross-regional and cross-observation gait scenarios, differences in feature distribution lead to insufficient generalization ability and difficulty in guaranteeing detection accuracy. Furthermore, existing domain adaptation methods struggle to effectively maintain the discriminative power of water category features across different regions and observation gaits when processing remote sensing images.

[0004] In addition, seasonal hydrological fluctuations in water dynamic monitoring are easily confused with real water change events. Existing change detection methods have relatively simple processing methods for high uncertainty areas, and it is difficult to simultaneously suppress pseudo-changes and preserve real changes, which affects the overall performance of water change detection in complex environments. Summary of the Invention

[0005] The main objective of this invention is to provide an image processing method for dynamic water monitoring, aiming to solve the technical problem that existing image processing methods suffer from the inconsistency of multi-source image radiation, insufficient generalization of gait features across regions and observations, and inaccurate suppression of false changes, which lead to the difficulty in achieving both accuracy and robustness in water change detection under complex environments.

[0006] In a first aspect, the present invention provides an image processing method for dynamic monitoring of water bodies, comprising:

[0007] The raw observation data stream is acquired, and a baseline map sequence is obtained after radiometric normalization anchoring and geometric registration.

[0008] The baseline map sequence is input into a time-series coding network, and the hydrological periodicity is used as a soft time-series constraint anchor point to encode and map the spatiotemporal dynamics of the baseline map sequence, thereby obtaining the primary water body disturbance identification results carrying change type labels.

[0009] Adversarial scene drift suppression is applied to the primary water disturbance identification results, scene-independent essential water body response features are extracted, highly ambiguous regions are identified, and scene-adaptive water disturbance identification results are obtained.

[0010] For highly ambiguous regions, multiple parallel inferences are performed using a random activation masking mechanism. Differentiated soft suppression is applied to each spatial location based on the predicted distribution and change type label of the inference results. The soft suppression results are then combined with the detection results of the remaining regions to obtain refined water disturbance identification results.

[0011] The refined water disturbance identification results are analyzed at the primitive level, target entity level, and scene semantic level, and the global uncertainty index is calculated.

[0012] The contribution coefficient is adaptively determined based on the global uncertainty index, and weighted fusion is performed to generate dynamic water monitoring results.

[0013] Optionally, the reference map sequence obtained after radiometric normalization anchoring and geometric registration includes:

[0014] Perform initial spatial matching on the raw observation data stream to obtain coarsely aligned observation frames;

[0015] Spectral stability index is calculated for each spatial location in the coarsely aligned observation frame to determine the spatial location of candidate pseudo-invariant features;

[0016] For candidate pseudo-invariant feature spatial locations, calculate the temporal and frequency domain co-coefficients between them and their spatial neighborhood spatial locations. Retain candidate spatial locations whose temporal and frequency domain co-coefficients both satisfy the preset conditions as pseudo-invariant feature spatial locations.

[0017] A radiometric normalization regression model between the image to be corrected and the reference image is established based on the spatial location of pseudo-invariant features. Radiometric normalization processing is performed on the coarsely aligned observation frames to obtain a radiometrically consistent image.

[0018] Feature points are extracted from radiometrically consistent images, and fine geometric registration is performed based on these feature points to obtain a reference image sequence.

[0019] Optionally, using the hydrological periodicity as a soft temporal constraint anchor point, the spatiotemporal dynamics of the baseline map sequence are encoded and mapped, including:

[0020] Spatial features are extracted from each observed gait image in the baseline image sequence and fused to obtain the spatial representation information flow corresponding to each observed gait.

[0021] The spatial representation information streams of adjacent observation gaits are subtracted point by point to obtain the difference perturbation field. The difference perturbation field is stacked and coupled with the original spatial representation information streams of the two adjacent frames in the feature channel dimension to obtain the dual observation gait interaction feature body of difference information and original spatial information.

[0022] Hydrological periodicity data is encoded into temporal position embedding vectors and injected into the temporal position encoding layer of a temporal coding network as soft constraint anchors. This process is used to temporally encode the dual-observation gait interaction feature body, resulting in temporal context features constrained by hydrological periodic semantics.

[0023] Within the feature space, the change displacement vector is calculated based on the temporal context features of adjacent observed gaits. The change displacement vector is then jointly judged in terms of change amplitude and change direction to generate a change type label, thus obtaining the primary water disturbance identification result carrying the change type label.

[0024] Optionally, the scene-adaptive water disturbance recognition results include:

[0025] Remote sensing images of labeled water body change samples are used as reference scene data, and unlabeled real-time observation data of the area to be monitored are used as target detection scene data.

[0026] Based on benchmark reference scene data and target detection scene data, and constrained by the cross-domain semantic feature center alignment of water bodies, non-water bodies and water-land transition zones, the feature extraction network is trained by adversarial scene drift to obtain a scene adaptation feature network.

[0027] The spatial representation information flow of the baseline map sequence and the primary water disturbance identification results are jointly input into the scene adaptation feature network to obtain the essential water body response feature map;

[0028] The reliability of feature membership between each spatial location in the essential water body response feature map and the semantic feature center of each category in the benchmark scene is calculated to mark highly ambiguous regions and obtain the scene-adaptive water body disturbance recognition results.

[0029] Optionally, the refined water disturbance identification results include:

[0030] Multiple forward inferences are performed on highly ambiguous regions using a random neuron masking mechanism to obtain multiple sets of predicted probability distributions.

[0031] The response distribution dispersion of each spatial location is calculated based on multiple sets of predicted probability distributions, and the initial pseudo-change suppression weight for that location is generated based on the response distribution dispersion.

[0032] The initial pseudo-change suppression weights are adjusted according to the change type label to obtain the secondary pseudo-change suppression weights;

[0033] Based on the secondary pseudo-change suppression weight, the response results of each spatial location in the highly ambiguous region are subjected to soft suppression processing. The soft suppression results are then combined with the scene adaptation water disturbance identification results of the high-confidence region to obtain the refined water disturbance identification results.

[0034] Optionally, the refined water disturbance identification results are parsed at the primitive level, target entity level, and scene semantic level, respectively, including:

[0035] The water body disturbance identification results are refined to extract basic-level water body change information to obtain a basic-level water body change map.

[0036] Eight-neighbor spatial connectivity analysis was performed on the elementary-level water body change map. Based on the spatial geometric properties of each connectivity region, the change type of each connectivity region was spatially refined to obtain the target entity-level water body change map.

[0037] Based on the target entity-level water body change map, the change type is semantically refined by combining geospatial prior information to obtain the scene semantic-level water body change map.

[0038] Based on the region type of each spatial location, corresponding uncertainty indicators are determined, resulting in global uncertainty indicators for three different granularity detection results.

[0039] Optionally, the generated water body dynamic monitoring results include:

[0040] Based on the global uncertainty index, the statistical mean of the spatial location uncertainty index corresponding to each granularity is calculated. The initial contribution coefficient of each granularity is constructed by the reciprocal of the statistical mean, and the initial contribution coefficients of the three granularities are normalized.

[0041] When the dispersion of the three granularity normalized initial contribution coefficients is not higher than the upper limit of dispersion, and the average uncertainty index of the three granularities is not higher than the preset upper limit of uncertainty, the scene semantic level water body change map is used as the output of the water body dynamic monitoring result.

[0042] When the dispersion of the normalized initial contribution coefficients of the three particle sizes is higher than the upper limit of dispersion, or when there is a scale in the average uncertainty index of the three particle sizes that is higher than the preset upper limit of uncertainty, the result of weighted fusion of the detection results of the three particle sizes according to the normalized initial contribution coefficients is used as the output of the dynamic monitoring result of the water body.

[0043] Optionally, the output of dynamic water body monitoring results includes:

[0044] The change type markers of each spatial location in the target entity-level water change map and the scene semantic-level water change map are mapped to the corresponding change probability values. The map is then rasterized into a probability map with the same spatial resolution as the primitive-level water change map. The predicted probability values ​​of each spatial location at the corresponding positions in the three granularities are multiplied by the normalized initial contribution coefficient of the corresponding granularity and then summed to obtain the probability value of each spatial location. Based on the probability values, the dynamic monitoring results of the water body are generated.

[0045] Optionally, determining the spatial location of candidate pseudo-invariant features and the spatial location of pseudo-invariant features includes:

[0046] Calculate the temporal variation coefficient of spectral reflectance at each spatial location and in each band of the coarsely aligned observation frame, obtain the spectral stability index based on the temporal variation coefficient, and compare and determine the spatial location of candidate pseudo-invariant features.

[0047] For each candidate pseudo-invariant feature spatial location, a time-series spectral sequence is extracted, and the cooperability coefficients in the time and frequency domains are calculated based on the time-series spectral sequence to determine the pseudo-invariant feature spatial location.

[0048] Secondly, the present invention provides an image processing system for dynamic monitoring of water bodies, comprising:

[0049] The data preprocessing module is used to acquire the raw observation data stream, and after radiometric normalization anchoring and geometric registration, a reference map sequence is obtained.

[0050] The encoding module is used to input the reference map sequence into the time-series encoding network, use the hydrological periodic law as a soft time-series constraint anchor point, encode and map the spatiotemporal dynamics of the reference map sequence, and obtain the primary water body disturbance identification result carrying the change type label;

[0051] The scene adaptation module is used to perform adversarial scene drift suppression on the initial water disturbance identification results, extract scene-independent essential water response features, identify highly ambiguous regions, and obtain scene-adapted water disturbance identification results.

[0052] The uncertainty reasoning module is used to perform multiple parallel reasonings on highly ambiguous regions through a random activation masking mechanism. Based on the predicted distribution and change type label of the reasoning results, differentiated soft suppression is applied to each spatial location. The soft suppression results are then merged with the detection results of the remaining regions to obtain refined water disturbance identification results.

[0053] The processing module is used to parse the refined water disturbance identification results at the primitive level, target entity level, and scene semantic level, and calculate the global uncertainty index.

[0054] The output module is used to adaptively determine the contribution coefficient based on the global uncertainty index, perform weighted fusion, and generate dynamic water monitoring results.

[0055] Thirdly, the present invention provides an electronic device, comprising:

[0056] Memory, used to store computer programs;

[0057] A processor is used to execute computer programs to implement the aforementioned image processing method for dynamic monitoring of water bodies.

[0058] Fourthly, the present invention provides a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned image processing method for dynamic monitoring of water bodies.

[0059] The beneficial effects of this invention are as follows: This invention effectively solves the coupling problem between inconsistencies in radiometrics from multi-source images and insufficient generalization of gait features across regions and observations. The accuracy of radiometric normalization is improved, and systematic errors are suppressed; the model can distinguish between periodic natural fluctuations in water bodies and real change events, reducing the interference of seasonal hydrological fluctuations on the detection results; it can still maintain effective water body change detection capabilities even under unlabeled target domain data conditions; the detection results simultaneously achieve both spurious change suppression and real change preservation; and the dynamic water body monitoring results possess both spatial accuracy and semantic integrity. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0061] Figure 1 This is a schematic flowchart of an image processing method for dynamic monitoring of water bodies provided in an embodiment of the present invention.

[0062] Figure 2 This is a flowchart of radiometric normalization anchoring and geometric registration for an image processing method for dynamic water monitoring provided in an embodiment of the present invention.

[0063] Figure 3 This is a flowchart illustrating the process of refining water disturbance identification results acquisition in an image processing method for dynamic water monitoring, as provided in an embodiment of the present invention.

[0064] Figure 4 This is a structural block diagram of an image processing system for dynamic water monitoring provided in an embodiment of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, 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.

[0066] Reference Figure 1This invention discloses an image processing method for dynamic monitoring of water bodies, which may include:

[0067] S1: Obtain the raw observation data stream, and obtain the baseline map sequence after radiometric normalization anchoring and geometric registration.

[0068] The raw observation data stream includes at least one of optical satellite imagery and synthetic aperture radar (SAR) imagery. Optical satellite imagery provides spectral information in the visible, near-infrared, and short-wave infrared bands, effectively distinguishing water bodies from non-water features; SAR imagery has all-weather observation capabilities and can acquire water distribution information even under cloud cover conditions, compensating for data gaps in optical imagery.

[0069] Specifically, the flowchart for radiation normalization anchoring and geometric registration is as follows: Figure 2 As shown, it includes the following sub-steps:

[0070] S1.1: Perform initial spatial matching on the original observation data stream to obtain coarsely aligned observation frames.

[0071] In one embodiment, the geographic coordinate information or ground control points in the image header file are used to perform preliminary spatial alignment of each observation gait image data stream to eliminate macroscopic geometric deviations between data and obtain coarsely aligned observation frames.

[0072] S1.2: Calculate the spectral stability index for each spatial location in the coarsely aligned observation frame to determine the spatial location of candidate pseudo-invariant features.

[0073] Specifically, for each spatial location in the coarse-aligned observation frame, the temporal variation coefficient of the spectral reflectance of each band is calculated on all gait images observed within the preset monitoring period. The weighted average of the temporal variation coefficients of each band is taken as the spectral stability index. The weight of each band can be set empirically based on the contribution of the band to water body identification, or an equal-weighted average can be used. Spatial locations with a spectral stability index lower than a first threshold are selected as candidate pseudo-invariant feature spatial locations. The first threshold can be set empirically or adaptively determined through statistical analysis of the stability index distribution.

[0074] S1.3: For candidate pseudo-invariant feature spatial locations, calculate the temporal and frequency domain co-coefficients between them and their spatial neighbors. Retain candidate spatial locations whose temporal and frequency domain co-coefficients both satisfy the preset conditions as pseudo-invariant feature spatial locations.

[0075] Specifically, for each candidate pseudo-invariant feature spatial location, its temporal spectral sequence is extracted, and the spatial location within an N×N spatial neighborhood window centered on the candidate spatial location is taken as the spatial neighborhood spatial location. N can be set empirically based on the image spatial resolution.

[0076] In the time domain, the Pearson correlation coefficient between the temporal spectral sequences of the candidate spatial location and its neighboring spatial locations is calculated as the temporal coherence coefficient. A higher temporal coherence coefficient indicates that the candidate spatial location and its neighboring spatial locations have similar temporal spectral variation trends, and are more likely to be true pseudo-invariant feature spatial locations. In the frequency domain, Discrete Fourier Transforms are performed on the temporal spectral sequences of the candidate spatial location and its neighboring spatial locations respectively. The reciprocal of the frequency difference between their dominant frequency components plus a minimum smoothing term is calculated as the frequency domain coherence coefficient to avoid division-by-zero anomalies when the dominant frequencies of the two spatial locations are completely identical. A higher frequency domain coherence coefficient indicates that the candidate spatial location and its neighboring spatial locations have similar spectral periodic variation structures.

[0077] For each candidate pseudo-invariant feature spatial location, if both the time-domain coherence coefficient and the frequency-domain coherence coefficient are higher than a preset time-domain threshold, the candidate spatial location is retained as a pseudo-invariant feature spatial location; otherwise, it is discarded. The preset time-domain threshold and preset frequency-domain threshold can be set empirically based on the ground cover distribution characteristics of the monitoring area, or adaptively determined through statistical analysis of the distribution of each metric value.

[0078] The technical advantages of the above two-step screening mechanism are as follows: First, based on the spectral stability index, spatial locations such as vegetation, water bodies, and bare soil that change over time are eliminated to obtain candidate pseudo-invariant feature spatial locations. Second, through joint screening using time-domain and frequency-domain synergy coefficients, isolated anomalous spatial locations and spatial locations with unstable spectral periodic structures are eliminated to obtain pseudo-invariant feature spatial locations. Compared to single-threshold screening, this two-step progressive screening improves the purity of pseudo-invariant feature spatial locations, thereby enhancing the accuracy of the radiation normalization regression model and reducing systematic errors.

[0079] S1.4: Based on the spatial location of pseudo-invariant features, a radiometric normalization regression model is established between the image to be corrected and the reference image. Radiometric normalization processing is performed on the coarsely aligned observation frames to obtain a radiometrically consistent image.

[0080] Specifically, based on the selected pseudo-invariant feature spatial locations, the spectral reflectance of the image to be corrected and the reference image at these spatial locations is extracted. Using the spectral reflectance of the reference image as the dependent variable and the spectral reflectance of the image to be corrected as the independent variable, linear regression models are established for each band. In accordance with the minimum sample size requirement for linear regression in statistics, the number of pseudo-invariant feature spatial locations participating in the regression for each band is no less than 30 to ensure the statistical reliability of the regression model.

[0081] Furthermore, the regression coefficients of each band are estimated using the least squares method, and the regression model is applied to all spatial locations of the observed gait images to be corrected, thus obtaining radiometrically consistent images.

[0082] S1.5: Extract feature points from the radiometrically consistent image, perform fine geometric registration based on the feature points, and obtain the reference image sequence.

[0083] Furthermore, feature points are extracted from the radiometrically consistent images, and feature point matching algorithms (such as SIFT, ORB, or phase correlation algorithms) are used to perform fine geometric registration on each observed gait image to eliminate the residual geometric deviation after coarse registration, improve the spatial alignment accuracy of each observed gait image, and obtain a reference image sequence.

[0084] S2: Input the baseline map sequence into the time-series coding network, use the hydrological periodicity as a soft time-series constraint anchor point, encode and map the spatiotemporal dynamics of the baseline map sequence, and obtain the primary water body disturbance identification results carrying the change type label.

[0085] Specifically, step S2 includes the following sub-steps:

[0086] S2.1: Extract spatial features from each observed gait image in the baseline image sequence and fuse them to obtain the spatial representation information flow corresponding to each observed gait.

[0087] In one embodiment, a shared-weight encoder extracts multi-scale spatial features frame by frame from each observed gait image in the reference image sequence to obtain shallow texture feature maps and deep semantic feature maps for each observed gait; feature pyramid fusion is performed on the shallow texture feature maps and deep semantic feature maps to obtain the spatial representation information flow corresponding to each observed gait.

[0088] S2.2: Perform point-by-point subtraction on the spatial representation information streams of adjacent observation gaits to obtain the difference perturbation field. Stack and couple the difference perturbation field with the original spatial representation information streams of the two adjacent frames in the feature channel dimension to obtain the dual observation gait interaction feature body of difference information and original spatial information.

[0089] Specifically, the spatial representation information streams of adjacent observation gaits are subtracted element by element to obtain a difference perturbation field reflecting the semantic and structural changes between the two observation gaits. The difference perturbation field and the original spatial representation information streams of the two adjacent observation gaits are stacked and spliced ​​along the channel dimension to form a dual-observation gait interaction feature body containing difference information and original spatial information of the dual observation gaits. The dual-observation gait interaction feature body retains the original spatial semantic information of each observation gait and the change information between the observation gaits, providing an input expression that takes into account both static features and dynamic changes for subsequent temporal coding.

[0090] S2.3: Encode the hydrological periodicity data into a temporal position embedding vector, inject it into the temporal position encoding layer of the temporal encoding network as a soft constraint anchor point, and perform temporal encoding on the dual-observation gait interaction feature body to obtain temporal context features constrained by hydrological periodic semantics.

[0091] Furthermore, historical hydrological data of the monitoring area are acquired, and the periodic variation patterns of water area and water level are extracted to construct hydrological periodicity data. The hydrological periodicity data is encoded into a time-series position embedding vector with the same dimension as the time-series position encoding. The time-series position embedding vector is added element-wise to the original time-series position encoding in the time-series encoding network to guide the network to learn time-series feature expressions that conform to the hydrological periodicity in a soft constraint form. Based on the time-series position encoding after injecting the hydrological periodicity, the dual-observation gait interaction feature body is time-series encoded to obtain time-series context features constrained by the hydrological periodicity semantics.

[0092] S2.4: Calculate the change displacement vector for the temporal context features of adjacent observed gaits in the feature space, jointly determine the change amplitude and change direction of the change displacement vector, generate change type label, and obtain the primary water disturbance identification result carrying the change type label.

[0093] Specifically, the difference between the temporal context features of adjacent observed gaits is calculated to obtain the displacement vectors of each spatial location. The magnitude of the displacement vector measures the magnitude of the change, and the direction of the change is represented by the direction of the change. The direction of the displacement vector reflects the evolution trend of the water state in the feature space. The displacement vectors of different change types exhibit different directional distribution patterns in the feature space. The classification head in the temporal coding network is used to determine the direction of the displacement vectors and identify their change type, thereby supporting the generation of subsequent change type labels.

[0094] Furthermore, based on the hydrological periodicity incorporated into the time-series coding network, the amplitude sequence of the changing displacement vector on the time axis is matched with the periodicity characteristics of the hydrological periodicity. Based on the change amplitude, change direction, and time-series matching results, change type labels are generated for each spatial location. Change type labels include, but are not limited to, periodic fluctuation, sudden expansion, gradual retreat, and boundary fluctuation. Based on the change type labels and the joint judgment results of change amplitude and change direction, the primary water disturbance identification results carrying change type labels are obtained.

[0095] By introducing change type markers to semantically distinguish the detection results, structured semantic guidance is provided for subsequent adversarial scene drift suppression and differential soft suppression, improving the targeting of subsequent steps for the spatial location of different change types.

[0096] S3: Perform adversarial scene drift suppression on the primary water disturbance identification results, extract scene-independent essential water response features, identify highly ambiguous regions, and obtain scene-adaptive water disturbance identification results.

[0097] Specifically, step S3 includes the following sub-steps:

[0098] S3.1: Use the remote sensing images of labeled water change samples as the benchmark reference scene data, and use the unlabeled real-time observation data of the area to be monitored as the target detection scene data.

[0099] S3.2: Based on the benchmark reference scene data and target detection scene data, and with the cross-domain semantic feature center alignment of water bodies, non-water bodies and water-land transition zones as constraints, the feature extraction network is trained by adversarial scene drift to obtain the scene adaptation feature network.

[0100] Specifically, within the feature space of the benchmark reference scenario, the mean values ​​of the domain features of three semantic categories—water bodies, non-water bodies, and water-land transition zones—are calculated based on the annotation information to construct the semantic prototype of the benchmark reference scenario; within the feature space of the target detection scenario, the mean values ​​of the feature vectors of each category are calculated by weighting the prediction results of the feature extraction network with the prediction confidence as the weight, to construct the semantic prototype of the target detection scenario.

[0101] Furthermore, the category alignment error between the benchmark reference scene semantic prototype and the target detection scene semantic prototype is used as the cross-domain semantic feature center alignment constraint. The feature extraction network is optimized by combining the cross-domain semantic feature center alignment constraint and the adversarial training loss. In the adversarial training, the domain discriminator aims to distinguish the feature source scene, while the feature extraction network aims to confuse the domain discriminator. The alternating optimization makes the feature extraction network gradually converge to extract the essential water body response features that are independent of the scene, thus obtaining the scene-adaptive feature network.

[0102] S3.3: Input the spatial representation information flow of the baseline map sequence and the primary water disturbance identification results into the scene adaptation feature network to obtain the essential water response feature map.

[0103] Furthermore, multi-scale features are extracted from each observed gait image in the baseline image sequence to obtain a multi-scale spatial representation information flow. This information flow is then input into the scene adaptation feature network along with the primary water disturbance identification results. The change type labels in the primary water disturbance identification results are converted into a spatial attention weight map through a learnable embedding mapping. The spatial attention weight map is used to weight the intermediate layer features of the scene adaptation feature network. The essential water response features at each scale are then fused across scales to obtain the essential water response feature map.

[0104] S3.4: Calculate the feature membership reliability between each spatial location in the essential water body response feature map and the semantic feature center of each category in the benchmark reference scene, in order to mark highly ambiguous areas and obtain the scene-adapted water body disturbance recognition results.

[0105] Specifically, for each spatial location in the essential water body response feature map, the alignment degree between the feature distribution of the reference scene and the target detection scene is measured based on the distance between its feature vector and the semantic feature center of each category of the reference scene semantic prototype. The normalized value of the distance closest to the category semantic center is used as the feature membership reliability of the spatial location. The higher the feature membership reliability, the better the alignment degree between the feature of the spatial location and the reference scene.

[0106] Furthermore, a reliability threshold is set. Spatial locations with feature membership reliability below the reliability threshold are classified as high-ambiguity regions, while spatial locations with feature membership reliability not below the reliability threshold are classified as high-confidence regions. This indicates that the essential water body response features of these spatial locations are highly aligned with the corresponding category semantic feature centers of the benchmark scene, and the impact of scene drift is relatively small, resulting in high reliability of their water disturbance identification results. Spatial locations with feature membership reliability below the reliability threshold are classified as high-ambiguity regions. This indicates that the essential water body response features of these spatial locations deviate significantly from the category semantic feature centers of the benchmark scene, and the cross-domain alignment effect is poor, resulting in significant uncertainty in their water disturbance identification results. These regions require special attention or differentiated processing in subsequent steps.

[0107] Furthermore, the feature membership reliability labeling results are combined with the detection results of the essential water body response feature map to obtain the scene-adaptive water body disturbance identification results with confidence labels.

[0108] The reliability threshold can be empirically set based on the characteristic distribution of the target detection scenario data, or adaptively determined through statistical analysis of the reliability distribution of each spatial location feature. For example, the mean, median, or a specified quantile (such as the top 25% quantile) of the reliability distribution of all spatial location features in the target detection scenario can be used as the reliability threshold. This allows the division between highly ambiguous and highly reliable regions to be dynamically adjusted according to the actual characteristic distribution of the target detection scenario data, avoiding the problem of insufficient adaptability of a fixed threshold in different scenarios.

[0109] By introducing cross-domain semantic feature center alignment constraints, cross-domain feature alignment is achieved at the category level, which improves the reliability of the model in water change detection scenarios without labeled targets. By evaluating the reliability of scene adaptation results per spatial location, the detection results are divided into highly ambiguous regions and highly reliable regions, providing a targeted basis for subsequent differentiated soft suppression.

[0110] S4: For highly ambiguous regions, multiple parallel inferences are performed using a random activation masking mechanism. Based on the predicted distribution and change type labels of the inference results, differentiated soft suppression is applied to each spatial location. The soft suppression results are then merged with the detection results of the remaining regions to obtain refined water disturbance identification results.

[0111] Specifically, the flowchart for obtaining refined water disturbance identification results is as follows: Figure 3 As shown, it includes the following sub-steps:

[0112] S4.1: For highly ambiguous regions, multiple forward inferences are performed using a random neuron masking mechanism to obtain multiple sets of predicted probability distributions.

[0113] In this step, for spatial locations marked as highly ambiguous regions in the scene adaptation water disturbance identification results, a random neuron masking mechanism is enabled. The feature extraction network is subjected to multiple independent random forward inferences with a preset inactivation probability. All predicted probability distributions of each spatial location in the multiple inferences are retained, resulting in multiple sets of predicted probability distributions for each spatial location.

[0114] The preset inactivation probability can be set empirically based on the network structure and target detection scenario data characteristics; the number of inferences can be set empirically based on the statistical stability requirements of uncertainty estimation and computational resources.

[0115] S4.2: Calculate the response distribution dispersion of the output at each spatial location based on multiple sets of predicted probability distributions, and generate the initial pseudo-change suppression weight for that location based on the response distribution dispersion.

[0116] Specifically, the mean of the prediction probability is calculated along the inference number dimension for multiple prediction probability distributions at each spatial location to obtain the average prediction probability distribution for each spatial location; the response distribution dispersion of each spatial location is calculated according to the Shannon entropy formula based on the average prediction probability distribution; and the difference between the response distribution dispersion of each spatial location and the preset response distribution dispersion threshold is normalized and used as the initial pseudo-change suppression weight for each spatial location.

[0117] The preset response distribution dispersion threshold can be set empirically based on the uncertainty distribution characteristics of the target detection scenario data, or adaptively determined by taking the preset quantile of the response distribution dispersion distribution at each spatial location.

[0118] S4.3: Adjust the initial pseudo-change suppression weights according to the category correlation based on the change type label to obtain the secondary pseudo-change suppression weights.

[0119] Specifically, for spatial locations labeled as periodic fluctuations, the pseudo-change basic suppression weight is increased by an amount negatively correlated with the response distribution dispersion; for spatial locations labeled as sudden expansions or gradual retreats, the pseudo-change basic suppression weight is decreased by an amount negatively correlated with the response distribution dispersion; for spatial locations labeled as boundary fluctuations, the difference between the response distribution dispersion and the mean of the response distribution dispersion in the spatial neighborhood is calculated, and the pseudo-change basic suppression weight is increased when the difference is greater than zero, decreased when it is less than zero, and kept unchanged when it is equal to zero; after the above differential adjustment, the secondary pseudo-change suppression weight corresponding to each spatial location is obtained.

[0120] Compared to methods that uniformly apply fixed suppression to highly ambiguous regions, this invention adopts a differentiated linkage strategy for different change types. For periodic fluctuations, the suppression weight is increased to suppress seasonal pseudo-changes; for sudden expansions and gradual retreats, the suppression weight is decreased to preserve real changes; and for boundary fluctuations, the suppression direction is adaptively determined based on the difference in the dispersion value of the relative response distribution in the spatial neighborhood. This refines the water disturbance identification results while simultaneously suppressing pseudo-changes and preserving real changes.

[0121] S4.4: Based on the secondary pseudo-change suppression weight, the response results of each spatial location in the highly ambiguous region are subjected to soft suppression processing. The soft suppression results are then combined with the scene adaptation water disturbance identification results in the high-confidence region to obtain the refined water disturbance identification results.

[0122] Furthermore, based on the secondary pseudo-change suppression weights corresponding to each spatial location, the detection probability values ​​of each spatial location within the highly ambiguous region are subjected to soft suppression processing to obtain the soft suppression detection results of the highly ambiguous region; the detection results of the high-confidence region are extracted from the scene-adapted water disturbance identification results and merged with the soft suppression detection results of the highly ambiguous region according to spatial location to obtain the refined water disturbance identification results covering the complete monitoring area.

[0123] S5: The refined water disturbance identification results are analyzed at the primitive level, target entity level, and scene semantic level, respectively, and the global uncertainty index is calculated.

[0124] S5.1: Extract basic-level water change information from the refined water disturbance identification results to obtain a basic-level water change map.

[0125] In one embodiment, the predicted probability of water body change at each spatial location in the refined water disturbance identification result is compared with a preset change probability threshold; spatial locations with a predicted probability of water body change higher than the preset change probability threshold are marked as changed spatial locations, and spatial locations with a predicted probability of water body change not higher than the preset change probability threshold are marked as unchanged spatial locations, thus obtaining a primitive-level water body change map.

[0126] The preset change probability threshold can be set based on experience with the characteristics of land features in the monitoring area, or adaptively determined through statistical analysis of the predicted probability distribution.

[0127] S5.2: Perform eight-neighbor spatial connectivity analysis on the elementary-level water body change map, and spatially refine the change type of each connectivity region based on the spatial geometric attributes of each connectivity region to obtain the target entity-level water body change map.

[0128] Specifically, the spatial locations of changes in the elementary-level water body change map are marked as connected regions. An eight-neighbor spatial connection method is used to divide spatially connected changes into the same connected region. The eight-neighbor spatial connection method considers the spatial connection relationships in the vertical, horizontal, left-right, and diagonal directions, and is suitable for actual scenarios where water body boundaries have diagonal connections.

[0129] Furthermore, spatial geometric attributes, including area, perimeter, and shape index, are extracted from each connected region. Based on these spatial geometric attributes and combined with the change type markers of the corresponding regions in the primary water disturbance identification results, the change types of each connected region are spatially refined according to the correspondence rules between area, perimeter, and shape index and each change type, resulting in a target entity-level water change map.

[0130] The shape index is defined as the ratio of the perimeter of the connected region to the square root of its area, and is used to describe the shape complexity of the connected region; the area and perimeter reflect the scale characteristics of the connected region.

[0131] S5.3: Based on the target entity-level water body change map, the change types are semantically refined by combining geospatial prior information to obtain the scene semantic-level water body change map.

[0132] Specifically, the system acquires prior geospatial information about the monitoring area, including hydrological unit boundaries and topographic data. The prior geospatial information is pre-loaded during system initialization; hydrological unit boundaries are derived from watershed delineation data; and topographic data is obtained based on a digital elevation model.

[0133] Furthermore, based on the boundaries of hydrological units, the spatial rationality of each connecting region in the target entity-level water change map is evaluated, and connecting regions whose spatial location is completely outside the boundaries of hydrological units are marked as suspected noise regions and removed. Based on topographic data, the topographic background of the remaining change regions is analyzed, and combined with topographic features, the change type markings of each change region are semantically refined according to the elevation and water flow direction to obtain a scene semantic-level water change map.

[0134] S5.4: Based on the region type of each spatial location, determine the corresponding uncertainty index to obtain the global uncertainty index of the three detection results with different granularities.

[0135] Specifically, for spatial locations in highly ambiguous regions, the dispersion value of the response distribution of the predicted probability distribution obtained from multiple random forward inferences in step S4 is used as the uncertainty index; for spatial locations in highly reliable regions, since there is no corresponding predicted probability distribution because the region has not undergone multiple random forward inferences, the result of subtracting the normalized feature membership reliability of each spatial location in step S3.4 from 1 is used as the uncertainty index for that spatial location, ensuring full spatial location coverage.

[0136] Since the two types of uncertainty indicators have different dimensional meanings, before calculating the subsequent mean, the dispersion value of the spatial location of the highly ambiguous region and the reliability complement value of the spatial location of the highly reliable region are normalized to ensure that they have a unified distribution range and ensure dimensional consistency.

[0137] For both target entity-level and scene semantic-level results, regardless of whether the spatial location belongs to a highly ambiguous or highly reliable region, the normalized uncertainty index is determined using the aforementioned correspondence method. The mean of all spatial location uncertainty indices within the corresponding connected region is then used as the region-level uncertainty index. Specifically, for spatial locations that have been geospatially corrected during scene semantic-level refinement, their uncertainty indices retain their original values ​​before correction to avoid introducing additional uncertainty estimation biases through geospatial prior correction. The technical effect of this multi-granularity processing is that, through progressive processing at the primitive level, target entity level, and scene semantic level, the water change detection results simultaneously possess spatial location-level accuracy, connected region spatial geometric attributes, and geospatial semantic information; and based on prediction uncertainty, a full spatial location coverage uncertainty index is constructed, providing a unified reliability quantification basis for subsequent adaptive fusion.

[0138] S6: Based on the global uncertainty index, the contribution coefficient is adaptively determined, and weighted fusion is performed to generate dynamic water monitoring results.

[0139] S6.1: Based on the global uncertainty index, calculate the statistical mean of the spatial location uncertainty index corresponding to each granularity, construct the initial contribution coefficient of each granularity with the reciprocal of the statistical mean, and normalize the initial contribution coefficients of the three granularities.

[0140] It should be noted that constructing the initial contribution coefficient using the reciprocal of the mean of the uncertainty index enables the granularity with lower uncertainty and more reliable results to obtain a larger contribution ratio in the fusion process, thereby achieving adaptive weighted fusion based on reliability.

[0141] S6.2: When the dispersion of the normalized initial contribution coefficients of the three granularities is not higher than the upper limit of dispersion, and the average uncertainty index of the three granularities is not higher than the preset upper limit of uncertainty, the scene semantic level water body change map is used as the output of the water body dynamic monitoring result.

[0142] It should be noted that when the dispersion of the initial contribution coefficients of the three granularities is not higher than the upper limit of dispersion, and the average uncertainty index of the three granularities is not higher than the preset upper limit of uncertainty, the reliability of the results of each granularity is similar and the overall credibility is high. The scene semantic level water change map with the most complete semantic expression is output first, which can provide the richest semantic information while ensuring the reliability of detection.

[0143] The upper limit of the degree of dispersion can be set empirically based on the credibility distribution characteristics of the results at each granularity in the actual monitoring scenario, and the upper limit of the preset uncertainty can be set empirically based on the data quality of the target detection scenario.

[0144] S6.3: When the dispersion of the normalized initial contribution coefficients of the three particle sizes is higher than the upper limit of dispersion, or when there is a scale in the average uncertainty index of the three particle sizes that is higher than the preset upper limit of uncertainty, the result of weighted fusion of the detection results of the three particle sizes according to the normalized initial contribution coefficients shall be used as the output of the dynamic monitoring result of the water body.

[0145] Specifically, the change type markers of each spatial location in the target entity-level water change map and the scene semantic-level water change map are mapped to corresponding change probability values. The map is then rasterized into a probability map with the same spatial resolution as the primitive-level water change map. The predicted probability values ​​of each spatial location at the corresponding positions in the three granularities are multiplied by the normalized initial contribution coefficient of the corresponding granularity and then summed to obtain the probability value of each spatial location. Based on the probability values, dynamic water monitoring results are generated.

[0146] When the cloud coverage rate in the monitoring area is too high, causing the overall uncertainty index of the results at all particle sizes to exceed the preset uncertainty limit, a data quality label is added to the output water dynamic monitoring results to indicate that the reliability of the results is limited.

[0147] The technical effect of the above-mentioned adaptive fusion mechanism is as follows: the fusion weight is adaptively determined based on the prediction uncertainty of the detection results of each particle size, so that the particle size with higher reliability can make a greater contribution in the fusion process; when the overall credibility of the results of each particle size is high and the reliability is similar, the semantic level result with the most complete semantic expression is output first; when there are differences in the reliability of the results of each particle size, the complementary information of the three particles is comprehensively utilized through weighted fusion, so that the dynamic monitoring results of water bodies take into account both spatial accuracy and semantic integrity.

[0148] Accordingly, refer to Figure 4 This invention also provides an image processing system for dynamic water body monitoring, which may include:

[0149] The data preprocessing module is used to acquire the raw observation data stream, and after radiometric normalization anchoring and geometric registration, a reference map sequence is obtained.

[0150] The encoding module is used to input the reference map sequence into the time-series encoding network, use the hydrological periodic law as a soft time-series constraint anchor point, encode and map the spatiotemporal dynamics of the reference map sequence, and obtain the primary water body disturbance identification result carrying the change type label;

[0151] The scene adaptation module is used to perform adversarial scene drift suppression on the initial water disturbance identification results, extract scene-independent essential water response features, identify highly ambiguous regions, and obtain scene-adapted water disturbance identification results.

[0152] The uncertainty reasoning module is used to perform multiple parallel reasonings on highly ambiguous regions through a random activation masking mechanism. Based on the predicted distribution and change type label of the reasoning results, differentiated soft suppression is applied to each spatial location. The soft suppression results are then merged with the detection results of the remaining regions to obtain refined water disturbance identification results.

[0153] The processing module is used to parse the refined water disturbance identification results at the primitive level, target entity level, and scene semantic level, and calculate the global uncertainty index.

[0154] The output module is used to adaptively determine the contribution coefficient based on the global uncertainty index, perform weighted fusion, and generate dynamic water monitoring results.

[0155] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the image processing method for dynamic water monitoring as described in any of the above embodiments.

[0156] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the image processing method for dynamic water monitoring as described in any of the above embodiments.

[0157] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0158] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image processing method for dynamic monitoring of water bodies, characterized in that, include: The raw observation data stream is acquired, and a baseline map sequence is obtained after radiometric normalization anchoring and geometric registration. The reference map sequence is input into a time-series coding network, and the hydrological periodicity is used as a soft time-series constraint anchor point to encode and map the spatiotemporal dynamics of the reference map sequence, thereby obtaining the primary water body disturbance identification result carrying the change type label. Adversarial scene drift suppression is applied to the primary water disturbance identification results to extract scene-independent essential water body response features, identify highly ambiguous regions, and obtain scene-adaptive water disturbance identification results. The highly ambiguous region is subjected to multiple parallel inferences through a random activation masking mechanism. Differentiated soft suppression is applied to each spatial location based on the predicted distribution of the inference results and the change type label. The soft suppression results are then combined with the detection results of the remaining regions to obtain the refined water disturbance identification results. The refined water disturbance identification results are analyzed at the primitive level, target entity level, and scene semantic level, respectively, and the global uncertainty index is calculated. Based on the aforementioned global uncertainty index, the contribution coefficient is adaptively determined, and weighted fusion is performed to generate dynamic water body monitoring results. Using hydrological periodicity as a soft temporal constraint anchor point, the spatiotemporal dynamics of the baseline map sequence are encoded and mapped, including: Spatial features are extracted from each observed gait image in the baseline image sequence and fused to obtain the spatial representation information flow corresponding to each observed gait. The spatial representation information streams of adjacent observation gaits are subtracted point by point to obtain the difference perturbation field. The difference perturbation field is stacked and coupled with the original spatial representation information streams of the two adjacent frames in the feature channel dimension to obtain the dual observation gait interaction feature body of difference information and original spatial information. Hydrological periodicity data is encoded into temporal position embedding vectors and injected into the temporal position encoding layer of the temporal encoding network as soft constraint anchors. The dual-observation gait interaction feature body is temporally encoded to obtain temporal context features constrained by hydrological periodic semantics. In the feature space, the change displacement vector is calculated for the temporal context features of adjacent observed gaits. The change displacement vector is then jointly judged for change amplitude and change direction to generate a change type label, thus obtaining the primary water disturbance identification result carrying the change type label. The refined water disturbance identification results include: Multiple forward inferences are performed on highly ambiguous regions using a random neuron masking mechanism to obtain multiple sets of predicted probability distributions. The response distribution dispersion of each spatial location is calculated based on the multiple sets of predicted probability distributions, and the initial pseudo-change suppression weight of that location is generated based on the response distribution dispersion. The initial pseudo-change suppression weights are adjusted according to the change type label to obtain the secondary pseudo-change suppression weights; Based on the secondary pseudo-change suppression weight, the response results of each spatial location in the highly ambiguous region are subjected to soft suppression processing. The soft suppression results are then combined with the scene adaptation water disturbance identification results of the high-confidence region to obtain the refined water disturbance identification results.

2. The image processing method for dynamic water body monitoring according to claim 1, characterized in that, The reference map sequence obtained after radiometric normalization anchoring and geometric registration includes: Perform initial spatial matching on the raw observation data stream to obtain coarsely aligned observation frames; Spectral stability index is calculated for each spatial location in the coarsely aligned observation frame to determine the spatial location of candidate pseudo-invariant features; For the candidate pseudo-invariant feature spatial locations, calculate the temporal and frequency domain co-co ... Based on the spatial location of the pseudo-invariant features, a radiometric normalization regression model is established between the image to be corrected and the reference image. The coarsely aligned observation frame is then radiometrically normalized to obtain a radiometrically consistent image. Feature points are extracted from the radiometrically consistent image, and fine geometric registration is performed based on the feature points to obtain a reference image sequence.

3. The image processing method for dynamic water body monitoring according to claim 1, characterized in that, The results of scene-adaptive water disturbance recognition include: Remote sensing images of labeled water body change samples are used as reference scene data, and unlabeled real-time observation data of the area to be monitored are used as target detection scene data. Based on the benchmark reference scene data and target detection scene data, and constrained by the cross-domain semantic feature center alignment of water bodies, non-water bodies and water-land transition zones, the feature extraction network is trained in adversarial scene drift to obtain a scene adaptation feature network. The spatial representation information flow of the baseline map sequence and the primary water body disturbance identification results are jointly input into the scene adaptation feature network to obtain the essential water body response feature map; The feature membership reliability between each spatial location in the essential water body response feature map and each category of semantic feature center in the benchmark reference scene is calculated to mark highly ambiguous regions and obtain the scene-adaptive water body disturbance recognition result.

4. The image processing method for dynamic water body monitoring according to claim 1, characterized in that, The refined water disturbance identification results are parsed at the primitive level, target entity level, and scene semantic level, respectively, including: The water body disturbance identification results are refined to extract basic-level water body change information to obtain a basic-level water body change map. Eight-neighbor spatial connectivity analysis is performed on the elementary-level water body change map. Based on the spatial geometric attributes of each connectivity region, the change type of each connectivity region is spatially refined to obtain the target entity-level water body change map. Based on the target entity-level water body change map, the change type is semantically refined by combining geospatial prior information to obtain a scene semantic-level water body change map. Based on the region type of each spatial location, corresponding uncertainty indicators are determined, resulting in global uncertainty indicators for three different granularity detection results.

5. The image processing method for dynamic water body monitoring according to claim 1, characterized in that, The generated water body dynamic monitoring results include: Based on the global uncertainty index, the statistical mean of the spatial location uncertainty index corresponding to each granularity is calculated, and the initial contribution coefficient of each granularity is constructed by the reciprocal of the statistical mean. The initial contribution coefficients of the three granularities are then normalized. When the dispersion of the three granularity normalized initial contribution coefficients is not higher than the upper limit of dispersion, and the average uncertainty index of the three granularities is not higher than the preset upper limit of uncertainty, the scene semantic level water body change map is used as the output of the water body dynamic monitoring result. When the dispersion of the normalized initial contribution coefficients of the three particle sizes is higher than the upper limit of dispersion, or when there is a scale in the average uncertainty index of the three particle sizes that is higher than the preset upper limit of uncertainty, the result of weighted fusion of the detection results of the three particle sizes according to the normalized initial contribution coefficients is used as the output of the dynamic monitoring result of the water body.

6. The image processing method for dynamic water body monitoring according to claim 5, characterized in that, The output of water body dynamic monitoring results includes: The change type markers of each spatial location in the target entity-level water change map and the scene semantic-level water change map are mapped to the corresponding change probability values. The map is then rasterized into a probability map with the same spatial resolution as the primitive-level water change map. The predicted probability values ​​of each spatial location at the corresponding positions in the three granularities are multiplied by the normalized initial contribution coefficient of the corresponding granularity and then summed to obtain the probability value of each spatial location. Based on the probability values, the dynamic monitoring results of the water body are generated.

7. The image processing method for dynamic water body monitoring according to claim 2, characterized in that, Determining the spatial location of candidate pseudo-invariant features and the spatial location of pseudo-invariant features includes: Calculate the temporal variation coefficient of spectral reflectance at each spatial location and in each band of the coarsely aligned observation frame, obtain the spectral stability index based on the temporal variation coefficient, and compare and determine the spatial location of candidate pseudo-invariant features. For each candidate pseudo-invariant feature spatial location, a time-series spectral sequence is extracted, and the cooperation coefficients in the time and frequency domains are calculated based on the time-series spectral sequence to determine the pseudo-invariant feature spatial location.

8. An image processing system for dynamic water body monitoring, employing the image processing method for dynamic water body monitoring as described in any one of claims 1 to 7, characterized in that, include: The data preprocessing module is used to acquire the raw observation data stream, and after radiometric normalization anchoring and geometric registration, a reference map sequence is obtained. The encoding module is used to input the reference map sequence into the time-series encoding network, use the hydrological periodic law as a soft time-series constraint anchor point, encode and map the spatiotemporal dynamics of the reference map sequence, and obtain the primary water body disturbance identification result carrying the change type label; The scene adaptation module is used to perform adversarial scene drift suppression on the initial water disturbance identification results, extract scene-independent essential water response features, identify highly ambiguous regions, and obtain scene-adapted water disturbance identification results. The uncertainty reasoning module is used to perform multiple parallel reasonings on highly ambiguous regions through a random activation masking mechanism. Based on the predicted distribution and change type label of the reasoning results, differentiated soft suppression is applied to each spatial location. The soft suppression results are then merged with the detection results of the remaining regions to obtain refined water disturbance identification results. The processing module is used to parse the refined water disturbance identification results at the primitive level, target entity level, and scene semantic level, and calculate the global uncertainty index. The output module is used to adaptively determine the contribution coefficient based on the global uncertainty index, perform weighted fusion, and generate dynamic water monitoring results.

Images