A method and system for monitoring surface water storage based on multi-source synergy of air, space, and water.

Abstract

This invention discloses a method and system for monitoring surface water storage based on multi-source collaboration across air, space, and water, relating to the field of water resources survey technology. The method includes: acquiring multi-source collaborative data of the target water body to generate a multi-source dataset with a unified spatiotemporal benchmark; automatically extracting and detecting changes in the surface water area using the water body index method and deep learning to obtain the water body boundary; extracting onshore and underwater topographic data using the water body boundary as a spatial constraint, inputting it into a deep learning fusion model based on an encoder-decoder structure to generate an integrated land-water digital elevation model; and monitoring the storage capacity based on the water body type using corresponding deep learning models, calculating the water storage capacity and dynamic changes for different water body types. This invention achieves efficient fusion and intelligent processing of multi-source data, significantly improving the accuracy and efficiency of surface water storage monitoring, and is suitable for refined dynamic monitoring of various water bodies such as rivers, reservoirs, and ponds.

Description

Technical Field

[0001] This invention relates to the field of water resources survey technology, and more specifically, to a method and system for monitoring surface water storage based on multi-source synergy of air, space, and water. Background Technology

[0002] Accurate surveys and dynamic monitoring of surface water storage are crucial foundations for water resource development and utilization, flood control and drought relief scheduling, and ecological environment assessment. Traditional manual field measurement methods suffer from low efficiency, high cost, and limited coverage, especially in complex terrain and deep water areas, failing to meet the needs of modern refined water resource management.

[0003] With the development of technologies such as remote sensing, UAVs, navigation and positioning and underwater exploration, a variety of technical means have been gradually applied to surface water surveys, such as satellite remote sensing for large-scale water body identification, UAV aerial surveys to obtain high-resolution images, GNSS to provide high-precision positioning, and unmanned surface vessel sounding to obtain underwater topography. However, at present, these technologies are mostly in the stage of independent or simple combination application, such as the binary fusion mode of "satellite + ground measurement" and "UAV + remote sensing", which have not yet formed a systematic and integrated collaborative monitoring system, and have the following obvious limitations: (1) Existing schemes are mostly limited to the simple superposition of two or three data sources, and have failed to achieve the organic coordination and deep integration of four sources of data: "space-based (satellite), air-based (UAV), ground-based (GNSS / total station), and water-based (unmanned surface vessel sounding)", resulting in obvious shortcomings in the spatial continuity, topographic integrity and calculation accuracy of the survey results. (2) Rivers, reservoirs, ponds and other water bodies have significant differences in morphological characteristics and spatial distribution. Existing technologies often use a uniform method, which cannot adapt to the survey needs of different water bodies, affecting the reliability and economy of the results. (3) Key steps such as water body boundary extraction, terrain modeling, and water volume calculation still rely heavily on manual intervention, which is inefficient and prone to errors, making it difficult to support large-scale, high-frequency dynamic monitoring.

[0004] Therefore, there is an urgent need to develop a method and system for monitoring surface water storage that can integrate multi-source data, achieve integrated coordination of air, land, and water, and possess intelligent processing capabilities, so as to improve the accuracy, efficiency, and applicability of the survey and provide reliable technical support for the scientific management of water resources. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes a method and system for monitoring surface water storage based on multi-source collaboration between air, space, and water. This method achieves efficient fusion and intelligent processing of multi-source data, significantly improving the accuracy and efficiency of surface water storage monitoring. It is applicable to the refined dynamic monitoring of various water bodies such as rivers, reservoirs, and ponds.

[0006] The first aspect of this invention provides a method for monitoring surface water storage based on multi-source coordination of air, space, and water, comprising the following steps:

[0007] S1: Acquire multi-source collaborative acquisition data of the target water body, preprocess the multi-source collaborative acquisition data and unify it with the spatiotemporal reference to generate a multi-source dataset with a unified spatiotemporal reference.

[0008] S2: Based on the aforementioned multi-source dataset, the water index method and deep learning are used to automatically extract and detect changes in the surface water area, thereby obtaining the water area boundary;

[0009] S3: Using the water boundary as a spatial constraint, extract the onshore and underwater topographic data from the multi-source dataset and input them into a deep learning fusion model based on an encoder-decoder structure to perform water-land topographic fusion and generate an integrated water-land digital elevation model.

[0010] S4: Based on the water body type, the storage capacity is monitored by using a deep learning model corresponding to the water body type based on the integrated land and water digital elevation model, and the water storage capacity and dynamic changes of different water body types are calculated.

[0011] In this scheme, multi-source collaborative data acquisition of the target water body is used, including

[0012] According to the preset preliminary survey plan, multi-source collaborative acquisition data of the target water body is obtained. The multi-source collaborative acquisition data includes macroscopic image data obtained by satellite remote sensing, high-resolution terrain and image data obtained by UAV aerial survey, control point and water surface elevation data obtained by ground measurement, and underwater terrain data obtained by underwater mapping.

[0013] Extract the metadata and feature descriptors corresponding to the multi-source collaborative acquisition data. The feature descriptors include image resolution, point cloud density, control point distribution uniformity, and water depth range.

[0014] Based on big data retrieval methods, surface water storage survey cases of different water body types are obtained to construct a historical experience database. Each survey case is used as a node, and water body type, geometric morphology features, data source features, environmental and working condition features, final contribution weight vector of each data source, and task objectives are used as node attributes. Feature similarity edges are constructed by calculating the feature similarity between nodes, and nodes are connected through the feature similarity edges to generate a graph corresponding to the historical experience database.

[0015] Based on the feature descriptor, the initial node location is performed in the graph using similarity. A preset number of nodes are selected as the starting point set for the walk. Starting from each walk node, a random walk guided by feature similarity and task objective is performed along the feature edges and task edges in the graph.

[0016] During the random walk, the weight vectors of each data source in the path nodes are recorded to generate a path weight sequence. The sequence is aggregated using an attention mechanism to generate a path weight vector. The weight vectors of all paths are integrated to generate an adaptive weight allocation scheme suitable for the target water body and the current task.

[0017] The adaptive weight allocation scheme is fed back to the data acquisition system, and subsequent data acquisition is dynamically adjusted based on the weights of each data source.

[0018] In this scheme, based on the aforementioned multi-source dataset, the water index method and deep learning are used to automatically extract and detect changes in the extent of surface water areas, resulting in water area boundaries, including:

[0019] Load the generated multi-source dataset with a unified spatiotemporal reference, and extract multi-temporal optical satellite images, multi-temporal synthetic aperture radar images, and high-resolution orthophotos acquired by UAVs.

[0020] For parallel computation of water index, vegetation index and building index for optical images, for extraction of backscattering intensity, coherence, texture features and polarization decomposition features for synthetic aperture radar images, the original spectral bands, index features, synthetic aperture radar features and UAV orthophotos are stacked at the pixel level to generate feature cubes.

[0021] The fusion feature cubes of all time phases are imported into a deep learning network containing a spectral index branch, a synthetic aperture radar feature branch, and a spatial context branch to generate a water body probability time series. A change detection method based on Bayesian update is then used to obtain the final determined water body probability map.

[0022] The water body probability map is binarized, and the binary grid is converted into a preliminary vector boundary. The preliminary vector boundary is spatially registered and compared with the coordinates of key boundary points obtained by high-resolution UAV orthophotos and ground-based measurements. Based on preset rules, the boundary is refined according to the comparison results, and the spatial vector boundary of the surface water area is output.

[0023] In this scheme, a deep learning network is constructed and trained, including a spectral index branch, a synthetic aperture radar feature branch, and a spatial context branch, comprising:

[0024] A preset adaptive threshold is used to mark water body sample points and non-water body sample points on the multi-temporal optical satellite images using water body index, constructing a training sample set, and supplementing it with high-resolution orthophotos acquired by UAVs by extracting water body boundary samples through edge detection.

[0025] A deep learning network is constructed that includes a spectral index branch, a synthetic aperture radar feature branch, and a spatial context branch. The network is trained using training samples. The spectral index branch is used to obtain the response patterns of different index combinations under different environments. The synthetic aperture radar feature branch is used to learn the performance of water bodies in radar images. The spatial context branch is used to learn the spatial structure features of water body boundaries.

[0026] Attention fusion is used to weight and fuse the output features of the three branches. The output of the spectral index branch is used as prior knowledge. The learning attention of the synthetic aperture radar feature branch and the spatial context branch is dynamically adjusted. The probability map of the output pixel as water body is tested. When the network performance meets the preset standard, the trained deep learning network is obtained.

[0027] In this scheme, the water boundary is used as a spatial constraint. Onshore and underwater topographic data are extracted from the multi-source dataset and input into a deep learning fusion model based on an encoder-decoder structure for water-land topography fusion, including:

[0028] Obtain the boundary of the land surface and water area, and extract the corresponding onshore and underwater topographic data from the multi-source dataset. Based on the boundary of the land surface and water area, establish a fusion buffer that extends to both sides of the land and water area according to a preset distance. Use the fusion buffer to spatially clip the onshore and underwater topographic data, and convert the clipped point cloud data into an initial digital elevation model.

[0029] A deep learning fusion model is established based on the encoder-decoder structure. Two symmetrical encoder branches are set up to process the initial digital elevation models of the onshore and underwater terrain, respectively. Multi-scale dilated convolution is used in each encoder to extract multi-scale features, and cross attention is used to associate the features of the corresponding scales of the two encoder branches.

[0030] The multi-scale features extracted from the two encoder branches are imported into the gated fusion unit to generate a dynamic fusion weight map for feature fusion. The fused multi-scale features are then imported into the decoder, where the original input data is initially classified into terrain to obtain a terrain semantic supervision signal.

[0031] The terrain semantic supervision signal is used to guide the reconstruction of the fused multi-scale features, and a complete integrated digital elevation model of land and water is output.

[0032] In this scheme, the training of a deep learning fusion model based on an encoder-decoder structure includes:

[0033] Collect complete measured data of land and water topography in the region, and perturb the region samples to construct paired samples of defect sample data and true value sample data;

[0034] The paired samples are input into the constructed deep learning fusion model to obtain the predicted complete integrated water and land digital elevation model. Based on the predicted complete integrated water and land digital elevation model, the real complete integrated water and land digital elevation model and the input water boundary, the elevation regression loss, terrain structure loss and boundary continuity loss are calculated respectively, and the total loss is obtained by summing them according to the weights.

[0035] The gradient of the total loss with respect to all weight parameters of the model is calculated by backpropagation, and the parameters are updated using an adaptive optimizer to minimize the total loss. After iterative training, the model is validated, and training ends when the performance meets the preset requirements.

[0036] In this scheme, for river water bodies, the integrated land-water digital elevation model, water area vector boundaries, and time-series hydrological data are input into a river network dynamic model based on a graph neural network to calculate the segmented storage capacity and total storage capacity changes in different river segments, including:

[0037] For river water bodies, the river centerline extracted based on the water area boundary divides the river into several continuous river segments. Each river segment is defined as a graph structure node. For each node, the average river width, average slope, cross-sectional area sequence, estimated riverbed roughness value, and channel storage capacity curve of the river segment are extracted from the integrated land and water digital elevation model to construct static features. The time series features of the river segment hydrological data are also extracted to construct dynamic features.

[0038] Directed edges are defined based on the upstream and downstream relationships of nodes. A river system graph structure is established based on the directed edges and nodes. The static features of each node and the dynamic features of one time step are merged and input into a multilayer perceptron to be encoded into a node state embedding vector.

[0039] The node state is embedded into a vector and input into a river network dynamic model based on a graph neural network. Graph convolution is used to pass messages based on the upstream node state and the dynamic weights of directed edges. Downstream nodes aggregate messages from upstream neighboring nodes and import them into a gated recurrent unit to update the node state.

[0040] The updated node states are imported into the fully connected network to obtain the water volume change value and outflow prediction value of the river segment at the current time step. The real-time total storage of the river segment is calculated based on the theoretical static volume of the river segment at the current water level. Finally, the storage change process line of each river segment and the total storage change process line of the entire river are output in the time series.

[0041] In this scheme, for the reservoir water body, the integrated land-water digital elevation model, water area vector boundary, and multi-period water level observation data are input into the reservoir capacity model fused with a spatiotemporal attention mechanism to obtain the reservoir capacity value and predict future storage volume, including:

[0042] In the integrated land and water digital elevation model of the reservoir, horizontal slices are made based on historical water level observation data and water area boundaries. The inundation range of the reservoir under the water level is calculated, an instantaneous water surface polygon is generated, the underwater topography of the instantaneous water surface polygon is extracted, a three-dimensional topographic instance of the reservoir at a specific water level is generated, and finally the reservoir topographic state sequence corresponding to the water level is obtained.

[0043] For each reservoir 3D terrain instance, calculate the corresponding geometric features and terrain features to construct a morphological state vector, and encapsulate the water level observation value, morphological state vector and the reservoir capacity value observed at the same time point into a spatiotemporal sample.

[0044] In the reservoir capacity model, the spatiotemporal samples are spatiotemporally fused and encoded to obtain the fusion state vector at each time point. Spatial morphology attention is introduced to learn the contribution weight of different topographic units of the reservoir to the total capacity under different water levels. Temporal memory attention is introduced to calculate the temporal dependence and memory state of the capacity change.

[0045] After processing by the cascaded attention mechanism, the enhanced spatiotemporal state vector at each time point is obtained. The reservoir capacity value at each time point is output using a fully connected layer, and the water level forecast value at future time is obtained. The morphological state vector corresponding to the water level forecast value is imported into the reservoir capacity model to predict the future storage volume at future time.

[0046] In this scheme, for pond water bodies, under the constraints of the integrated land-water digital elevation model and the water area vector boundary, a water depth inversion model based on remote sensing imagery and a pixel volume accumulation algorithm are used to calculate the storage capacity of each pond in the area, including:

[0047] Based on the water area vector boundary, bottom topographic data corresponding to each pit is extracted from the integrated land and water digital elevation model, and spatially registered with the remote sensing image. According to the bottom topographic data and the estimated water surface elevation at the imaging time, the pseudo water depth true value is calculated for each pixel in the water area of ​​each pit to form a training sample set.

[0048] A regional adaptive water depth inversion deep learning model is constructed, with pixel spectral features and bottom elevation as inputs, and soft constraints based on the physical equation of water optical transmission are introduced. The model is trained using the training sample set to learn the mapping relationship from spectrum and topographic features to water depth.

[0049] The trained water depth inversion model is applied to the target pond water body, and the water depth value of each pixel in each pond is obtained by batch inversion. The inverted water depth of all pixels within the water boundary is multiplied by the pixel area to obtain the micro-volume of each pixel. The micro-volumes of all pixels are accumulated to obtain the total storage capacity of the pond.

[0050] The second aspect of the present invention provides a surface water storage monitoring system based on multi-source collaboration of air, space, land and water. The system includes: a multi-source collaborative acquisition and adaptive planning module, a multi-source heterogeneous data preprocessing and fusion management module, an intelligent water area boundary extraction and change detection module, an integrated land and water terrain intelligent fusion modeling module, a differentiated water storage intelligent calculation and prediction module, and a visualization analysis and decision-making module.

[0051] The multi-source collaborative acquisition and adaptive planning module receives user input of water body type, target area and accuracy requirements to perform adaptive task planning and acquire multi-source collaborative acquisition data of the target water body.

[0052] The multi-source heterogeneous data preprocessing and fusion management module preprocesses and unifies the spatiotemporal reference of the multi-source collaboratively collected data to generate a multi-source dataset with a unified spatiotemporal reference.

[0053] The intelligent water area boundary extraction and change detection module, based on the multi-source dataset, uses the water index method and deep learning to automatically extract and detect changes in the surface water area, obtain the water area boundary, and identify the change area and intensity.

[0054] The integrated land and water terrain intelligent fusion modeling module uses the water boundary as a spatial constraint, extracts onshore and underwater terrain data from the multi-source dataset, and inputs them into a deep learning fusion model based on an encoder-decoder structure to perform land and water terrain fusion, generate an integrated land and water digital elevation model, and generate a data confidence map to quantitatively evaluate the reliability of terrain data in different regions.

[0055] The differentiated water storage intelligent calculation and prediction module uses the river dynamic model sub-engine, reservoir capacity model sub-engine and pit and pond survey model sub-engine corresponding to the water body type based on the integrated land and water digital elevation model to monitor the storage and calculate the water storage and dynamic changes of different water body types.

[0056] The visualization analysis and decision-making module dynamically displays water body boundaries, terrain models, water storage volume and change processes, and provides storage data, early warning information and analysis reports through API.

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0058] This invention constructs a four-in-one collaborative data acquisition system encompassing "sky, air, land, and water," and combines it with an adaptive weight allocation mechanism to overcome the limitations of traditional survey methods, such as single data sources and incomplete coverage. Based on water body type and survey objectives, it intelligently schedules and integrates multi-source heterogeneous data, achieving an organic combination of large-scale macro-monitoring and local high-precision detection. This ensures the breadth of the survey and the detailed accuracy of key areas, significantly improving the comprehensiveness and relevance of data acquisition.

[0059] By employing automated water boundary extraction based on a multi-branch collaborative training network, terrain fusion combining physical constraints and deep learning, and specialized models such as graph neural networks and spatiotemporal attention mechanisms tailored to different water body types, this solution achieves intelligent processing across the entire data-to-information chain. This enhances the automation and accuracy of water body extent identification, terrain modeling, and water volume calculation, reduces reliance on human experience, and enables the model to learn complex hydrological and topographical relationships and spatiotemporal dynamic patterns, revealing the dynamic changes and underlying mechanisms of water bodies that are difficult to capture using traditional methods. Furthermore, this solution can meet the specific management needs of different types of water bodies, providing more refined and reliable data support than ever before for practical operations such as water resource allocation, flood forecasting, drought assessment, and ecological water replenishment. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0061] Figure 1 A flowchart of a method for monitoring surface water storage based on multi-source synergy of air, space, land, and water is shown.

[0062] Figure 2 A flowchart illustrating the process of obtaining the water boundary of the surface water area is shown;

[0063] Figure 3 A flowchart illustrating the construction of an integrated land and water digital elevation model is shown.

[0064] Figure 4 A block diagram of a surface water storage monitoring system based on multi-source synergy of air, space, land, and water is shown. Detailed Implementation

[0065] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0066] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0067] like Figure 1As shown, this embodiment provides a method for monitoring surface water storage based on multi-source synergy of air, space, land, and water, including:

[0068] S1: Acquire multi-source collaborative acquisition data of the target water body, preprocess the multi-source collaborative acquisition data and unify it with the spatiotemporal reference to generate a multi-source dataset with a unified spatiotemporal reference.

[0069] S2: Based on the aforementioned multi-source dataset, the water index method and deep learning are used to automatically extract and detect changes in the surface water area, thereby obtaining the water area boundary;

[0070] S3: Using the water boundary as a spatial constraint, extract the onshore and underwater topographic data from the multi-source dataset and input them into a deep learning fusion model based on an encoder-decoder structure to perform water-land topographic fusion and generate an integrated water-land digital elevation model.

[0071] S4: Based on the water body type, the storage capacity is monitored by using a deep learning model corresponding to the water body type based on the integrated land and water digital elevation model, and the water storage capacity and dynamic changes of different water body types are calculated.

[0072] It should be noted that, according to the pre-set preliminary survey plan, multi-source collaborative acquisition data of the target water body is obtained. This multi-source collaborative acquisition data includes macroscopic image data acquired through satellite remote sensing, high-resolution topographic and image data acquired through UAV aerial surveying, control point and water surface elevation data acquired through ground surveying, and underwater topographic data acquired through underwater mapping. Metadata and feature descriptors corresponding to the multi-source collaborative acquisition data are extracted. The feature descriptors include image resolution, point cloud density (average number of points per square meter of UAV lidar point cloud and underwater bathymetry point cloud), control point distribution uniformity (calculated by the Gini coefficient of the control point locations), and water depth range (obtained statistically from the preliminary underwater bathymetry data). The maximum, minimum, and average water depth of the water body are determined. Based on big data retrieval methods and literature mining technology, surface water storage survey cases of different water body types (rivers, reservoirs, ponds) are obtained to construct a historical experience database. Each survey case is used as a node, and the water body type, geometric morphological features (length-to-width ratio, surface area, shoreline development coefficient, etc.), data source features (the aforementioned feature descriptors), environmental and working condition features (meteorological conditions, water turbidity, operating cost range, etc.), the final contribution weight vector of each data source, and the task objective (survey accuracy level) are used as node attributes. The feature similarity between nodes is calculated to construct feature similarity edges. The nodes are connected through the feature similarity edges to generate the graph corresponding to the historical experience database.

[0073] Initial node localization is performed in the graph using similarity based on the feature descriptors. A preset number of nodes are selected as the starting point set for the walk. Starting from each walking node, random walks are performed along feature edges and task edges in the graph, guided by feature similarity and task objectives. Feature similarity-guided walks tend to jump to neighboring nodes connected by feature similarity edges, while task objective-guided walks tend to jump to neighboring nodes whose task objective attributes better match the current investigation task objective. During the random walk, the weight vectors of each data source in the path nodes are recorded to generate a path weight sequence. An attention mechanism is used to aggregate the sequence, distributing attention scores to each historical weight vector in the sequence. Through weighted summation, the entire sequence is aggregated into path weight vectors representing each path. The weight vectors of all paths are integrated, such as by taking the mean or weighted median, to generate an adaptive weight allocation scheme suitable for the target water body and the current task. The adaptive weight allocation scheme is fed back to data acquisition, and subsequent data acquisition is dynamically adjusted according to the weights of each data source.

[0074] It should be noted that, as Figure 2 As shown, a multi-source dataset with a unified spatiotemporal reference is loaded, and multi-temporal optical satellite images, multi-temporal synthetic aperture radar (SAR) images, and high-resolution orthorectified images acquired by UAVs are extracted. For optical images, water body indices, vegetation indices, and building indices are calculated in parallel. For SAR images, backscattering intensity, coherence, texture features, and polarization decomposition features are extracted. The original spectral bands, index features, SAR features, and UAV orthorectified images are stacked at the pixel level to generate feature cubes. The fused feature cubes of all time phases are imported into a deep learning network containing a spectral index branch, a SAR feature branch, and a spatial context branch to generate a water body probability time series. A change detection method based on Bayesian updates is used to obtain the final determined water body probability map. The spectral index branch takes the calculated water body index as input to learn the response law of different index combinations under different environments. The SAR feature branch takes the backscattering, coherence, and other features extracted by SAR as input to learn the performance of water bodies in radar images. The spatial context branch takes high-resolution images as input and learns the spatial structure features near the water body boundary through a convolutional neural network. The water body probability map is binarized, and the binary grid is converted into a preliminary vector boundary. This preliminary vector boundary is then spatially registered and compared with the coordinates of key boundary points obtained from high-resolution UAV orthophotos and ground-based measurements. Based on preset rules, the boundary is refined according to the comparison results, and the spatial vector boundary of the surface water area is output. Preferred preset rules include that where the boundary is clearly discernible in the UAV image, it should converge towards the water-land boundary as interpreted by the image; and where there are key control points, the boundary must pass through or be adjacent to those points.

[0075] A deep learning network comprising a spectral index branch, a synthetic aperture radar (SAR) feature branch, and a spatial context branch was constructed and trained. An adaptive threshold was preset, and water body sample points and non-water body sample points were marked on multi-temporal optical satellite imagery using water body indices to construct a training sample set. High-resolution orthophotos acquired by UAVs were used to supplement the training sample set by extracting water body boundary samples through edge detection. The deep learning network, comprising the spectral index branch, SAR feature branch, and spatial context branch, was trained using the training samples. The spectral index branch was used to obtain the response patterns of different index combinations under different environments. The SAR feature branch was used to learn the representation of water bodies in radar imagery, and the spatial context branch was used to learn the spatial structural features of water body boundaries. During training, the three branches shared some low-level features. Attention fusion was used to weightedly fuse the output features of the three branches. The output of the spectral index branch was used as prior knowledge to dynamically adjust the learning attention of the SAR feature branch and the spatial context branch. A probability map of pixels representing water bodies was output for testing. When the network performance met the preset standards, the trained deep learning network was obtained.

[0076] It should be noted that, as Figure 3 As shown, the boundary of the surface water area is obtained, and the corresponding onshore and underwater topographic data are extracted from the multi-source dataset. Based on the boundary of the surface water area, a fusion buffer extending to both sides of the land and water is established according to a preset distance of 50-100 meters. The fusion buffer is used to spatially clip the onshore and underwater topographic data, and the clipped point cloud data is converted into an initial digital elevation model. A deep learning fusion model is established based on an encoder-decoder structure. The deep learning fusion model has the ability to understand terrain semantics. Two symmetrical encoder branches are set to process the initial digital elevation models of the onshore and underwater topographic data respectively. Multi-scale dilated convolution is used in each encoder to extract multi-scale features. The shallow network obtains local terrain texture, and the deep network understands the terrain structure by increasing the receptive field. Cross attention is used to associate the features of the corresponding scales of the two encoder branches, so that the onshore topographic features match the possible continuous structure in the underwater topographic features. Multi-scale features extracted from the two encoder branches are imported into a gated fusion unit to generate a dynamic fusion weight map for feature fusion. For areas with clear land-water boundaries, the original data is used; for areas with ambiguous boundaries, contextual information learned from the opposite branch is relied upon. The fused multi-scale features are then imported into a decoder, where the original input data undergoes preliminary terrain classification, roughly dividing the point cloud into steep slopes, gentle slopes, and flat land, thus obtaining a terrain semantic supervision signal. This signal guides the reconstruction of the fused multi-scale features, ensuring the reconstruction process conforms to natural terrain patterns. A complete integrated land-water digital elevation model (DEM) is output, and a data confidence map is generated to indicate the reliability of the elevation values ​​at each grid point.

[0077] A deep learning fusion model based on an encoder-decoder structure is trained. Complete regional samples of measured land and water topography are collected. These regional samples are perturbed to construct paired samples of defective and ground truth samples. For example, the water area is cropped from the complete terrain to simulate missing underwater data, and underwater point clouds are sparsified or noise is added to simulate different depth sounding accuracy conditions. The paired samples are input into the constructed deep learning fusion model to obtain a predicted complete land and water integrated digital elevation model (DEM). Based on the predicted complete land and water integrated DEM, the actual complete land and water integrated DEM, and the input water boundary, elevation regression loss, terrain structure loss, and boundary continuity loss are calculated respectively, and the total loss is obtained by summing them according to their weights. The elevation regression loss ensures that the DEM generated by the model is as close as possible to the actual terrain data in terms of absolute elevation values. In areas where the actual DEM is known, the elevation difference between each effective grid point between the model-predicted DEM and the actual DEM is calculated. The terrain structure loss ensures that the generated terrain is consistent with the real terrain in both microscopic and macroscopic morphology. Gradients in the east-west and north-south directions are calculated for both the predicted and real DEMs, resulting in two gradient fields. The L2 norm loss between these two gradient fields is calculated, and the curvature of the predicted and real DEMs is further calculated and compared. The terrain structure loss is obtained by weighted sum of the gradient loss and curvature loss. The boundary continuity loss ensures a smooth transition of terrain on both sides of the land-water boundary. A boundary buffer mask is generated using the input water boundary vector. Within the mask region, the elevation gradient in the predicted DEM perpendicular to the boundary direction is calculated. This gradient is compared with a predefined gentle gradient distribution, and the steep gradient at the boundary is penalized by minimizing the difference. The elevation difference between corresponding point pairs on both sides of the boundary line is calculated and minimized.

[0078] The gradient of the total loss with respect to all weight parameters of the model is calculated by backpropagation, and the parameters are updated using an adaptive optimizer to minimize the total loss. During training, the input terrain tiles are randomly rotated, slightly scaled, and random noise is added to increase the diversity of the data. After iterative training, the model is validated, and training ends when the performance meets the preset requirements.

[0079] For river bodies, the integrated land-water digital elevation model, water area vector boundaries, and time-series hydrological data are input into a river network dynamic model based on a graph neural network to calculate the segmented storage capacity and total storage capacity changes of different river segments. For river bodies, based on the river centerline extracted from the water area boundaries, the river is divided into several continuous river segments according to landmarks such as natural river channel turning points, hydrological stations, and confluence points. Each river segment is defined as a graph structure node. For each node, the average river width, average slope, cross-sectional area sequence, estimated riverbed roughness, and channel storage capacity curve are extracted from the integrated land-water digital elevation model to construct static features. Temporal features of the river segment's hydrological data, including upstream inlet flow, local rainfall, evaporation, and downstream outlet water level, are also extracted to construct dynamic features. Directed edges are defined based on the upstream and downstream relationships of the nodes. Based on the river channel meandering and slope changes calculated from the DEM, the river channel morphology connecting two river segments is obtained, and real-time hydraulic conditions are obtained based on the flow velocity estimated from the upstream and downstream flow differences. Edge weights are obtained based on the river channel morphology and hydraulic conditions. A river system graph structure is established based on the directed edges and nodes. The static features of each node and the dynamic features of one time step are merged and input into a multilayer perceptron to be encoded into a node state embedding vector. The node state embedding vector is input into a river network dynamic model based on a graph neural network. Graph convolution is used to pass messages based on the upstream node state and the dynamic weights of the directed edges. The messages include the upstream node state, the dynamic weights of the edges, and the soft constraints of the mass conservation principle. Downstream nodes aggregate messages from upstream neighboring nodes and import them into a gated recurrent unit to update the node state. The updated node state is imported into a fully connected network to obtain the water volume change value and outflow prediction value of the river segment at the current time step. The real-time total storage capacity of the river segment is calculated based on the theoretical static volume of the river segment at the current water level. Finally, the storage capacity change process line of each river segment and the total storage capacity change process line of the entire river are output on the time series.

[0080] For the reservoir water body, the integrated land-water digital elevation model, water area vector boundary and multi-period water level observation data are input into the reservoir capacity model with spatiotemporal attention mechanism. The outflow predicted by the model is used as one of the inputs to the downstream node of the next time step to obtain the reservoir capacity value and predict the future storage volume. In the integrated land-water digital elevation model of the reservoir, horizontal slicing is performed based on historical water level observation data and water area boundaries to calculate the inundation range of the reservoir under the water level, generating instantaneous water surface polygons. The underwater topography of these instantaneous water surface polygons is extracted to generate 3D reservoir topography instances at specific water levels. This slicing operation is repeated for each historical water level to ultimately obtain a reservoir topography state sequence corresponding to the water level, containing the overall topography of the reservoir area under the water level. For each 3D reservoir topography instance, corresponding geometric and topographic features are calculated to construct a morphological state vector. The geometric features include the water surface area at the current water level, the length of the reservoir shoreline, the reservoir shoreline complexity index, and the inundation length of major tributaries. The topographic features include the average reservoir bottom elevation calculated based on the current inundation area DEM, the reservoir topographic relief, and the rate of change of the area enclosed by specific contour lines with elevation. The water level observation values, morphological state vectors, and concurrently observed reservoir capacity values ​​at each time point are encapsulated as spatiotemporal samples to construct a training dataset.

[0081] In the reservoir capacity model, spatiotemporal fusion encoding is performed on the spatiotemporal samples to obtain the fusion state vector at each time point. This spatiotemporal fusion encoding includes time position encoding and morphological feature encoding. Time position encoding adds a sine-cosine position code to the timestamp of each sample to perceive the order and period of the samples in the time series. Morphological feature encoding maps the morphological state vector to a high-dimensional morphological embedding vector. The time position code, water level value, and morphological embedding vector are concatenated to obtain the fusion state vector at each time point. The model divides the reservoir topography into regular grids in horizontal space. The morphological feature vector contains the area-elevation relationship information of the grids. Spatial morphological attention is introduced, dynamically calculating the attention weight of each topographic grid based on the current water level and global context. The model learns the contribution weight of different topographic units to the total reservoir capacity at different water levels. A Transformer encoder architecture is used to introduce temporal memory attention. Through a self-attention mechanism, the state of all time points in the historical sequence is obtained, and a gating mechanism is introduced to determine how far back and how strong the past state information is influencing the current reservoir capacity calculation. The temporal dependence and memory state of the changes are considered. After processing through a cascaded attention mechanism, an enhanced spatiotemporal state vector is obtained for each time point, which integrates the current form, spatial weights, and effective historical memory. The reservoir capacity value at each time point is output using a fully connected layer, and the water level forecast value for future time points is obtained. The morphological state vector corresponding to the water level forecast value is imported into the reservoir capacity model to predict the future storage volume. In addition, it is coupled with a hydrological model, which provides the future inflow process. The reservoir capacity model iteratively predicts the future water level and reservoir capacity changes based on the current spatiotemporal state and the future inflow.

[0082] For ponds, under the constraints of the integrated land-water digital elevation model and the water area vector boundary, a water depth inversion model based on remote sensing imagery and a pixel volume accumulation algorithm are used to calculate the storage capacity of each pond within the region. This includes: using the water area vector boundary as a precise spatial mask, extracting the bottom topographic data corresponding to each pond from the integrated land-water digital elevation model; extracting spectral information of the same pond range from multi-source remote sensing images of the same phase; and spatially registering the integrated land-water digital elevation model grid with the remote sensing image pixels. For each pond, the bottom elevation of each pixel location is calculated using the corresponding integrated land-water digital elevation data. The water surface elevation of the pond at the imaging time is obtained from the attributes of the water area vector boundary through local area horizontal plane fitting. Based on the bottom topographic data and the estimated water surface elevation at the imaging time, pseudo-depth ground truth values ​​are calculated for pixels within each pond, forming a training sample set.

[0083] A regional adaptive water depth inversion deep learning model is constructed, using pixel spectral features and bottom elevation as inputs. Soft constraints based on the physical equations of optical transmission in water are introduced to force the network to predict water depth that satisfies the fundamental physical laws of attenuation and scattering with the optical signal. For example, the network needs to learn that in clear water, the reflectance of a specific band is approximately linearly related to the logarithm of water depth; while in turbid water, the relationship is modulated by the concentration of suspended solids. The model is trained using the training sample set to measure the mean square error between the predicted water depth and the pseudo-true water depth, learning the mapping relationship from spectral and topographic features to water depth, and generating a regionally adaptive physically enhanced inversion generator. The trained water depth inversion model is applied to the target pond water body, batch inverting to obtain the water depth values ​​of each pixel within each pond. The inverted water depth of all pixels within the water boundary is multiplied by the pixel area to obtain the micro-volume of each pixel. The micro-volumes of all pixels are accumulated to obtain the total storage capacity of the pond.

[0084] like Figure 4 As shown, the second embodiment of the present invention provides a surface water storage monitoring system based on multi-source collaboration of air, land, and water. The system includes: a multi-source collaborative acquisition and adaptive planning module, a multi-source heterogeneous data preprocessing and fusion management module, an intelligent water area boundary extraction and change detection module, an integrated land and water terrain intelligent fusion modeling module, a differentiated water storage intelligent calculation and prediction module, and a visualization analysis and decision-making module.

[0085] The multi-source collaborative acquisition and adaptive planning module receives user input regarding water body type, target area, and accuracy requirements to perform adaptive task planning and acquire multi-source collaborative acquisition data of the target water body.

[0086] The multi-source heterogeneous data preprocessing and fusion management module preprocesses the multi-source collaboratively collected data and unifies it with the spatiotemporal reference to generate a multi-source dataset with a unified spatiotemporal reference.

[0087] The intelligent water area boundary extraction and change detection module, based on the multi-source dataset, uses the water body index method and deep learning to automatically extract and detect changes in the surface water area, obtain the water area boundary, and identify the change area and intensity.

[0088] The integrated land and water terrain intelligent fusion modeling module uses the water boundary as a spatial constraint, extracts onshore and underwater terrain data from the multi-source dataset, and inputs them into a deep learning fusion model based on an encoder-decoder structure to perform land and water terrain fusion, generate an integrated land and water digital elevation model, and generate a data confidence map to quantitatively evaluate the reliability of terrain data in different regions.

[0089] The differentiated water storage intelligent calculation and prediction module, based on the integrated land-water digital elevation model, uses river dynamic model sub-engines, reservoir capacity model sub-engines, and pond survey model sub-engines corresponding to the water body type to monitor storage and calculate the water storage and dynamic changes for different water body types. The river dynamic model sub-engine, constructed based on a graph neural network, takes into account the integrated land-water DEM, river water body boundaries, and time-series hydrological data to simulate the dynamic transmission and distribution of water in the river segment, calculating the segmented storage of the river segment and the overall water storage of the entire river. The total storage volume changes over time; the reservoir capacity model sub-engine is constructed based on a spatiotemporal morphological memory attention network, inputting integrated land-water DEM, reservoir boundaries, and multi-period water level data, learning the complex relationship between reservoir morphology and capacity, obtaining reservoir capacity, and supporting the prediction of future water storage; the pit and pond survey model sub-engine is constructed based on a physics-guided deep learning water depth inversion model, under the constraints of integrated land-water DEM and boundaries, quickly inverting the water depth distribution of all pits and ponds in the area, and batch calculating the storage volume of each pit and pond through pixel volume accumulation addition to achieve regional survey.

[0090] The visualization analysis and decision-making module uses various methods such as two-dimensional maps, three-dimensional scenes, and time-series charts to dynamically display water body boundaries, terrain models, water storage volume and change processes. It also provides storage data, early warning information and analysis reports to water resource management, flood control and drought relief, ecological assessment and other businesses through API.

[0091] The third embodiment of the present invention provides a computer-readable storage medium, which includes a method program for monitoring surface water storage based on multi-source coordination of air, space, and water. When the method program is executed by a processor, it implements the steps of the method for monitoring surface water storage based on multi-source coordination of air, space, and water.

[0092] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0093] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0094] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring surface water storage based on multi-source synergy of air, space, and water, characterized in that, Includes the following steps: S1: Acquire multi-source collaborative acquisition data of the target water body, preprocess the multi-source collaborative acquisition data and unify it with the spatiotemporal reference to generate a multi-source dataset with a unified spatiotemporal reference. S2: Based on the aforementioned multi-source dataset, the water index method and deep learning are used to automatically extract and detect changes in the surface water area, thereby obtaining the water area boundary; S3: Using the water boundary as a spatial constraint, extract the onshore and underwater topographic data from the multi-source dataset and input them into a deep learning fusion model based on an encoder-decoder structure to perform water-land topographic fusion and generate an integrated water-land digital elevation model. S4: Based on the water body type, the storage capacity is monitored by using a deep learning model corresponding to the water body type based on the integrated land and water digital elevation model, and the water storage capacity and dynamic changes of different water body types are calculated. Acquire multi-source collaborative data of the target water body, including: According to the preset preliminary survey plan, multi-source collaborative acquisition data of the target water body is obtained. The multi-source collaborative acquisition data includes macroscopic image data obtained by satellite remote sensing, high-resolution terrain and image data obtained by UAV aerial survey, control point and water surface elevation data obtained by ground measurement, and underwater terrain data obtained by underwater mapping. Extract the metadata and feature descriptors corresponding to the multi-source collaborative acquisition data. The feature descriptors include image resolution, point cloud density, control point distribution uniformity, and water depth range. Based on big data retrieval methods, surface water storage survey cases of different water body types are obtained to construct a historical experience database. Each survey case is used as a node, and water body type, geometric morphology features, data source features, environmental and working condition features, final contribution weight vector of each data source, and task objectives are used as node attributes. Feature similarity edges are constructed by calculating the feature similarity between nodes, and nodes are connected through the feature similarity edges to generate a graph corresponding to the historical experience database. Based on the feature descriptor, the initial node location is performed in the graph using similarity. A preset number of nodes are selected as the starting point set for the walk. Starting from each walk node, a random walk guided by feature similarity and task objective is performed along the feature edges and task edges in the graph. During the random walk, the weight vectors of each data source in the path nodes are recorded to generate a path weight sequence. The sequence is aggregated using an attention mechanism to generate a path weight vector. The weight vectors of all paths are integrated to generate an adaptive weight allocation scheme suitable for the target water body and the current task. The adaptive weight allocation scheme is fed back to the data acquisition system, and subsequent data acquisition is dynamically adjusted based on the weight of each data source.

2. The method for monitoring surface water storage based on multi-source synergy of air, space, land, and water as described in claim 1, characterized in that, Based on the aforementioned multi-source dataset, the water index method and deep learning are used to automatically extract and detect changes in the extent of surface water bodies, resulting in water body boundaries, including: Load the generated multi-source dataset with a unified spatiotemporal reference, and extract multi-temporal optical satellite images, multi-temporal synthetic aperture radar images, and high-resolution orthophotos acquired by UAVs. For parallel computation of water index, vegetation index and building index for optical images, for extraction of backscattering intensity, coherence, texture features and polarization decomposition features for synthetic aperture radar images, the original spectral bands, index features, synthetic aperture radar features and UAV orthophotos are stacked at the pixel level to generate feature cubes. The fusion feature cubes of all time phases are imported into a deep learning network containing a spectral index branch, a synthetic aperture radar feature branch, and a spatial context branch to generate a water body probability time series. A change detection method based on Bayesian update is then used to obtain the final determined water body probability map. The water body probability map is binarized, and the binary grid is converted into a preliminary vector boundary. The preliminary vector boundary is spatially registered and compared with the coordinates of key boundary points obtained by high-resolution UAV orthophotos and ground-based measurements. Based on preset rules, the boundary is refined according to the comparison results, and the spatial vector boundary of the surface water area is output.

3. The method for monitoring surface water storage based on multi-source synergy of air, space, and water as described in claim 2, characterized in that, Construct and train a deep learning network that includes a spectral index branch, a synthetic aperture radar feature branch, and a spatial context branch, including: A preset adaptive threshold is used to mark water body sample points and non-water body sample points on the multi-temporal optical satellite images using water body index, constructing a training sample set, and supplementing it with high-resolution orthophotos acquired by UAVs by extracting water body boundary samples through edge detection. A deep learning network is constructed that includes a spectral index branch, a synthetic aperture radar feature branch, and a spatial context branch. The network is trained using training samples. The spectral index branch is used to obtain the response patterns of different index combinations under different environments. The synthetic aperture radar feature branch is used to learn the performance of water bodies in radar images. The spatial context branch is used to learn the spatial structure features of water body boundaries. Attention fusion is used to weight and fuse the output features of the three branches. The output of the spectral index branch is used as prior knowledge. The learning attention of the synthetic aperture radar feature branch and the spatial context branch is dynamically adjusted. The probability map of the output pixel as water body is tested. When the network performance meets the preset standard, the trained deep learning network is obtained.

4. The method for monitoring surface water storage based on multi-source synergy of air, space, and water as described in claim 1, characterized in that, Using the water boundary as a spatial constraint, onshore and underwater topographic data are extracted from the multi-source dataset and input into a deep learning fusion model based on an encoder-decoder structure for water-land topographic fusion, including: Obtain the boundary of the land surface and water area, and extract the corresponding onshore and underwater topographic data from the multi-source dataset. Based on the boundary of the land surface and water area, establish a fusion buffer that extends to both sides of the land and water area according to a preset distance. Use the fusion buffer to spatially clip the onshore and underwater topographic data, and convert the clipped point cloud data into an initial digital elevation model. A deep learning fusion model is established based on the encoder-decoder structure. Two symmetrical encoder branches are set up to process the initial digital elevation models of the onshore and underwater terrain, respectively. Multi-scale dilated convolution is used in each encoder to extract multi-scale features, and cross attention is used to associate the features of the corresponding scales of the two encoder branches. The multi-scale features extracted from the two encoder branches are imported into the gated fusion unit to generate a dynamic fusion weight map for feature fusion. The fused multi-scale features are then imported into the decoder, where the original input data is initially classified into terrain to obtain a terrain semantic supervision signal. The terrain semantic supervision signal is used to guide the reconstruction of the fused multi-scale features, and a complete integrated digital elevation model of land and water is output.

5. A method for monitoring surface water storage based on multi-source synergy of air, space, and water as described in claim 4, characterized in that, Training a deep learning fusion model based on an encoder-decoder architecture includes: Collect complete measured data of land and water topography in the region, and perturb the region samples to construct paired samples of defect sample data and true value sample data; The paired samples are input into the constructed deep learning fusion model to obtain the predicted complete integrated water and land digital elevation model. Based on the predicted complete integrated water and land digital elevation model, the real complete integrated water and land digital elevation model and the input water boundary, the elevation regression loss, terrain structure loss and boundary continuity loss are calculated respectively, and the total loss is obtained by summing them according to the weights. The gradient of the total loss with respect to all weight parameters of the model is calculated by backpropagation, and the parameters are updated using an adaptive optimizer to minimize the total loss. After iterative training, the model is validated, and training ends when the performance meets the preset requirements.

6. The method for monitoring surface water storage based on multi-source synergy of air, space, and water as described in claim 1, characterized in that, For river bodies, the integrated land-water digital elevation model, water area vector boundaries, and time-series hydrological data are input into a river network dynamic model based on a graph neural network to calculate the segmented storage capacity and total storage capacity changes in different river segments, including: For river water bodies, the river centerline extracted based on the water area boundary divides the river into several continuous river segments. Each river segment is defined as a graph structure node. For each node, the average river width, average slope, cross-sectional area sequence, estimated riverbed roughness value, and channel storage capacity curve of the river segment are extracted from the integrated land and water digital elevation model to construct static features. The time series features of the river segment hydrological data are also extracted to construct dynamic features. Directed edges are defined based on the upstream and downstream relationships of nodes. A river system graph structure is established based on the directed edges and nodes. The static features of each node and the dynamic features of one time step are merged and input into a multilayer perceptron to be encoded into a node state embedding vector. The node state is embedded into a vector and input into a river network dynamic model based on a graph neural network. Graph convolution is used to pass messages based on the upstream node state and the dynamic weights of directed edges. Downstream nodes aggregate messages from upstream neighboring nodes and import them into a gated recurrent unit to update the node state. The updated node states are imported into the fully connected network to obtain the water volume change value and outflow prediction value of the river segment at the current time step. The real-time total storage of the river segment is calculated based on the theoretical static volume of the river segment at the current water level. Finally, the storage change process line of each river segment and the total storage change process line of the entire river are output in the time series.

7. The method for monitoring surface water storage based on multi-source synergy of air, space, land, and water as described in claim 1, characterized in that, For the reservoir water body, the integrated land-water digital elevation model, water area vector boundary, and multi-period water level observation data are input into the reservoir capacity model fused with a spatiotemporal attention mechanism to obtain the reservoir capacity value and predict future storage volume, including: In the integrated land and water digital elevation model of the reservoir, horizontal slices are made based on historical water level observation data and water area boundaries. The inundation range of the reservoir under the water level is calculated, an instantaneous water surface polygon is generated, the underwater topography of the instantaneous water surface polygon is extracted, a three-dimensional topographic instance of the reservoir at a specific water level is generated, and finally the reservoir topographic state sequence corresponding to the water level is obtained. For each reservoir 3D terrain instance, calculate the corresponding geometric features and terrain features to construct a morphological state vector, and encapsulate the water level observation value, morphological state vector and the reservoir capacity value observed at the same time point into a spatiotemporal sample. In the reservoir capacity model, the spatiotemporal samples are spatiotemporally fused and encoded to obtain the fusion state vector at each time point. Spatial morphology attention is introduced to learn the contribution weight of different topographic units of the reservoir to the total capacity under different water levels. Temporal memory attention is introduced to calculate the temporal dependence and memory state of the capacity change. After processing by the cascaded attention mechanism, the enhanced spatiotemporal state vector at each time point is obtained. The reservoir capacity value at each time point is output using a fully connected layer, and the water level forecast value at future time is obtained. The morphological state vector corresponding to the water level forecast value is imported into the reservoir capacity model to predict the future storage volume at future time.

8. A method for monitoring surface water storage based on multi-source synergy of air, space, and water as described in claim 1, characterized in that, For ponds and pools, under the constraints of the integrated land-water digital elevation model and the water area vector boundary, a water depth inversion model based on remote sensing imagery and a pixel volume accumulation algorithm are used to calculate the storage capacity of each pond and pool within the area, including: Based on the water area vector boundary, bottom topographic data corresponding to each pit is extracted from the integrated land and water digital elevation model, and spatially registered with the remote sensing image. According to the bottom topographic data and the estimated water surface elevation at the imaging time, the pseudo water depth true value is calculated for each pixel in the water area of ​​each pit to form a training sample set. A regional adaptive water depth inversion deep learning model is constructed, with pixel spectral features and bottom elevation as inputs, and soft constraints based on the physical equation of water optical transmission are introduced. The model is trained using the training sample set to learn the mapping relationship from spectrum and topographic features to water depth. The trained water depth inversion model is applied to the target pond water body, and the water depth value of each pixel in each pond is obtained by batch inversion. The inverted water depth of all pixels within the water boundary is multiplied by the pixel area to obtain the micro-volume of each pixel. The micro-volumes of all pixels are accumulated to obtain the total storage capacity of the pond.

9. A surface water storage monitoring system based on multi-source synergy of air, space, and water, characterized in that, To implement the surface water storage monitoring method based on multi-source collaboration of air, land, and water as described in any one of claims 1-8, the system includes: a multi-source collaborative acquisition and adaptive planning module, a multi-source heterogeneous data preprocessing and fusion management module, an intelligent water area boundary extraction and change detection module, an integrated land and water terrain intelligent fusion modeling module, a differentiated water storage intelligent calculation and prediction module, and a visualization analysis and decision-making module. The multi-source collaborative acquisition and adaptive planning module receives user input of water body type, target area and accuracy requirements to perform adaptive task planning and acquire multi-source collaborative acquisition data of the target water body. The multi-source heterogeneous data preprocessing and fusion management module preprocesses and unifies the spatiotemporal reference of the multi-source collaboratively collected data to generate a multi-source dataset with a unified spatiotemporal reference. The intelligent water area boundary extraction and change detection module, based on the multi-source dataset, uses the water index method and deep learning to automatically extract and detect changes in the surface water area, obtain the water area boundary, and identify the change area and intensity. The integrated land and water terrain intelligent fusion modeling module uses the water boundary as a spatial constraint, extracts onshore and underwater terrain data from the multi-source dataset, and inputs them into a deep learning fusion model based on an encoder-decoder structure to perform land and water terrain fusion, generate an integrated land and water digital elevation model, and generate a data confidence map to quantitatively evaluate the reliability of terrain data in different regions. The differentiated water storage intelligent calculation and prediction module uses the river dynamic model sub-engine, reservoir capacity model sub-engine and pit and pond survey model sub-engine corresponding to the water body type based on the integrated land and water digital elevation model to monitor the storage and calculate the water storage and dynamic changes of different water body types. The visualization analysis and decision-making module dynamically displays water body boundaries, terrain models, water storage volume and change processes, and provides storage data, early warning information and analysis reports through API.

Images