A seamless daily-scale surface water filling method based on precipitation constraint

Abstract

The present application provides a seamless daily scale surface water filling method based on precipitation constraint, comprising: obtaining optical remote sensing image, synthetic aperture radar image and daily scale precipitation sequence data of a study area; obtaining corresponding fusion water body probability graph based on the optical remote sensing image and the synthetic aperture radar image; dividing the fusion water body probability graph of the precipitation date into multiple constraint groups; for the fusion water body probability graph with missing pixels in each constraint group, traversing each missing pixel in the fusion water body probability graph, and sequentially searching the remaining fusion water body probability graph in order of high to low similarity to fill the missing pixel with effective water body probability value; for non-precipitation dates, linear interpolation is carried out based on the reconstructed fusion water body probability graph of adjacent precipitation dates to obtain the water body probability graph of non-precipitation days; threshold segmentation is carried out to obtain seamless daily scale surface water data. The seamless daily scale surface water data with continuity and hydrological rationality can be obtained.

Description

Technical Field

[0001] This invention relates to the field of remote sensing hydrological monitoring technology, and is particularly applicable to continuous surface water mapping in cloudy and rainy areas. Specifically, it relates to a seamless daily-scale surface water filling method based on precipitation constraints. Background Technology

[0002] Surface water is a crucial hydrological element for maintaining the balance of terrestrial ecosystems, identifying meteorological disasters, and supporting water resource management. With the intensification of global climate change and the increasing human activities such as water diversion, agricultural expansion, and reservoir regulation, natural water bodies have experienced varying degrees of shrinkage, fragmentation, and changes in their spatial and temporal distribution in many regions. High-resolution spatiotemporal monitoring of surface water fluctuations can not only provide fundamental data for assessing the impacts of human activities, refining water resource management, and issuing flood and drought early warnings, but also contribute to understanding watershed hydrological cycles and climate feedback mechanisms.

[0003] In some regions with dense river networks and numerous lakes, precipitation exhibits a significant seasonal concentration, making them typical areas influenced by rainfall input, evaporation loss, and human intervention. Surface water in these regions responds rapidly to short-term rainfall, often displaying dynamic changes of rapid expansion and contraction. Therefore, obtaining continuous, diurnal-scale, and reliable surface water observation data is crucial for regional water resource allocation, drought identification, and hydrological anomaly monitoring. However, limitations imposed by persistent cloud and rain, complex surface conditions, and satellite observation capabilities mean that while existing long-term surface water products can be used for trend analysis, they still struggle to provide continuous, gap-free, diurnal-scale high-resolution water body sequences, thus restricting the detailed characterization of short-term hydrological processes and drought development.

[0004] Currently, surface water remote sensing monitoring mainly relies on two types of data sources: optical remote sensing and synthetic aperture radar (SAR). Optical sensors such as MODIS (Moderate Resolution Imaging Spectroradiometer) have high temporal resolution but low spatial resolution, making it difficult to identify small water bodies. The Landsat series satellites can provide long historical observation records, but the revisit period is long, making it difficult to achieve high-frequency dynamic monitoring. Although Sentinel-2 has high spatial resolution and a short revisit period, it is still difficult to achieve continuous diurnal scale monitoring alone in cloudy and rainy areas. In contrast, Sentinel-1 data has all-weather, all-day observation capabilities and can penetrate clouds and precipitation, compensating for the shortcomings of optical imagery to some extent. However, SAR imagery is susceptible to surface scattering characteristics, geometric distortion, and complex ground backgrounds, and misclassification may still occur in highly heterogeneous areas. At the same time, differences in orbital period, imaging angle, and spatial resolution among different sensors can also lead to inconsistencies in observation results for the same area.

[0005] In recent years, methods for remediating gaps in global surface water products and water bodies have been continuously developing. For example, the Dynamic World global surface water product based on Landsat imagery can provide near real-time land cover classification at a resolution of 10 meters, including water body categories. However, such products mainly rely on optical imagery and are easily affected by cloud cover, leading to observation gaps. To improve data integrity, existing studies have proposed methods such as spatiotemporal neighborhood similarity, image similarity reconstruction, stepwise incompleteness, and self-supervised learning. Some studies have also attempted to fuse optical and SAR imagery to generate bi-monthly scale water body products.

[0006] While the aforementioned methods improve the completeness of surface water product data, most still rely primarily on empirical statistics or similarity rules, lacking physical constraints that reflect hydrological processes. In particular, they rarely incorporate the direct impact of precipitation on surface water changes into the filling process. Furthermore, existing methods often fill local gaps or specific periods, failing to establish a continuous and stable daily-scale surface water product production framework. Their temporal resolution is mostly limited to monthly or bi-monthly scales, which is insufficient to meet the needs of short-term drought and flood responses and refined water resource management. Therefore, it is necessary to develop a method that integrates optical and SAR multi-source remote sensing data, incorporates precipitation constraints, and can generate seamless daily-scale surface water sequences to improve the dynamic monitoring capabilities of surface water in complex cloud and rain regions. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a seamless daily-scale surface water filling method based on precipitation constraints. Leveraging the spectral recognition advantages of optical remote sensing imagery and the all-weather observation advantages of synthetic aperture radar imagery, a fused water body probability map is generated through a water body probability prediction model. Precipitation intensity and seasonal attributes of daily-scale precipitation sequence data are used as physical constraints in the filling process. Pixel-by-pixel retrieval filling and temporal interpolation are performed at the probability level, thereby obtaining continuous and hydrologically sound seamless daily-scale surface water data.

[0008] To achieve the above objectives, the specific technical solution of the present invention is as follows:

[0009] In a first aspect, the present invention provides a seamless daily-scale surface water refill method based on precipitation constraints, comprising:

[0010] Acquire optical remote sensing images, synthetic aperture radar images, and diurnal precipitation sequence data of the study area;

[0011] Based on the optical remote sensing image and the synthetic aperture radar image, an optical water probability map and a synthetic aperture radar water probability map for each date are generated using a water body probability prediction model. The optical water probability map and the synthetic aperture radar water probability map for the same date are then fused together to obtain the corresponding fused water probability map.

[0012] Based on the precipitation intensity and seasonal attributes of the daily precipitation sequence data, the fused water body probability map of precipitation dates is divided into multiple constraint groups;

[0013] For each of the constraint groups containing missing pixels in the merged water body probability map, the similarity between the merged water body probability map and the other merged water body probability maps in the same constraint group is calculated in the effective common region. Each missing pixel in the merged water body probability map is traversed, and the other merged water body probability maps are searched sequentially according to the similarity from high to low. When the other merged water body probability map has a valid water body probability value at the corresponding pixel position, the missing pixel is filled with the valid water body probability value to obtain the reconstructed merged water body probability map.

[0014] For non-precipitation dates, linear interpolation is performed on the reconstructed and merged water body probability map based on adjacent precipitation dates to obtain the water body probability map for non-precipitation days;

[0015] Threshold segmentation is performed on the reconstructed and merged water body probability map and the water body probability map on non-precipitation days to obtain seamless daily-scale surface water data.

[0016] Furthermore, the formula for calculating the similarity is:

[0017] ;

[0018] in, Indicates similarity; and Represents the weighting coefficient, and Greater than ; Represents the structural similarity index; The similarity component based on mean squared error is calculated as follows: , This represents the mean square error.

[0019] Furthermore, the precipitation intensity is divided into seven levels based on 24-hour precipitation: no precipitation, light rain, moderate rain, heavy rain, rainstorm, torrential rain, and extremely heavy rain.

[0020] Furthermore, the seasonal attributes include the flood season and the non-flood season.

[0021] Furthermore, the optical remote sensing image includes surface reflectance data formed by fusing the Landsat series and Sentinel-2 series, and the input bands include six bands: red light, green light, blue light, near-infrared, shortwave infrared 1 and shortwave infrared 2.

[0022] The synthetic aperture radar imagery includes terrain-corrected Sentinel-1 GRD data, with input features including two polarization channels: VV and VH.

[0023] Furthermore, the water body probability prediction model adopts a dual-source U-Net model, which is used to receive the optical remote sensing image and the synthetic aperture radar image, respectively;

[0024] Each of the U-Net models includes an encoder, a bottleneck layer, a decoder, and a convolutional layer; the encoder includes a multi-level convolutional module and a max-pooling layer; the decoder includes a transposed convolutional upsampling layer and skip connections connected to the corresponding layer of the encoder.

[0025] Furthermore, the linear interpolation includes:

[0026] Using the reconstructed and merged water body probability map of adjacent precipitation dates as endpoints, linear interpolation is performed on non-precipitation dates between endpoints according to date distance, and the interpolation results are constrained to a probability range of 0 to 1.

[0027] Furthermore, before performing linear interpolation on the reconstructed and merged water body probability map based on adjacent precipitation dates for non-precipitation dates to obtain the water body probability map for non-precipitation days, the method further includes:

[0028] After the filling is completed, if there are still missing pixels in the reconstructed and merged water body probability map in any of the constraint groups, then based on the time position of the reconstructed and merged water body probability map with missing pixels, the reconstructed and merged water body probability maps with adjacent time are searched step by step forward and backward, and the missing pixels are retrieved and filled pixel by pixel according to the degree of temporal proximity.

[0029] Furthermore, the threshold segmentation uses the bimodal method to determine the threshold, and the reconstructed and merged water body probability map and the non-precipitation day water body probability map are converted into a binary surface water map as seamless daily-scale surface water data.

[0030] Secondly, the present invention provides a seamless daily-scale surface water refill system based on precipitation constraints, comprising:

[0031] The data acquisition module is used to acquire optical remote sensing images, synthetic aperture radar images, and daily precipitation sequence data of the study area;

[0032] The probability prediction and fusion module is used to generate optical water probability maps and synthetic aperture radar water probability maps for each date based on the optical remote sensing image and the synthetic aperture radar image, using a water probability prediction model, and to perform complementary fusion of the optical water probability map and the synthetic aperture radar water probability map for the same date to obtain the corresponding fused water probability map.

[0033] The precipitation constraint grouping module is used to divide the fused water body probability map of the precipitation date into multiple constraint groups based on the precipitation intensity and seasonal attributes of the daily precipitation sequence data.

[0034] The similarity imputation module is used to calculate the similarity between the merged water body probability map and other merged water body probability maps in the same constraint group in the effective common region for each merged water body probability map with missing pixels in each constraint group; iterates through each missing pixel in the merged water body probability map, and sequentially searches the other merged water body probability maps according to the similarity from high to low; when the other merged water body probability map has a valid water body probability value at the corresponding pixel position, the missing pixel is filled with the valid water body probability value to obtain the reconstructed merged water body probability map;

[0035] The temporal interpolation module is used to perform linear interpolation on the reconstructed and merged water body probability map based on adjacent precipitation dates for non-precipitation dates to obtain the water body probability map for non-precipitation days.

[0036] The data output module is used to perform threshold segmentation on the reconstructed and merged water body probability map and the non-precipitation day water body probability map to obtain seamless daily-scale surface water data.

[0037] Compared with the prior art, the advantages of the present invention are:

[0038] (1) This invention uses the complementary fusion of optical remote sensing images and synthetic aperture radar images at the probability level, and uses precipitation intensity and seasonal attributes as physical constraints to fill in the time series, increasing the effective pixel ratio of the study area from 27.04% of the original optical image to 99.71%, which is 41.51% higher than the traditional method. It achieves almost complete coverage of daily continuous observation, effectively solves the problem of discontinuity in long-term observation caused by clouds, rainfall and satellite orbits, and is suitable for cloudy and rainy areas and areas with rapid hydrological response, significantly improving the spatiotemporal integrity and data availability of surface water products.

[0039] (2) Traditional gap filling methods mostly rely on statistical similarity or temporal proximity, lacking constraints on the formation process of surface water. This invention uses precipitation intensity and seasonal attributes as grouping constraints, so that the probability map of the merged water body with missing pixels is in the same or similar hydrological and meteorological background as the other probability maps of the merged water body, ensuring that the filling results conform to the physical mechanism of the response law of surface water to precipitation input.

[0040] (3) This invention does not directly perform hard classification on images, but rather performs fusion and completion at the probability level throughout the process, from water body probability prediction, multi-source probability fusion, pixel-by-pixel probability filling within the constraint group to non-precipitation day probability interpolation. It outputs pixel-level water body probability instead of hard classification results, effectively preserving the gradual change information of water body boundaries and pixel-level uncertainty, avoiding boundary ambiguity and category confusion caused by binary hard classification, and improving spatial accuracy and stability under complex terrain conditions.

[0041] (4) By constructing a surface water replenishment method based on precipitation constraints, this invention realizes seamless daily-scale continuous production of surface water products in complex cloud and rain areas. This product can reliably capture the rapid rise and fall of surface water caused by heavy rainfall, short-term drought, irrigation and drainage and water conservancy scheduling, providing a key data foundation for early identification of short-term drought, flood response monitoring, reservoir, lake and pond scheduling assessment and agricultural water resources management. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart of a seamless daily-scale surface water refill method based on precipitation constraints provided in an embodiment of the present invention;

[0044] Figure 2A schematic diagram of the architecture of the dual-source U-Net model provided in an embodiment of the present invention;

[0045] Figure 3 A flowchart illustrating the precipitation-constrained filling method provided in an embodiment of the present invention;

[0046] Figure 4 This is a comparison chart for missing simulation verification provided in an embodiment of the present invention, wherein, Figure 4 a represents the original true value, simulated gaps, and seamless surface water under striped missing values; Figure 4 b represents the original true value, simulated gaps, and seamless surface water under the condition of missing rectangular blocks; Figure 4 c and Figure 4 d represents the original true value, simulated gap, and seamless surface water under circular defects; Figure 4 e is a comparison chart of the calculation results of the quantitative accuracy index;

[0047] Figure 5 A comparison chart showing the correlation between the surface water area filled using the method of this embodiment and the measured water level at Jianli Hydrological Station;

[0048] Figure 6 This is a comparison diagram of the coverage pixel ratio before and after filling in the method of this invention and the conventional method;

[0049] Figure 7 This is a comparison chart showing the quantitative accuracy index calculation results of the method of this embodiment of the invention with those of Comparative Examples 1, 2 and 3. Detailed Implementation

[0050] To enable those skilled in the art to clearly and completely understand the technical solution of the present invention, the present invention will be further described in detail below with reference to embodiments. Obviously, the embodiments described herein are only for explaining the present invention and are not intended to limit the scope of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0051] To address the challenges in existing surface water remote sensing monitoring, such as data gaps caused by cloud and rain obstruction in optical imagery, misclassification of synthetic aperture radar (SAR) imagery under complex surface conditions, and the lack of hydrophysical constraints (especially precipitation-driven mechanisms) in current infilling methods, which hinder the generation of continuous, diurnal, and seamless high-resolution surface water sequences, this invention proposes a seamless diurnal surface water infilling method based on precipitation constraints. Leveraging the spectral recognition advantages of optical remote sensing imagery and the all-weather observation advantages of SAR imagery, this method generates a fused water probability map using a water body probability prediction model. Precipitation intensity and seasonal attributes from diurnal precipitation sequence data are used as physical constraints in the infilling process. Pixel-by-pixel infilling and temporal interpolation are performed at the probabilistic level, resulting in continuous and hydrologically sound seamless diurnal surface water data. This significantly improves the completeness and accuracy of dynamic surface water monitoring in cloudy and rainy areas. The infilled results exhibit high hydrological consistency with measured water levels and can effectively support applications such as short-term drought and flood response, water resource allocation, and hydrological anomaly monitoring.

[0052] Example 1

[0053] This invention proposes a seamless daily-scale surface water refill method based on precipitation constraints, see reference. Figure 1 Specifically, this may include the following steps:

[0054] S1. Data Collection and Preprocessing. Acquire optical remote sensing images of the study area (i.e., Figure 1 HLS remote sensing imagery and synthetic aperture radar imagery (i.e., HLS remote sensing imagery) Figure 1 The data includes Sentinel-1 remote sensing images and diurnal precipitation sequence data. The optical remote sensing images consist of surface reflectance data (such as HLS products) fused from the Landsat and Sentinel-2 series, hereinafter referred to as HLS remote sensing images. These provide multispectral water body identification information, with input bands including red, green, blue, near-infrared, shortwave infrared 1, and shortwave infrared 2. The synthetic aperture radar images include topographically corrected Sentinel-1 GRD (Ground Range Detected) data, hereinafter referred to as Sentinel-1 SAR images, used to fill observational gaps under cloud and rain conditions. Input features include two polarization channels: VV and VH. The diurnal precipitation sequence data includes precipitation intensity and seasonal attributes to provide 24-hour precipitation information, used to construct physical constraints for subsequent filling processes. In addition, this embodiment also acquires the Global Surface Water Mapping Layers (GSW) dataset (JRC Global Surface Water Mapping Layers) and auxiliary data for accuracy verification or hydrological consistency analysis, including high-resolution Planet imagery and daily water level data from hydrological stations.

[0055] In this embodiment, input images for all dates (including HLS remote sensing images and Sentinel-1 SAR images) undergo preprocessing including coordinate system 1, spatial resampling, band selection, and effective observation mask generation. For example, coordinate system 1 is set to GCS_WGS_1984 (geographic coordinate system, based on the WGS84 ellipsoid); HLS remote sensing images and Sentinel-1 SAR images are uniformly resampled to a preset spatial resolution (e.g., 30 m), and an effective observation mask is constructed for each date's image; the maximum possible water body range constraint is constructed using the intersection of the maximum inundation extent layer of the global surface water dataset and the effective observation area; six bands are selected for HLS remote sensing images: red, green, blue, near-infrared, shortwave infrared 1, and shortwave infrared 2; and two bands are selected for Sentinel-1 SAR images: VV and VH.

[0056] S2. Water Body Probability Prediction. Based on optical remote sensing imagery and synthetic aperture radar (SAR) imagery, an optical water body probability map and a SAR water body probability map for each date are generated using a water body probability prediction model. The optical water body probability map and the SAR water body probability map for the same date are then complementaryly fused to obtain the corresponding fused water body probability map.

[0057] Preferably, see Figure 2This embodiment employs a dual-source U-Net model for water body probabilistic prediction, receiving optical remote sensing imagery and synthetic aperture radar (SAR) imagery respectively. To adapt to different data sources, the number of input channels is set as a flexible parameter, supporting dual-channel Sentinel-1 SAR imagery and six-channel HLS remote sensing imagery. Each U-Net model includes an encoder, a bottleneck layer, a decoder, and convolutional layers. The encoder comprises multi-level convolutional modules and max-pooling layers, specifically four convolutional modules. Each module contains two 3×3 convolutional layers followed by a ReLU activation function, and max-pooling layers are applied after each module to progressively downsample the feature map and capture hierarchical features. The number of feature channels increases successively from 64 to 128, 256, and 512. The bottleneck layer, with 1024 convolutional filters, connects the encoder and decoder, providing high-level global features crucial for refined water body extraction. The decoder consists of transposed convolutional upsampling layers and skip connections connected to the corresponding encoder layers. The transposed convolutional upsampling layers progressively upsample the feature maps, and the skip connections link them to the corresponding encoder feature maps, ensuring the preservation of fine spatial details and contextual semantic information during reconstruction. The number of feature channels in the decoder decreases sequentially from 512 to 64, gradually restoring spatial resolution. Finally, a 1×1 convolutional layer projects the feature maps onto a single-channel water probability map, generating pixel-by-pixel surface water probabilities; that is, the output pixel value represents the probability that the pixel belongs to surface water.

[0058] In this embodiment, the dual-source U-Net model described above was trained using the following method to acquire water sample data, including Sentinel-1 and Sentinel-2 images and their corresponding binarized water masks. All images were segmented into 256×256 pixel sizes and randomly divided into training (70%), validation (20%), and test (10%) sets. The model training parameters were set as follows: batch size of 8, learning rate of 0.0001, binary cross-entropy as the loss function, and epoch of 50. Since HLS remote sensing images contain six bands and Sentinel-1 SAR images contain two polarization channels, VV and VH, two separate U-Net models were trained: one with six optical remote sensing image input channels and the other with two synthetic aperture radar image input channels.

[0059] Based on the water body probability prediction model trained above, HLS remote sensing imagery (i.e. Figure 2 HLS time-series images) and Sentinel-1 SAR images (i.e. Figure 2 The S1 time-series imagery is input into the water body probability prediction model to obtain the optical water body probability map for each date (i.e., Figure 2 HLS water probability map and synthetic aperture radar water probability map (i.e. Figure 2 In the S1 water body probability map, the predicted region in each water body probability map is constrained by the maximum possible water body range and the intersection of the effective observation region of each image.

[0060] Furthermore, when gaps occur in the optical water probability map due to clouds, cloud shadows, rainfall, or invalid sensor observations, they are supplemented using a synthetic aperture radar water probability map from the same day. Since Sentinel-1 SAR imagery can acquire surface information under cloud and rain conditions, this step can significantly increase the proportion of effective pixels in areas with frequent precipitation, resulting in a fused water probability map from the same date.

[0061] S3. Water body probabilistic infilling based on precipitation constraints. This embodiment introduces precipitation intensity and seasonal attributes as physical constraints, divides the fused water body probability map into multiple constraint groups, performs similarity infilling within each constraint group, and then supplements it with temporal proximity infilling and linear interpolation to finally achieve seamless reconstruction of daily-scale surface water data. Specifically, it includes the following steps:

[0062] S31. Based on the precipitation intensity and seasonal attributes of daily precipitation sequence data, the merged water body probability map of precipitation dates is divided into multiple constraint groups. Regarding precipitation intensity, it can be classified according to precipitation amount. For example, the "Precipitation Classification" standard (GB / T 28592-2012) can be followed. This standard defines seven precipitation levels based on 24-hour precipitation: 0-0.1 mm, 0.1-10 mm, 10-25 mm, 25-50 mm, 50-100 mm, 100-250 mm, and >250 mm, corresponding to no precipitation (or scattered light rain), light rain, moderate rain, heavy rain, torrential rain, extremely heavy rain, and exceptionally heavy rain, respectively. Regarding seasonal attributes, the flood season and non-flood season can be used as the representative quantities of seasonal attributes. (See also...) Figure 3 For all days with precipitation (daily rainfall > 0.1 mm), the above classification criteria can be applied to obtain a maximum of twelve constraint groups (6 precipitation levels × 2 seasons). Non-precipitation days are not included in the grouping and will be handled through subsequent linear interpolation. Of course, the actual number of groups should be determined based on the actual situation; for example, if there is no rainfall in the 10-25 mm range in the study area, then ten constraint groups will be obtained.

[0063] Taking the Jianghan Plain as an example, this region has a dense river network and numerous lakes and ponds. Surface water is significantly affected by precipitation, evaporation, agricultural irrigation, and water conservancy management, and it belongs to the Yangtze River Basin. According to the definition provided by the China Meteorological Administration, the flood season in the Yangtze River Basin is from May to October, and the remaining months are the non-flood season. The study period is from 2015 to 2024. Based on the actual conditions of the study area, some rainfall levels are missing. Therefore, for all dates with precipitation (i.e., daily rainfall greater than 0.1 mm), eight constraint groups were obtained according to the above definition.

[0064] Precipitation is the primary driver of surface water expansion. Under the same precipitation intensity level, the spatial response patterns of surface water (such as inundation range and water connectivity) exhibit statistical similarity. Simultaneously, there are significant differences in soil saturation, evaporation intensity, vegetation cover, and human interventions (such as reservoir discharge and farmland irrigation) between flood season and non-flood season. Classification based on seasonal attributes can further eliminate interference from changes in the environmental background. The combination of precipitation intensity and seasonal attributes constructs physical constraints that conform to hydrological patterns, grouping fused water body probability maps with similar hydrological response characteristics into the same constraint group. This ensures that the reference image and the image to be infilled in the subsequent infilling process have similar background conditions and dynamic patterns, thereby improving the rationality of similarity matching and infilling accuracy.

[0065] S32, Similarity-based cross-image filling.

[0066] First, for the merged water body probability map (i.e., the image to be filled mentioned above) with missing pixels in each constraint group, the similarity between the merged water body probability map and the other merged water body probability maps (i.e., the reference images mentioned above) in the effective common region is calculated. The similarity index is specifically defined as follows: [Image definition missing]. (i.e., the image to be filled) and the image The mean square error of (any reference image) in the effective common region is :

[0067] (1)

[0068] In the above formula (1), This represents the total number of pixels within the valid common area; and Representing images respectively and images In the The probability value of water bodies at each pixel.

[0069] To standardize the units, the mean squared error is transformed into a similarity form (a larger value indicates greater similarity), and similarity components based on the mean squared error are defined. for:

[0070] (2)

[0071] Considering that error-based metrics alone are insufficient to reflect the structural features of images, a structural similarity index is further introduced. Its calculation form is:

[0072] (3)

[0073] In the above formula (3), and Representing images , The mean over the effective common region; and Indicates variance; Represents covariance; and represents the stability constant, exemplarily taken as 0.0001 and 0.0009 respectively.

[0074] The similarity score for this embodiment is constructed by weighting the structural similarity index and the similarity component based on mean squared error. The formula for calculating the similarity score is as follows:

[0075] (4)

[0076] In the above formula (4), Indicates similarity; Represents the structural similarity index; This represents the similarity components based on mean squared error; and Represents the weighting coefficient, and Greater than To highlight the consistency of the image's spatial structure, for example, , In other words, when calculating similarity, the structural similarity index has the main weight, while the similarity component based on mean square error is used as an auxiliary factor, so as to balance the consistency of image structure and the accuracy of spectral values.

[0077] Then, refer to Figure 3 The process iterates through each missing pixel in the merged water body probability map, and sequentially searches the other merged water body probability maps according to the similarity from high to low. When the other merged water body probability maps have a valid water body probability value at the corresponding pixel position, the missing pixel is filled with the valid water body probability value to obtain the reconstructed merged water body probability map.

[0078] Furthermore, in this embodiment, each completed reconstructed and merged water body probability map is reassigned to its respective constraint group and participates in subsequent filling as other merged water body probability maps, so as to enrich the available information in the constraint group and improve the filling effect of subsequent large-area continuous gap areas.

[0079] S33. After completing step S32 above, if any of the reconstructed and merged water body probability maps in any constraint group still have missing pixels, then based on the time position of the reconstructed and merged water body probability map with missing pixels, progressively search forward and backward for reconstructed and merged water body probability maps with adjacent time periods, and retrieve and fill missing pixels pixel by pixel according to the degree of temporal proximity. This method is based on the assumption of temporal series continuity, that is, adjacent temporal phase images have high similarity in spatial distribution and probability characteristics. By prioritizing the temporal phase with the closest time distance, it maximizes the use of the continuous information of the time series, thereby improving the filling effect of residual missing areas.

[0080] S34. For non-precipitation days, linear interpolation is performed on the reconstructed and merged water body probability maps based on adjacent precipitation days to obtain the water body probability map for non-precipitation days. See details in [link to documentation]. Figure 3 Using the reconstructed and merged water body probability maps of adjacent precipitation dates as endpoints, linear interpolation is performed on non-precipitation dates between endpoints based on date distance, and the interpolation results are constrained to a probability range of 0 to 1. This step is based on the assumption of temporal continuity of water body probability, that is, the water body probability between two adjacent precipitation events gradually changes over time. By generating water body probability maps for non-precipitation days through linear interpolation, both the temporal continuity at the daily scale and the avoidance of interpolation results exceeding the physically reasonable range are ensured.

[0081] S35. Threshold segmentation is performed on the reconstructed and merged water body probability map and the non-precipitation daily water body probability map to obtain seamless daily-scale surface water data. Preferably, the threshold segmentation uses the bimodal method to determine the threshold, and the reconstructed and merged water body probability map and the non-precipitation daily water body probability map are converted into binary surface water maps as seamless daily-scale surface water data.

[0082] Compared to existing infilling methods that rely solely on spatiotemporal statistical similarity, this method introduces precipitation intensity and seasonal attributes as hydrophysical constraints into the infilling process for the first time. This ensures that the image to be infilled has similar background conditions and dynamic patterns to the reference image, significantly improving the rationality and accuracy of the infilling. At the same time, through a progressive infilling strategy that progresses from intra-group similarity infilling to temporal proximity infilling and then to linear interpolation, complete coverage from precipitation days to non-precipitation days is achieved. This method can generate continuous and stable daily-scale surface water data, making it particularly suitable for plain water network areas with frequent cloud cover and rain and human activity.

[0083] Step S4: Accuracy verification and hydrological consistency assessment.

[0084] (1) Accuracy verification

[0085] To verify the reliability of this method, in this embodiment, high-resolution reference images from different seasons are selected as independent validation data, and the classification results are evaluated using overall precision, accuracy, recall, and F1 score. Taking the Jianghan Plain as an example, a grid is randomly selected in each of the four seasons: spring (May 2, 2022), summer (June 9, 2023), autumn (October 24, 2021), and winter (December 28, 2024). High-resolution Planet images with complete coverage and cloud cover less than 10% are downloaded, and their NDWI (Normalized Difference Water Index) is calculated. The OTSU method (an adaptive thresholding algorithm) is then applied to obtain a threshold for binarization, resulting in the final surface water raster, which is used as the high-resolution ground truth to validate the seamless surface water data. For each period, a stratified random sampling rule (1500 sampling points, including 1000 water bodies and 500 non-water bodies) is designed to construct a confusion matrix. The overall precision, accuracy, recall, and F1 score were calculated, as shown in Table 1 below. It can be seen that the scores of these evaluation indicators are all high, indicating the accuracy of the method in this embodiment.

[0086] Table 1. Accuracy indicators of surface water data from refilled and higher-resolution imagery at different dates.

[0087]

[0088] (2) Missing simulation verification

[0089] Furthermore, the recovery ability of this method for different deletion morphologies was tested by randomly simulating strip-shaped, block-shaped, or circular deletion regions. Figure 4 As shown, taking the Jianghan Plain as an example, the first row contains the original ground truth (the purple area represents the actual distribution of surface water, and the white area is the background). Specifically, dates with no missing data in the original image were selected (2018-09-15, 2021-11-10, 2023-03-13, and 2024-01-27 correspond to...). Figure 4 (a, 4b, 4c, 4d) use the original complete image as the truth value; Figure 4 The second line simulates gaps and sets three types of missing data: strip-shaped missing data, rectangular missing data, and circular missing data, which are used to simulate three missing data scenarios: scan line gaps, data block corruption, and natural cloud occlusion, respectively. Figure 4The third row shows the seamless surface water reconstructed using the method of this invention (the dark blue area represents the restored surface water). By comparing the original ground truth with the seamless surface water, it can be found that the reconstruction results maintain a high degree of consistency with the original ground truth in complex landforms such as the main channel of the river, the main body of the lake, the confluence of river networks, and meandering river sections. No systematic boundary misalignment, water body shrinkage, or category confusion due to differences in the morphology of missing areas occurs. Furthermore, the method of this invention demonstrates excellent filling and restoration capabilities under various typical missing morphologies, including strip-shaped, rectangular block-shaped, and circular shapes. This proves that the method of this embodiment can achieve high-precision spatial reconstruction under various missing morphologies, with missing areas being continuously and reasonably restored, and excellent average accuracy. Quantitative accuracy indicators were calculated for the above three scenarios, and the results are as follows: Figure 4 As shown in e, the overall accuracy (OA), precision, recall, and F1 score reached 0.99, 0.98, 0.99, and 0.99, respectively. It is worth noting that the accuracy for the September 15, 2018 date was slightly lower (0.97), mainly reflecting that the model over-identified a small number of water bodies in some boundary areas of the image in that period, i.e., a slight false positive. Its recall and F1 score remained at 0.99, indicating that the model was generally able to capture water body information completely without significant missed detections. In summary, the overall method can balance accuracy and completeness when facing missing data of different shapes and scales, demonstrating strong robustness and generalization ability.

[0090] (3) Hydrological consistency test

[0091] To evaluate the hydrological consistency of the filling results, a buffer zone was established around the Jianli hydrological station in the Jianghan Plain, and time-series data of surface water area within a 5km buffer zone near the Jianli station were acquired. This data was then compared with daily measured water level data for hydrological consistency verification. To analyze the consistency between the two at different time scales, a time series decomposition method was used to decompose the original series into trend, seasonal, and residual terms for comparative analysis. (See also...) Figure 5 By comprehensively comparing and analyzing the long-term trend, seasonal fluctuations, and short-term residuals across three time scales, the seamless daily-scale surface water area reconstructed in this invention shows good consistency with the measured water level at the Jianli hydrological station in terms of long-term trend, seasonal fluctuations, and short-term residuals. In particular, the Pearson correlation coefficient reaches 0.60 and passes the significance test, indicating that the method of this embodiment can effectively preserve the seasonal driving characteristics of the actual hydrological process and respond well to the seasonal rise and fall of river water levels. Therefore, the seamless daily-scale surface water data obtained based on the method of this invention can truly reflect the actual hydrological process, possessing high hydrological consistency and physical rationality, and can provide a reliable data foundation for short-term drought and flood identification and water resource allocation assessment.

[0092] See Figure 6 This is a comparison of the coverage pixel ratio before and after filling in the data using the method of this embodiment and the traditional method. Taking the Jianghan Plain as an example, the effective pixel ratio of the original optical image in the study area is only 27.04% on average, indicating that due to factors such as climate conditions, observation limitations, and data quality, there are significant spatial gaps in optical observations, and surface water information is incomplete. The traditional method (i.e. Figure 6 In the S1 method, Sentinel-1 imagery is typically used for time-series completion. However, although the effective pixel ratio after SAR image completion is 31.16% higher than the original result, improving data gaps to some extent, it is still only 58.20%. This indicates that relying solely on radar image completion is insufficient to recover continuous and complete temporal information. Furthermore, after using the method proposed in this embodiment for completion, the average effective pixel ratio reaches 99.71%, a further improvement of 41.51% compared to the traditional method S1. Only a few dates still have large gaps, and the overall completion effect is significant.

[0093] Furthermore, to further evaluate the contribution of the seasonal constraint of precipitation and the similarity calculation method to the accuracy of the water body filling results, three experiments were designed as controls, where Ours method represents the method used in this embodiment.

[0094] Comparative Example 1

[0095] Comparative Example 1 did not use precipitation constraints (i.e., it did not set precipitation intensity and seasonal attributes as physical constraints). For missing pixels in the image, it directly filled them according to the similarity of all images within the ten days in time. The similarity was calculated using the above formula (4), which is consistent with Ours' method.

[0096] Comparative Example 2

[0097] Comparative Example 2 uses precipitation constraints, but the similarity calculation sets the weight coefficient of the similarity component based on mean square error to 1, and sets the weight coefficient of the structural similarity index to 0, that is, only the similarity component based on mean square error is calculated. .

[0098] Comparative Example 3

[0099] Example 3 uses precipitation constraints, but the similarity calculation sets the weight coefficient corresponding to the structural similarity index to 1, and sets the weight coefficient of the similarity component based on mean square error to 0, that is, only the structural similarity index is calculated. .

[0100] To eliminate the influence of seasonal factors on accuracy and evaluation results, 16 images covering all four seasons (spring, summer, autumn, and winter) were randomly selected from four experimental grids. The selection of these dates required that the Planet high-resolution image coverage be above 95% and cloud cover below 10% to ensure high-quality reference images. NDWI of the images was calculated, and the surface water obtained by thresholding using the OTSU method was used as a high-resolution map to validate the surface water data from different experiments. Finally, a stratified random sampling rule (1500 sampling points, including 1000 water bodies and 500 non-water bodies) was designed for all temporal results to construct a confusion matrix, and four accuracy evaluation metrics were calculated, including overall accuracy (OA), precision, recall, and F1 score.

[0101] Table 2 Comparison of different methods and accuracy evaluation indicators

[0102]

[0103] As shown in Table 2 above, the accuracy of the filled water body obtained using Comparative Example 1 is relatively low, with an overall precision (OA) of only 0.7892, a recall of only 0.6921, and an F1 score of 0.8086. After incorporating precipitation constraints, the accuracy of the water body filling results is significantly improved. Combined with... Figure 7 Specifically, Comparative Example 1 showed low recall and F1 score on most dates, indicating that relying solely on image similarity for matching is easily affected by seasonal water level changes and short-term precipitation events, leading to the omission of some real water bodies. In contrast, Comparative Examples 2, 3, and Ours, which have precipitation constraints, had F1 scores exceeding 0.90 on most dates, with some even reaching above 0.95, except for a few validation dates where the F1 score was slightly below 0.9. This shows that seasonal precipitation information can effectively constrain the selection of reference images and improve the consistency between the infill results and the actual water body distribution. Regarding different similarity calculation methods, using only the similarity component based on mean squared error (Comparative Example 2) and only the structural similarity index (Comparative Example 3) both achieved good accuracy, but Comparative Example 2's accuracy was slightly lower than Comparative Example 3. Ours method shows a slight improvement in accuracy compared to Comparative Examples 2 and 3. This indicates that the similarity component based on mean square error focuses more on pixel grayscale differences, while the structural similarity index pays more attention to local structural information. It has certain advantages in maintaining water body boundaries and spatial morphology. At the same time, considering spectral differences and spatial structural similarity helps to improve the reliability of reference image matching.

[0104] Combination Figure 7Looking at the changes in accuracy across different dates, there are certain fluctuations in the evaluation metrics across different regions and dates. This is related to factors such as regional water area, image quality, and local ground cover background. For the three metrics of overall accuracy, recall, and F1 score, Scheme 1 still exhibits the lowest accuracy across all dates, while the accuracy is significantly improved after adding precipitation constraints. However, for the precision metric, the higher precision under the unconstrained condition is because the number of pixels filled is limited, resulting in significant incomplete coverage and thus severe missed detections, leading to a lower recall but higher precision.

[0105] In summary, both the precipitation constraint and the composite similarity metric set in this embodiment have a positive effect on the accuracy of water body filling, with the precipitation constraint having a more significant impact on improving filling accuracy. Compared with methods without precipitation constraints or with a single similarity index, the comprehensive method adopted in this paper shows higher stability and accuracy in different regions and on different dates, verifying the effectiveness of introducing hydrological process constraints for gap filling.

[0106] The present invention has the following significant beneficial effects:

[0107] (1) This invention uses the complementary fusion of optical remote sensing images and synthetic aperture radar images at the probability level, and uses precipitation intensity and seasonal attributes as physical constraints to fill in the time series, increasing the effective pixel ratio of the study area from 27.04% of the original optical image to 99.71%, which is 41.51% higher than the traditional method. It achieves almost complete coverage of daily continuous observation, effectively solves the problem of discontinuity in long-term observation caused by clouds, rainfall and satellite orbits, and is suitable for cloudy and rainy areas and areas with rapid hydrological response, significantly improving the spatiotemporal integrity and data availability of surface water products.

[0108] (2) Traditional gap-filling methods often rely on statistical similarity or temporal proximity, lacking constraints on the surface water formation process. This invention uses precipitation intensity and flood / non-flood season attributes as grouping constraints, ensuring that the reference image and the image to be filled are in the same or similar hydrological and meteorological background, thus ensuring that the filling results physically conform to the response law of surface water to precipitation input. After hydrological consistency verification, the reconstructed water area and the measured water level have good consistency in long-term trends, seasonal fluctuations, and short-term residuals, with a Pearson correlation coefficient of 0.60. The seamless daily-scale surface water data obtained by the method of this invention can truly reflect the real hydrological process, with high hydrological consistency and physical rationality, and can provide a reliable data foundation for short-term drought and flood identification and water resource allocation assessment.

[0109] (3) This invention does not directly perform hard classification on images, but rather performs fusion and completion at the probability level throughout the process, from water body probability prediction, multi-source probability fusion, pixel-by-pixel probability filling within the constraint group to non-precipitation day probability interpolation. It outputs pixel-level water body probability instead of hard classification results, effectively preserving the gradual change information of water body boundaries and pixel-level uncertainty, avoiding boundary ambiguity and category confusion caused by binary hard classification, and improving spatial accuracy and stability under complex terrain conditions.

[0110] (4) This invention constructs a similarity index primarily based on structural similarity and secondarily on similarity components based on mean square error. When screening reference images, it can maintain the integrity of the spatial morphology of water bodies such as rivers and lakes, and avoid filling distortion caused by spectral value drift. Through the above comparative examples 2 and 3, this weighted combination outperforms the single-index scheme in all four indicators: overall accuracy, precision, recall, and F1 score, and exhibits the least seasonal fluctuation.

[0111] (5) By constructing a surface water replenishment method based on precipitation constraints, this invention realizes seamless daily-scale continuous production of surface water products in complex cloud and rain areas. This product can reliably capture the rapid rise and fall of surface water caused by heavy rainfall, short-term drought, irrigation and drainage and water conservancy scheduling, providing a key data foundation for early identification of short-term drought, flood response monitoring, reservoir, lake and pond scheduling assessment and agricultural water resources management.

[0112] Example 2

[0113] Based on the same inventive concept, this invention proposes a seamless daily-scale surface water recharge system based on precipitation constraints, comprising:

[0114] The data acquisition module is used to acquire optical remote sensing images, synthetic aperture radar images, and daily precipitation sequence data of the study area;

[0115] The probability prediction and fusion module is used to generate optical water probability maps and synthetic aperture radar water probability maps for each date based on optical remote sensing images and synthetic aperture radar images using a water probability prediction model. It also performs complementary fusion of the optical water probability map and synthetic aperture radar water probability map for the same date to obtain the corresponding fused water probability map.

[0116] The precipitation constraint grouping module is used to divide the fused water body probability map of precipitation dates into multiple constraint groups based on the precipitation intensity and seasonal attributes of daily precipitation sequence data.

[0117] The similarity imputation module is used to calculate the similarity between the merged water body probability map and other merged water body probability maps in the same constraint group in the effective common region for each merged water body probability map with missing pixels in each constraint group. It iterates through each missing pixel in the merged water body probability map and searches other merged water body probability maps in descending order of similarity. When other merged water body probability maps have a valid water body probability value at the corresponding pixel position, the missing pixel is filled with the valid water body probability value to obtain the reconstructed merged water body probability map.

[0118] The temporal interpolation module is used to perform linear interpolation on the reconstructed and merged water body probability map based on adjacent precipitation dates for non-precipitation dates, so as to obtain the water body probability map for non-precipitation days.

[0119] The data output module is used to perform threshold segmentation on the reconstructed and merged water body probability map and the non-precipitation day water body probability map to obtain seamless daily-scale surface water data.

[0120] The seamless daily-scale surface water refill system based on precipitation constraints provided in this embodiment of the invention has a similar implementation principle and technical effect to that of Embodiment 1, and will not be repeated here.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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 or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A seamless daily-scale surface water refill method based on precipitation constraints, characterized in that, include: Acquire optical remote sensing images, synthetic aperture radar images, and diurnal precipitation sequence data of the study area; Based on the optical remote sensing image and the synthetic aperture radar image, an optical water probability map and a synthetic aperture radar water probability map for each date are generated using a water body probability prediction model. The optical water probability map and the synthetic aperture radar water probability map for the same date are then fused together to obtain the corresponding fused water probability map. Based on the precipitation intensity and seasonal attributes of the daily precipitation sequence data, the fused water body probability map of precipitation dates is divided into multiple constraint groups; For each of the constraint groups containing missing pixels in the merged water body probability map, the similarity between the merged water body probability map and the other merged water body probability maps in the same constraint group is calculated in the effective common region. Each missing pixel in the merged water body probability map is traversed, and the other merged water body probability maps are searched sequentially according to the similarity from high to low. When the other merged water body probability map has a valid water body probability value at the corresponding pixel position, the missing pixel is filled with the valid water body probability value to obtain the reconstructed merged water body probability map. For non-precipitation dates, linear interpolation is performed on the reconstructed and merged water body probability map based on adjacent precipitation dates to obtain the water body probability map for non-precipitation days; Threshold segmentation is performed on the reconstructed and merged water body probability map and the water body probability map on non-precipitation days to obtain seamless daily-scale surface water data.

2. The seamless daily-scale surface water refill method based on precipitation constraints according to claim 1, characterized in that, The formula for calculating the similarity is: ； in, Indicates similarity; and Represents the weighting coefficient, and Greater than ; Represents the structural similarity index; The similarity component based on mean squared error is calculated as follows: , This represents the mean square error.

3. The seamless daily-scale surface water refill method based on precipitation constraints according to claim 1, characterized in that, The precipitation intensity is divided into seven levels based on 24-hour precipitation: no precipitation, light rain, moderate rain, heavy rain, rainstorm, heavy rainstorm, and extremely heavy rainstorm.

4. The seamless daily-scale surface water refill method based on precipitation constraints according to claim 1, characterized in that, The seasonal attributes include the flood season and the non-flood season.

5. The seamless daily-scale surface water refill method based on precipitation constraints according to claim 1, characterized in that, The optical remote sensing imagery includes surface reflectance data formed by fusing the Landsat series and Sentinel-2 series, with input bands including six bands: red, green, blue, near-infrared, shortwave infrared 1, and shortwave infrared 2. The synthetic aperture radar imagery includes terrain-corrected Sentinel-1 GRD data, with input features including two polarization channels: VV and VH.

6. The seamless daily-scale surface water refill method based on precipitation constraints according to claim 1, characterized in that, The water body probability prediction model adopts a dual-source U-Net model, which is used to receive the optical remote sensing image and the synthetic aperture radar image, respectively. Each of the U-Net models includes an encoder, a bottleneck layer, a decoder, and a convolutional layer; the encoder includes a multi-level convolutional module and a max-pooling layer; the decoder includes a transposed convolutional upsampling layer and skip connections connected to the corresponding layer of the encoder.

7. The seamless daily-scale surface water refill method based on precipitation constraints according to claim 1, characterized in that, The linear interpolation includes: Using the reconstructed and merged water body probability map of adjacent precipitation dates as endpoints, linear interpolation is performed on non-precipitation dates between endpoints according to date distance, and the interpolation results are constrained to a probability range of 0 to 1.

8. The seamless daily-scale surface water refill method based on precipitation constraints according to claim 1, characterized in that, Before performing linear interpolation on the reconstructed and merged water body probability map based on adjacent precipitation dates for non-precipitation dates to obtain the water body probability map for non-precipitation days, the method further includes: After the filling is completed, if there are still missing pixels in the reconstructed and merged water body probability map in any of the constraint groups, then based on the time position of the reconstructed and merged water body probability map with missing pixels, the reconstructed and merged water body probability maps with adjacent time are searched step by step forward and backward, and the missing pixels are retrieved and filled pixel by pixel according to the degree of temporal proximity.

9. The seamless daily-scale surface water refill method based on precipitation constraints according to claim 1, characterized in that, The threshold segmentation uses the bimodal method to determine the threshold, and the reconstructed and merged water body probability map and the non-precipitation day water body probability map are converted into a binary surface water map as seamless daily-scale surface water data.

10. A seamless daily-scale surface water recharge system based on precipitation constraints, characterized in that, include: The data acquisition module is used to acquire optical remote sensing images, synthetic aperture radar images, and daily precipitation sequence data of the study area; The probability prediction and fusion module is used to generate optical water probability maps and synthetic aperture radar water probability maps for each date based on the optical remote sensing image and the synthetic aperture radar image, using a water probability prediction model, and to perform complementary fusion of the optical water probability map and the synthetic aperture radar water probability map for the same date to obtain the corresponding fused water probability map. The precipitation constraint grouping module is used to divide the fused water body probability map of precipitation dates into multiple constraint groups based on the precipitation intensity and seasonal attributes of the daily precipitation sequence data. The similarity imputation module is used to calculate the similarity between the merged water body probability map and other merged water body probability maps in the same constraint group in the effective common region for each merged water body probability map with missing pixels in each constraint group; iterates through each missing pixel in the merged water body probability map, and sequentially searches the other merged water body probability maps according to the similarity from high to low; when the other merged water body probability map has a valid water body probability value at the corresponding pixel position, the missing pixel is filled with the valid water body probability value to obtain the reconstructed merged water body probability map; The temporal interpolation module is used to perform linear interpolation on the reconstructed and merged water body probability map based on adjacent precipitation dates for non-precipitation dates to obtain the water body probability map for non-precipitation days. The data output module is used to perform threshold segmentation on the reconstructed and merged water body probability map and the non-precipitation day water body probability map to obtain seamless daily-scale surface water data.

Images