Water body region identification method and apparatus, storage medium, and computer program product

Abstract

Embodiments of the present disclosure disclose a water body region identification method and device, a storage medium and a computer program product. The method comprises: acquiring a first resolution remote sensing image at a previous time; performing image upscaling on the first resolution remote sensing image to generate a multi-scale image sequence; the multi-scale image sequence comprises a plurality of different scale images; performing ground object classification and segmentation to determine ground object classification results and ground object segmentation objects corresponding to the different scale images; based on the ground object classification results, the ground object segmentation objects, the first resolution remote sensing image at the previous time and a second resolution remote sensing image at the previous time, unmixing a second resolution remote sensing image at a subsequent time into initial remote sensing images corresponding to the different scale images; performing step-by-step incremental fusion of image residuals on the initial remote sensing images corresponding to the different scale images to obtain predicted remote sensing images corresponding to the different scale images at the subsequent time; and identifying a water body region from the predicted remote sensing images.

Description

Technical Field

[0001] This disclosure relates to the field of remote sensing image technology, specifically to a method, apparatus, storage medium, and computer program product for water body area identification. Background Technology

[0002] Against the backdrop of global climate change, extreme weather events are becoming increasingly frequent, and floods have become a major threat to human life and property. Timely and accurate monitoring of floods is crucial for developing effective disaster response strategies and mitigating losses. However, current ground-based monitoring using remote sensing satellites faces a technical bottleneck: high temporal resolution and high spatial resolution cannot be achieved simultaneously. This predicament makes it difficult to identify flood boundaries and flood flow trends with high frequency and precision in flood monitoring, severely restricting accurate disaster assessment and the scientific deployment of emergency measures. Existing flood monitoring methods have significant shortcomings in resolution and accuracy, making it difficult to meet the complex and ever-changing needs of practical applications. Summary of the Invention

[0003] This disclosure provides a method, apparatus, storage medium, and computer program product for identifying water areas.

[0004] In a first aspect, this disclosure provides a method for identifying water areas, comprising:

[0005] Acquire first-resolution remote sensing images from preceding time points;

[0006] The first-resolution remote sensing image is upscaled to generate a multi-scale image sequence; the multi-scale image sequence includes multiple images at different scales.

[0007] Perform land cover classification and segmentation on the multi-scale image sequence to determine the land cover classification results and land cover segmentation objects corresponding to images at different scales;

[0008] Based on the land cover classification results, land cover segmentation objects, first-resolution remote sensing images from previous time periods and second-resolution remote sensing images from previous time periods, the second-resolution remote sensing images from subsequent time periods are demixed to form the initial remote sensing images corresponding to the different scale images from subsequent time periods.

[0009] The initial remote sensing images corresponding to the images at different scales are fused stepwise with increasing image residuals to obtain the predicted remote sensing images corresponding to the images at different scales at subsequent time steps.

[0010] Identify water areas from the predicted remote sensing images.

[0011] Further, the multi-scale image sequence is subjected to land cover classification and segmentation to determine the land cover classification results and land cover segmentation objects corresponding to images at different scales, including:

[0012] Pixel clustering is performed on the images at different scales respectively;

[0013] The normalized water index is used to constrain the clustering results to obtain the land cover classification results of the images at different scales;

[0014] Image segmentation is performed on the images at different scales to obtain the land feature segmentation objects of the images at different scales;

[0015] For the images at different scales, identify the water bodies within the land feature segmentation objects;

[0016] The water body object is pixel-corrected using a pre-built dynamic flow direction model.

[0017] Furthermore, pixel correction is performed on the water body object using a pre-built dynamic flow direction model, including:

[0018] The dynamic flow direction model is used to determine the water flow direction of each pixel within the water object;

[0019] The boundary of the water body object is corrected based on the water flow direction.

[0020] Further, based on the land cover classification results corresponding to the images at different scales, the land cover segmentation objects, the first-resolution remote sensing image from the preceding time step, and the second-resolution remote sensing image from the preceding time step, the second-resolution remote sensing image from the subsequent time step is demixed to form the initial remote sensing image corresponding to the images at different scales in the subsequent time step, including:

[0021] Using the second-resolution remote sensing images from the preceding and subsequent time periods, the type change index within a pixel is calculated.

[0022] By utilizing the intra-pixel type change index and the land cover segmentation objects corresponding to the images at different scales, the intra-image type change index of land cover segmentation objects of different land cover types in the images at different scales is optimized.

[0023] Based on the in-image type change index and the multi-scale image sequence, the initial remote sensing images corresponding to the images at different scales at the subsequent time steps are predicted.

[0024] Further, the initial remote sensing images corresponding to the images at different scales are fused stepwise with increasing image residuals to obtain predicted remote sensing images corresponding to images at different scales at subsequent time steps, including:

[0025] The multi-scale image sequence is traversed in ascending order of scale, and the object-level residual of each object segmentation in the residual fusion image of the initial remote sensing image and the lower-level scale image corresponding to the current traversed scale image is calculated; wherein, the residual fusion image of the lower-level scale image of the smallest scale image in the multi-scale image sequence is the second resolution remote sensing image of the previous time step.

[0026] The object-level residual is fused to the initial remote sensing image corresponding to the current traversal scale image to obtain the residual fused image corresponding to the current traversal scale image;

[0027] Calculate the pixel-level residuals of the pixels in the residual fused image corresponding to the current traversed scale image and the residual fused image of the lower-level scale image;

[0028] The pixel-level residuals are fused into the residual fused image corresponding to the image at the current traversal scale.

[0029] After the traversal is completed, the residual fused image corresponding to the largest scale image in the multi-scale image sequence is determined as the predicted remote sensing image.

[0030] Further, the object-level residuals of each land cover segmentation object in the residual fusion image of the initial remote sensing image and the lower-level scale image corresponding to the current traversal scale image are calculated, including:

[0031] Calculate the land cover type consistency weight between pixels in the image at the current traversal scale and other pixels within their respective land cover segmentation objects;

[0032] Using the land cover type consistency weight and the first pixel residual of the pixels within the land cover segmentation object, the object-level residual of the land cover segmentation object in the current traversal scale image is calculated; wherein, the first pixel residual is calculated based on the residual fusion image of the initial remote sensing image and the lower-level scale image corresponding to the current traversal scale image.

[0033] Further, the pixel-level residuals of pixels in the residual fused image corresponding to the current traversed scale image and the residual fused image of the lower-level scale image are calculated, including:

[0034] The contribution weight of the similar pixels is calculated using the spatial distance between the target pixel and similar pixels in the image at the current traversal scale; the similar pixels are pixels selected by setting conditions within a low-resolution window centered on the target pixel.

[0035] Using the contribution weight and the second pixel residual of the target pixel, the pixel-level residual of the target pixel in the current traversal scale image is calculated; the second pixel residual is calculated based on the residual fusion image corresponding to the current traversal scale image and the residual fusion image of the lower-level scale image.

[0036] Secondly, this disclosure provides a water area identification device, comprising:

[0037] The acquisition module is configured to acquire the first-resolution remote sensing imagery from the preceding time step.

[0038] The generation module is configured to upscale the first-resolution remote sensing image to generate a multi-scale image sequence; the multi-scale image sequence includes multiple images at different scales.

[0039] The determination module is configured to perform land cover classification and segmentation on the multi-scale image sequence, and determine the land cover classification results and land cover segmentation objects corresponding to images at different scales;

[0040] The demixing module is configured to demix the second-resolution remote sensing image of the subsequent time step into the initial remote sensing image corresponding to the images of different scales at the subsequent time step based on the land cover classification results, land cover segmentation objects, the first-resolution remote sensing image of the preceding time step and the second-resolution remote sensing image of the preceding time step.

[0041] The fusion module is configured to perform incremental fusion of image residuals on the initial remote sensing images corresponding to the images at different scales to obtain predicted remote sensing images corresponding to images at different scales at subsequent time steps.

[0042] The identification module is configured to identify water areas from the predicted remote sensing image.

[0043] The function can be implemented by hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above function.

[0044] In one possible design, the above-described device includes a memory and a processor. The memory stores one or more computer instructions that support the device in performing the corresponding methods described above, and the processor is configured to execute the computer instructions stored in the memory. The device may also include a communication interface for communicating with other devices or communication networks.

[0045] Thirdly, embodiments of this disclosure provide an electronic device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described in any of the above aspects.

[0046] Fourthly, embodiments of this disclosure provide a computer-readable storage medium for storing computer instructions used by any of the above-described devices, which, when executed by a processor, are used to implement the methods described in any of the above aspects.

[0047] Fifthly, embodiments of this disclosure provide a computer program product comprising computer instructions which, when executed by a processor, are used to implement the methods described in any of the preceding aspects.

[0048] The technical solutions provided in this disclosure can include the following beneficial effects:

[0049] This embodiment achieves efficient water body segmentation from images by combining water body characteristics and clustering methods. This scheme only requires combining high- and low-resolution data from preceding time steps to achieve high-resolution extraction of flood water bodies, thus ensuring the accuracy of flood boundaries. Unlike traditional methods, this embodiment effectively handles boundary blurring between objects during boundary extraction, improving the accuracy of water body extraction and reducing the impact of boundary blurring or image fusion errors.

[0050] This embodiment addresses the boundary ambiguity issue in transitional regions through object-oriented segmentation technology while maintaining high spatiotemporal resolution, meeting the demand for refined boundaries in flood monitoring. By deeply fusing water body characteristics and image data, this method can provide high-precision flood boundary information even with complex terrain boundaries, thus offering more reliable data support for emergency response and post-disaster assessment.

[0051] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0052] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments, taken in conjunction with the accompanying drawings. In the drawings:

[0053] Figure 1 A flowchart of a water area identification method according to an embodiment of the present disclosure is shown.

[0054] Figure 2 A flowchart illustrating the process of water area identification according to an embodiment of the present disclosure is shown.

[0055] Figure 3A The image of a lake A according to an embodiment of this disclosure is shown as a result of the land feature classification.

[0056] Figure 3B It shows Figure 3A The image shown shows the ground feature segmentation results.

[0057] Figure 4 This diagram illustrates how the D-finity algorithm calculates the flow direction.

[0058] Figures 5A-5D A schematic diagram showing experimental data effects according to an embodiment of the present disclosure is provided.

[0059] Figures 6A-6D Showing the Figures 5A-5D The image shown is a schematic diagram of the effect after extracting water objects using NDWI.

[0060] Figures 7A-7B This illustrates an embodiment of the present disclosure. Figures 5A-5D The image shown is a schematic diagram of the effect obtained after multi-scale solution and the effect after water extraction and correction.

[0061] Figure 8 A structural block diagram of a water area identification device according to an embodiment of the present disclosure is shown.

[0062] Figure 9 This is a schematic diagram of the structure of an electronic device suitable for implementing a water area identification method according to an embodiment of the present disclosure. Detailed Implementation

[0063] In the following, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings to enable those skilled in the art to readily implement them. Furthermore, for clarity, portions unrelated to the description of the exemplary embodiments have been omitted from the drawings.

[0064] In this disclosure, it should be understood that terms such as “comprising” or “having” are intended to indicate the presence of features, figures, steps, behaviors, components, parts or combinations thereof disclosed in this specification, and do not preclude the possibility of the presence or addition of one or more other features, figures, steps, behaviors, components, parts or combinations thereof.

[0065] It should also be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0066] The existing technology has the following problems:

[0067] 1. Traditional satellite imagery is limited by the revisit cycle, and high-resolution imagery (such as MODIS) has insufficient spatial resolution, making it difficult to identify flood boundaries.

[0068] 2. Image interpretation methods based on a single sensor are susceptible to cloud interference, which leads to a decrease in the reliability of monitoring results.

[0069] 3. Existing super-resolution reconstruction technology is prone to artifacts in complex terrain areas, affecting the accuracy of water boundary extraction.

[0070] 4. Conventional image classification methods do not fully consider hydrological and geographical features, resulting in a high misclassification rate of inundated areas.

[0071] This disclosure aims to overcome the limitations of existing technologies and provide a water area identification method based on image classification and super-resolution technology. This method is applied to the identification and extraction of flood areas to solve the problem of flood area identification caused by resolution and observation time limitations. It enables real-time dynamic monitoring and accurate extraction of floods, providing accurate data support for disaster assessment and emergency response.

[0072] The method disclosed herein, through object-level processing, preserves the structure of image data and performs incremental fusion at multiple scales, effectively helping the model improve its detail recovery capability during residual fusion. To address the issue of inaccurate boundary extraction during traditional high- and low-resolution image fusion for resolution enhancement, this disclosure introduces a D-finity dynamic flow direction model based on a digital elevation model (DEM) to assist in the identification and reconstruction of floodwater areas, significantly improving the accuracy of flood area extraction. This method facilitates real-time monitoring and accurate extraction of floodwaters during flood disasters, providing more accurate data support for disaster assessment and emergency response.

[0073] The details of the embodiments of this disclosure are described in detail below through specific examples.

[0074] Figure 1 A flowchart illustrating a water area identification method according to an embodiment of this disclosure is shown. Figure 1 As shown, the water body area identification method includes the following steps:

[0075] In step S101, the first resolution remote sensing image of the preceding time step is acquired;

[0076] In step S102, the first resolution remote sensing image is upscaled to generate a multi-scale image sequence; the multi-scale image sequence includes multiple images at different scales.

[0077] In step S103, the multi-scale image sequence is classified and segmented to determine the classification results and segmentation objects of the land features corresponding to images at different scales.

[0078] In step S104, based on the land cover classification results, land cover segmentation objects, the first resolution remote sensing image of the preceding time and the second resolution remote sensing image of the preceding time, the second resolution remote sensing image of the subsequent time is demixed into the initial remote sensing image corresponding to the different scale images of the subsequent time.

[0079] In step S105, the initial remote sensing images corresponding to the images at different scales are fused stepwise with increasing image residuals to obtain the predicted remote sensing images corresponding to the images at different scales at subsequent time steps.

[0080] In step S106, water areas are identified from the predicted remote sensing image.

[0081] In this embodiment, the first-resolution remote sensing image at the preceding time step (T0) can be a high-resolution remote sensing image, which is existing data. It can be predicted based on the second-resolution remote sensing image at time step T0-1, i.e., the low-resolution remote sensing image, or it can be acquired by remote sensing satellites monitoring the ground. Both the second-resolution remote sensing image at the preceding time step and the second-resolution remote sensing image at the following time step (T1) can be obtained by remote sensing satellite monitoring. The second-resolution remote sensing image is a low-resolution remote sensing image, meaning that the resolution of the first-resolution remote sensing image is higher than that of the second-resolution remote sensing image. Through the above method of this disclosure, it is not necessary to simultaneously detect high temporal resolution and high spatial resolution remote sensing satellite monitoring. Instead, the high-resolution remote sensing image at the following time step can be predicted using the low-resolution and high-resolution remote sensing images at the preceding time step and the low-resolution remote sensing image at the following time step. Then, water areas can be identified from the predicted high-resolution remote sensing images for flood area monitoring.

[0082] In some embodiments, a multi-scale image sequence generated using a first-resolution remote sensing image may include multiple images of different scales, the scale of which decreases progressively from the first-resolution remote sensing image. The specific number of images of different scales and the magnitude of scale reduction may be determined based on the actual application, and this disclosure does not impose any limitations on this.

[0083] Figure 2 The diagram illustrates a flowchart of a water area identification process according to an embodiment of the present disclosure. The diagram uses an example of a generated multi-scale image sequence including 100*100, 50*50, and 20*20, where 100*100 can be a first-resolution remote sensing image at time T0 (i.e., the preceding time).

[0084] In some embodiments, land cover classification and segmentation are performed on each scale image in a multi-scale image sequence to determine the land cover segmentation and classification objects in each image. Land cover segmentation objects include water bodies and non-water bodies; non-water bodies include objects such as buildings, roads, grasslands, and mountains.

[0085] In some embodiments, the initial predicted remote sensing image for a subsequent time period is a high-resolution remote sensing image predicted based on the temporal differences between the high-resolution image (i.e., the second-resolution remote sensing image) of the preceding time period and the low-resolution remote sensing image of the subsequent time period in terms of the object type of the ground feature segmentation.

[0086] Unmixing the second-resolution remote sensing image at subsequent time steps can be understood as follows: because the low-resolution remote sensing image contains mixed pixels representing different types of land cover segments, unmixing is necessary to ensure that the pixels in the predicted high-resolution remote sensing image at subsequent time steps accurately reflect these segmented land cover types. In this unmixing process, the pixel differences between the preceding and subsequent second-resolution remote sensing images are determined. Based on these pixel differences and the types of land cover segments, a change index for different land cover segmentation types in the image is determined. Then, based on this change index and the second-resolution remote sensing image at subsequent time steps, a first-resolution remote sensing image at subsequent time steps is predicted. For ease of description, the predicted high-resolution remote sensing image at subsequent time steps is referred to as the initial remote sensing image. This unmixing process predicts the corresponding initial remote sensing image for each scale image in the multi-scale image sequence.

[0087] To ensure that the high-resolution remote sensing images obtained in subsequent time steps can be further refined, this embodiment also performs incremental fusion of image residuals on the unmixed initial predicted remote sensing images to obtain the final predicted remote sensing images. This predicted remote sensing image serves as the first-resolution remote sensing image for subsequent time steps, from which water areas can be identified; it can also be used to predict high-resolution remote sensing images for a further subsequent time step (i.e., time T2).

[0088] This embodiment significantly improves the extraction accuracy of boundaries between different land cover classification and segmentation types by introducing a multi-scale fusion technique based on object classification. Furthermore, a multi-scale unmixing and fusion strategy is employed to effectively reduce unmixing errors caused by excessive differences in image resolution. Through multi-scale processing, information can be fully fused at different resolutions, mitigating the impact of resolution differences on the unmixing results and thus improving the model's performance in extracting complex land cover boundaries.

[0089] In an optional implementation of this embodiment, step S103, which is the step of performing land cover classification and segmentation on the multi-scale image sequence to determine the land cover classification results and land cover segmentation objects corresponding to images at different scales, further includes the following steps:

[0090] Pixel clustering is performed on the images at different scales respectively;

[0091] The normalized water index is used to constrain the clustering results to obtain the land cover classification results of the images at different scales;

[0092] Image segmentation is performed on the images at different scales to obtain the land feature segmentation objects of the images at different scales;

[0093] For the images at different scales, identify the water bodies within the land feature segmentation objects;

[0094] The water body object is pixel-corrected using a pre-built dynamic flow direction model.

[0095] In this optional implementation, the purpose of image upscaling is to construct a multi-scale analysis space through Laplacian pyramid decomposition of the image, providing a resolution-adaptive representation basis for multi-level feature fusion.

[0096] In some embodiments, the Laplacian pyramid algorithm upscales the first-resolution remote sensing image from a preceding time step as follows:

[0097] First, the first-resolution remote sensing image from the preceding time step is downsampled using a Gaussian kernel to generate a multi-scale image sequence. Second, the high-frequency detail features of each scale layer are progressively gradientd using the Laplacian difference operator. Compared to the traditional pyramid algorithm, this embodiment introduces a spatial continuity constraint term, which allows the decomposition process to maintain the structural integrity of geographic objects while reducing resolution, providing refined geographic semantic consistency feature primitives for subsequent data fusion.

[0098] For each scale image in the multi-scale image sequence obtained by upscaling, this embodiment performs the following classification and segmentation:

[0099] To address the issues of local feature loss, edge blurring, and high-frequency detail degradation in traditional image super-resolution methods at the pixel level (especially when dealing with regions with complex textures and varied structures, where a globally uniform weight allocation strategy is difficult to adapt to the differentiated needs of different types of regions), this disclosure proposes a multimodal segmentation correction framework guided by land cover classification. This framework solves the core problem of detail reconstruction distortion through object-level feature decoupling, thereby achieving high-precision reconstruction of flood areas. The process mainly consists of three stages: land cover classification, object segmentation, and water body object correction.

[0100] The following section introduces the stages of land cover classification.

[0101] First, based on multi-scale image sequences, a spectral-texture joint feature space (K-means) is constructed. An improved K-means algorithm is then used to cluster land cover features in images at different scales within the multi-scale image sequences. Since this disclosure focuses on water bodies, to effectively improve water body identification accuracy, a Normalized Difference Water Index (NDWI) is introduced to constrain the cluster center initialization process on top of the traditional K-means algorithm. By minimizing intra-cluster differences, remote sensing image pixels are clustered into different land cover types, thereby obtaining a refined distribution of land cover types. Figure 3A The image shows the feature classification results of a lake A according to an embodiment of this disclosure. This feature classification process can obtain fine classification results for different feature types such as water bodies, buildings, and roads; that is, the process can obtain the feature type of each pixel in images at different scales.

[0102] The following describes the object segmentation stage.

[0103] This embodiment employs semantically guided image instance segmentation: It utilizes a pre-trained SAM (Segment Anything Model) to segment images at different scales within a multi-scale image sequence, dynamically adjusting segmentation boundaries to extract multiple land cover objects from images at different scales. This stage yields the land cover object to which each pixel belongs in the images at different scales, and the information for each land cover object. Figure 3B It shows Figure 3A The image shown corresponds to the land cover segmentation results. These segmented land cover objects represent different land cover types within the region. Each segmented land cover object may contain multiple land cover types obtained from various land cover classification stages, such as partial water areas or partial vegetation areas. This method helps to reduce fragmented patches at the water-land interface.

[0104] The following describes the calibration phase for water bodies.

[0105] Considering that the main purpose of this disclosure is to identify and extract water bodies, in order to further improve the accuracy of water body object extraction, DEM data is introduced to construct a D-finity dynamic flow direction model to achieve spectral-topographic joint correction. Then, area threshold and connected domain analysis are performed on the corrected water body object, and morphological-topological constraint optimization is carried out to ensure that the object topology conforms to the water flow law.

[0106] In an optional implementation of this embodiment, the step of performing pixel correction on the water object using a pre-built dynamic flow direction model further includes the following steps:

[0107] The dynamic flow direction model is used to determine the water flow direction of each pixel within the water object;

[0108] The boundary of the water body object is corrected based on the water flow direction.

[0109] In this embodiment, the correction stage for the water body specifically includes:

[0110] First, the NDWI of each scale image in the multi-scale image sequence is calculated, and the calculated NDWI value is compared with a threshold to filter possible water bodies in the land cover segmentation results of each scale image.

[0111] Secondly, the D-finity algorithm is used to construct a flow direction model based on DEM data, and further refinements are made for the area where the water body is located, taking into account geographical features. For example... Figure 4 The diagram illustrates the D-finity algorithm's method for calculating flow direction. The D-finity algorithm analyzes the steepest descent direction of each pixel on the DEM (Digital Elevation Model), and then constructs a flow direction model of the region based on this descent direction and topographical features of the water flow. This flow direction model can simulate the direction of water flow and the distribution of the watershed, automatically identifying upstream and downstream areas of a water body (such as...). Figure 4 As shown, the D-finity algorithm divides the 3×3 neighborhood into eight triangular faces and selects the steepest slope direction as the flow direction.

[0112] Based on this flow direction model, topographic correction is performed on water bodies initially selected by NDWI. This correction process includes: for objects marked as non-water bodies, if there is water flowing into the adjacent pixels of the water body object's boundary pixels (i.e., adjacent pixels located outside the water body object's boundary pixels), the land cover object segmentation type of that adjacent pixel is corrected to water body, thus classifying it as part of the water body object and expanding its boundary. This combined spectral and topographic information correction effectively eliminates misclassifications caused by topographic shadows or high-humidity soil. In this embodiment, with the assistance of the flow direction model, when an upstream area is identified as a water body, the model automatically classifies the corresponding downstream area as a water body area to correct the segmentation boundary. Simultaneously, the use of DEM data helps to more accurately correct boundaries in areas with complex topography, avoiding errors in remote sensing imagery from affecting the accurate extraction of water body boundaries.

[0113] Finally, morphological and topological optimizations were performed on the water bodies after spectral-topography correction. First, excessively small water bodies were filtered out based on area thresholds (to remove noise specks), ensuring the extracted results primarily consisted of meaningful water areas. Then, connectivity analysis was performed to fill in missing water areas within the water bodies (determined based on land cover classification results), ensuring spatial continuity. Finally, morphological operations (such as dilation, erosion, and boundary smoothing) were applied to optimize the water body shape, and topological constraints were used to ensure the spatial structure of the water bodies (such as branch connectivity) conformed to the connectivity characteristics of real water systems and the consistency of flow direction.

[0114] Through the above three stages, this disclosure achieves boundary correction for water bodies in remote sensing images. In areas with complex topography, joint spectral-topographic correction can more accurately correct water body boundaries, avoiding extraction biases caused by topographic undulations or remote sensing image errors. Morphological-topology constraint optimization ensures that the shape and connectivity of water bodies conform to natural water flow patterns, thereby improving the accuracy and reliability of water body boundary extraction. This method exhibits high accuracy and robustness in practical hydrogeographic applications and can be used to support the precise extraction and analysis of information on rivers, lakes, and other water bodies.

[0115] In an optional implementation of this embodiment, step S104, which involves demixing the second-resolution remote sensing image of the subsequent time period into the initial remote sensing image corresponding to the different scale images based on the land cover classification results, land cover segmentation objects, the first-resolution remote sensing image of the preceding time period, and the second-resolution remote sensing image of the preceding time period, further includes the following steps:

[0116] Using the second-resolution remote sensing images from the preceding and subsequent time periods, the type change index within a pixel is calculated.

[0117] By utilizing the intra-pixel type change index and the land cover segmentation objects corresponding to the images at different scales, the intra-image type change index of land cover segmentation objects of different land cover types in the images at different scales is optimized.

[0118] Based on the in-image type change index and the multi-scale image sequence, the initial remote sensing images corresponding to the images at different scales at the subsequent time steps are predicted.

[0119] In this embodiment, to address the issue of blurred change detection caused by mixed pixels in low-resolution remote sensing images, a spatiotemporally coupled object demixing model is established:

[0120] First, decoupling of changes: using the land cover classification results and land cover segmentation objects of the multi-scale image sequence obtained from the high-resolution remote sensing image at time T0, and the low-resolution remote sensing image data at times T0 and T1, the low-resolution remote sensing image at time T1 is demixed into the high-resolution remote sensing image at time T1.

[0121] ΔL=L T1 -L T0

[0122] Among them, L T0 For the low-resolution remote sensing image at the preceding time T0 (i.e., the second-resolution remote sensing image), L T1 ΔL represents the low-resolution remote sensing image at time T1, and ΔL represents the temporal difference between the low-resolution remote sensing images from T0 to T1.

[0123] According to spatial decomposition theory, assuming that land cover changes are uniform within the region corresponding to low-resolution pixels, the temporal variation of low-resolution pixels can be modeled as a linear combination of the temporal variations of high-resolution pixels for each land cover type within the corresponding region, i.e.:

[0124]

[0125] Where, ΔL xy ΔL represents the intra-pixel type change index of the land cover type within the region corresponding to a low-resolution pixel over time. xy The pixel value at coordinates (x, y) in ΔL; c represents the land cover type, f cxy ΔF represents the proportion of land cover type c within the region corresponding to the low-resolution pixel (this proportion can be determined based on the land cover types of each high-resolution pixel within the region corresponding to the low-resolution pixel, and the land cover types of each high-resolution pixel are obtained from the land cover classification results of the multi-scale image sequence). C n represents the in-image type change index of the land cover type c in the first-resolution remote sensing image at time T0 over time. c This represents the number of land cover types c. By combining the land cover classification results and land cover segmentation objects of pixels in the multi-scale image sequence in step 103, as well as the temporal difference ΔL between the preceding and following time series, the intra-image type change index of each land cover type in the entire image can be calculated, that is, the different degrees of change of different land cover types over time.

[0126] Secondly, in order to obtain the optimal ΔF C The solution can be calculated using the Adaptive Dynamic Momentum Optimization Algorithm (ADMA algorithm) in this embodiment.

[0127] The adaptive dynamic momentum optimization solution process is as follows:

[0128] The ADMA method is a momentum optimization-based approach that introduces a weighted sum of historical gradients during gradient descent to find local minima more quickly, avoiding the limitation of relying on the current gradient at every step. ADMA combines the advantages of momentum optimization with an adaptive dynamic adjustment mechanism, further improving convergence speed and stability.

[0129] This embodiment uses the ADMA algorithm to solve ΔF. C , to represent the in-image type change of the land cover type c in the first-resolution remote sensing image of the entire preceding time step;

[0130] The algorithm takes as input the land cover classification and segmentation results of multi-scale image sequences and the temporal difference between the second-resolution remote sensing images at time T0 and T1.

[0131] The momentum acceleration process of this algorithm incorporates a momentum term into the model update, making the optimization process smoother and accelerating convergence. During the calculation, ΔF... C For the parameters to be optimized, we first construct a loss function (objective function) to measure the current ΔF. C Is the solution optimal?

[0132]

[0133] Object space constraints: Land feature segmentation objects from multi-scale image sequences are added as regularization terms to the loss function to constrain model performance. These include spatial consistency regularization and spatial smoothness regularization. The loss function after adding regularization is:

[0134]

[0135] in and To customize the regularization parameters, C1 is the residual value after correcting the change value by weighting the distance of each pixel in each object segmentation object to the object center point, thereby reducing the degree to which the change value of each object boundary is affected by the change of surrounding objects. This is because the closer the area is to the object boundary, the more easily it is disturbed by the change of other objects. In order to keep the change within the same object area consistent, this spatial consistency regularization is added; C2 is the square of the gradient of the change value, which is used to smooth the changes in the image to avoid unnatural mottles in the changed image.

[0136] To minimize this loss function, for time step t, given ΔF at this moment... C gradient g t :

[0137]

[0138] Calculate the parameter ΔF for this time. C First moment m t (i.e., gradient g) t (mean) and second moment v t (i.e., gradient g) t The uncentered variance is used to represent momentum.

[0139] m t =β1·m t-1 +(1-β1)·g t

[0140]

[0141] β1 and β2 are hyperparameters, which are used to obtain corrected moment values ​​to update the parameters.

[0142] Adaptive learning rate: The ADAM algorithm adaptively adjusts the learning rate of each parameter by taking into account the first and second moments of the gradient, which helps to converge faster.

[0143]

[0144] Where α is the learning rate, set to 0.001 here, and ∈ is a minimum value added to prevent division by zero; for example, it could be 10. -8 Thus, the updated parameter ΔF is calculated. C .

[0145] Finally, high-resolution time series prediction is performed:

[0146] ΔF obtained through solution C Predictions are made based on high-resolution remote sensing images at time T1.

[0147] H T1 =H T0 +ΔF

[0148] Wherein, ΔF is the indices of in-image type change for each land cover type. C The calculated land cover type change index, H, for the entire image. T0 For the first resolution remote sensing image at the preceding time T0, H T1 This is the first-resolution remote sensing image for the predicted subsequent time step T1, referred to here as the initial remote sensing image. It should be noted that ΔF... C It is obtained by solving each image at each scale in the multi-scale image sequence separately. Therefore, for each scale image, the initial remote sensing image H is predicted accordingly. T1 .

[0149] In an optional implementation of this embodiment, step S105, which involves progressively increasing the fusion of image residuals from the initial remote sensing images corresponding to the images at different scales to obtain predicted remote sensing images corresponding to images at different scales at subsequent time steps, further includes the following steps:

[0150] The multi-scale image sequence is traversed in ascending order of scale, and the object-level residual of each object segmentation in the residual fusion image of the initial remote sensing image and the lower-level scale image corresponding to the current traversed scale image is calculated; wherein, the residual fusion image of the lower-level scale image of the smallest scale image in the multi-scale image sequence is the second resolution remote sensing image of the previous time step.

[0151] The object-level residual is fused to the initial remote sensing image corresponding to the current traversal scale image to obtain the residual fused image corresponding to the current traversal scale image;

[0152] Calculate the pixel-level residuals of the pixels in the residual fused image corresponding to the current traversed scale image and the residual fused image of the lower-level scale image;

[0153] The pixel-level residuals are fused into the residual fused image corresponding to the image at the current traversal scale.

[0154] After the traversal is completed, the residual fused image corresponding to the largest scale image in the multi-scale image sequence is determined as the predicted remote sensing image.

[0155] In this optional embodiment, the initial predicted images of image objects at different scales in the multi-scale image sequence are sequentially subjected to residual fusion, from object-based fusion to pixel-based fusion, achieving layer-by-layer fusion and correction to obtain the predicted remote sensing images at subsequent time steps. This ensures that uneven image residual fusion will not occur due to excessive differences in image scale, and ensures that the predicted remote sensing images at subsequent time steps are gradually refined. Figure 2 Taking the multi-scale image sequence shown as an example, first, the 20*20 image is traversed. Using the second-resolution remote sensing image at time T1, the object-level residual and pixel-level residual fusion is performed on the initial prediction image corresponding to the 20*20 image to obtain the residual fused image corresponding to the 20*20 image. Then, the 50*50 image is traversed. After the scale of the residual fused image corresponding to the 20*20 image is increased to 50*50, it is used to perform object-level residual and pixel-level residual fusion on the initial remote sensing image corresponding to the 50*50 image to obtain the residual fused image corresponding to the 50*50 image. Then, the 100*100 image is traversed, and so on, until all scale images are traversed. The obtained residual fused image is the final predicted high-resolution remote sensing image at time T1.

[0156] Regarding residual fusion based on land feature segmentation objects, we can start from the perspective of the numerical values ​​of the entire land feature segmentation object, select the principal components of the land feature segmentation object to calculate the residual change values, so as to represent the changes that have occurred in the entire land feature segmentation object, and superimpose them on the initial remote sensing image of the previous stage to improve the prediction accuracy of each land feature segmentation object.

[0157] The calculation of object-level residuals of various object segments in the residual fusion image of the initial remote sensing image and the lower-level scale image corresponding to the current traversal scale image includes:

[0158] Calculate the land cover type consistency weight between pixels in the image at the current traversal scale and other pixels within their respective land cover segmentation objects;

[0159] Using the land cover type consistency weight and the first pixel residual of the pixels within the land cover segmentation object, the object-level residual of the land cover segmentation object in the current traversal scale image is calculated; wherein, the first pixel residual is calculated based on the residual fusion image of the initial remote sensing image and the lower-level scale image corresponding to the current traversal scale image.

[0160] First, for each high-resolution pixel in the currently traversed scale image sequence, calculate the normalized spatial distance DC between its center point and the center point of its corresponding low-resolution pixel:

[0161]

[0162] Where x and y are the pixel coordinates in the image at the current traversal scale, x c y c This represents the center point coordinates within the low-resolution pixel window corresponding to pixel (x, y), and s is the ratio of the low resolution corresponding to the low-resolution window to the pixel resolution in the image at the current traversal scale. This low-resolution window depends on the low-resolution remote sensing image, which can be the second-resolution remote sensing image at time T0 or T1.

[0163] Object homogeneity index (OHI) for each high-resolution pixel:

[0164]

[0165] Where M represents the number of high-resolution pixels within a low-resolution pixel window, where the high-resolution pixels are the pixels in the image at the current traversal scale, and I k This represents the attributes of each high-resolution pixel. When it belongs to the same land cover type as its corresponding low-resolution pixel, then I... k =1, otherwise I k=0, this low-resolution pixel window depends on the low-resolution remote sensing image, which can be the second-resolution remote sensing image at time T0 or T1. The OHI (Object Type Inconsistency) is calculated with the high-resolution pixel as the center and the low-resolution pixel as the window size. It represents the land cover type consistency between a high-resolution pixel and its surrounding pixels; the larger the value, the higher the land cover type consistency within the pixel.

[0166] Theoretically, the more homogeneous the land cover within a low-resolution pixel, the higher the similarity between the high-resolution residual and the actual residual of the object segmented in that location. Therefore, by combining spatial distance DC and OHI, the object residual index (ORI) of each high-resolution pixel is calculated. The higher the residual index, the more representative the residual of that high-resolution pixel is of the actual residual.

[0167] ORI = OHI / DC

[0168] The land cover type consistency weight W of each high-resolution pixel within its local land cover segmentation object is calculated. k The overall residuals of each feature segmentation object are calculated:

[0169]

[0170] Where N represents the number of high-resolution pixels within the same feature segmentation object.

[0171]

[0172] Among them, R K Let represent the first pixel residual of each high-resolution pixel, where the prediction image residual ΔH is obtained by subtracting the initial prediction image obtained in the first stage from the low-resolution image at subsequent time steps. K This is used to recover the residuals within each feature segmentation object in this stage:

[0173] ΔH K =L T1 -H T1

[0174]

[0175] Among them, R K Indicates ΔH K The first pixel residual of each high-resolution pixel in the middle, H T1 These are the initial predicted images corresponding to images at different scales, L T1 These are low-resolution images from later time steps; L corresponds to images at different scales. T1 Unlike the smallest scale image, L T1Given the second-resolution remote sensing image at time T1, if the scale of the input second-resolution remote sensing image at time T1 differs from the minimum scale, it needs to be converted to the minimum scale before calculating the residuals; for images not at the minimum scale, L T1 The residual fused image is the image at a lower scale than the smallest scale image. For example, if the smallest scale image is 20*20, then its corresponding L... T1 Let L be the second-resolution remote sensing image at time T1. If the second-resolution remote sensing image at time T1 is a 10*10 scale image, then it is upscaled to 20*20 and the difference is taken with the 20*20 scale image in the multi-scale image sequence to obtain 20*20. For a non-minimum scale image of 50*50, its corresponding L... T1 This is the residual fused image corresponding to a 20*20 scale image. This residual fused image is the image obtained after object residual fusion and pixel residual fusion, which will be described in detail below.

[0176] Based on the above formula, the principal residual value R for each land feature segmentation object can be calculated. o This is combined with the initial predicted imagery calculated in the first stage to improve object-level accuracy.

[0177] H T1-O =H T1 +R o

[0178] Among them, H T1-O This represents the residual fused image corresponding to the current scale image at the corrected subsequent time step.

[0179] The following section introduces pixel-level residual fusion.

[0180] In previous stages, it was assumed that the land cover types within the segmented objects remained unchanged. However, considering that the primary task is to restore flood resolution, there must be a prerequisite that the land cover type has changed to water. Therefore, pixel-level residual calculations are performed on the images to recover pixels in each segmented object where the land cover type has changed. The pixel-level residuals of each pixel are obtained by weighted summation of the residuals of its neighboring similar pixels.

[0181] Furthermore, according to Tobler's first geographical law, similar pixels that are spatially closer to the currently targeted pixel contribute more to the combined residual than similar pixels that are farther away from the currently targeted pixel.

[0182] Therefore, calculating the pixel-level residuals of pixels in the residual fused image corresponding to the current traversal scale image and the residual fused image of the lower-level scale image includes:

[0183] The contribution weight of the similar pixels is calculated using the spatial distance between the target pixel and similar pixels in the image at the current traversal scale; the similar pixels are pixels selected by setting conditions within a low-resolution window centered on the target pixel.

[0184] Using the contribution weight and the second pixel residual of the target pixel, the pixel-level residual of the target pixel in the current traversal scale image is calculated; the second pixel residual is calculated based on the residual fusion image corresponding to the current traversal scale image and the residual fusion image of the lower-level scale image.

[0185] To identify similar pixels, this disclosure primarily relies on the following criteria: Within a low-resolution window centered on the high-resolution pixel (i.e., the target pixel), pixels whose spectral differences satisfy a set condition (threshold condition) are defined as similar pixels. This low-resolution window depends on a low-resolution remote sensing image, which can be a second-resolution remote sensing image at time T0 or T1. Next, by calculating the spatial distance between the target pixel and its neighboring pixels, the influence weight of these similar pixels on the target pixel residual can be further quantified. This weight calculation considers the spatial distance from the similar pixel to the center pixel; the closer the distance, the greater the contribution weight of the similar pixel.

[0186]

[0187] Among them, X S Y S Let X and Y represent the coordinates of similar pixels in the currently traversed images of the multi-scale image sequence, and let D represent the coordinates of the target pixel in the currently traversed images of the multi-scale image sequence. S It represents the spatial distance from similar pixels to the target pixel.

[0188]

[0189] The contribution weight W of similar pixels is calculated using the above formula. S This method uses weighted fusion to reconstruct details in images caused by variations in land cover types, thereby improving the accuracy and quality of high-resolution images. N represents the number of high-resolution pixels within a low-resolution pixel window. Here, high-resolution pixels refer to pixels in the image at the current traversal scale, and the low-resolution pixel window depends on the low-resolution remote sensing image, which can be a second-resolution remote sensing image at time T0 or T1.

[0190] By analyzing the residual fused image obtained in the second stage and the low-resolution image L from subsequent time steps... T1 By subtracting, we obtain the image residual ΔH during the current traversal process. S This is used to recover the residuals within each feature segmentation object in that stage.

[0191] ΔH S =L T1 -H T1-O

[0192]

[0193] Among them, R I R represents the pixel-level residual calculated in this stage. s Indicates that in ΔH S The second pixel residual is used to calculate the final prediction result H of the high-resolution image at time T1 for that image scale using the following formula. T1-F :

[0194] H T1-F =H T1-O +R I

[0195] For ease of description, H T1-F This is called the residual fused image corresponding to the current traversal scale image. In the next traversal scale image, this H... T1-F This is the residual fusion image of an image at a lower scale.

[0196] In this multi-level residual fusion process, image sizes are gradually fused from low to high scales, allowing for more accurate calculation of the residuals between low-resolution and high-resolution pixels, especially pixel-level residuals. By progressively increasing the fusion scale from smallest to largest, residual calculation is first performed on the larger low-resolution image, followed by the gradual introduction of high-resolution detail information, resulting in more accurate restoration of high-resolution pixels. Furthermore, this multi-scale, progressive merging method fully utilizes spatial information at different scales, optimizing residuals at each level to more effectively restore image details and structure.

[0197] This method not only improves image resolution but also enhances the robustness and prediction accuracy of the model, especially in regions with complex spatial structures and variations.

[0198] This disclosure combines the above methods, firstly by calculating the NDWI of the image to determine the water body region, and then by combining the DEM flow direction model to correct the water body region, thereby accurately identifying and extracting the water body region.

[0199] To verify the accuracy and effectiveness of the high-resolution flood extraction method based on multi-scale super-resolution of classification and segmentation combined with flow direction model, the inventors conducted method experiments, and the experimental data used are as follows: Figures 5A-5DThe image shows Landsat (high-resolution data, 30M) and MODIS (low-resolution data, 2KM) data for a lake on April 25, 2020, and August 10, 2020. Water areas were extracted from the data based on the Normalized Difference Water Index (NDWI) formula, yielding high- and low-resolution water area extraction results for the different time points. Figures 6A-6D As shown in the figure. The high-resolution predicted image obtained after classifying, segmenting, and multi-scale solving the experimental data shown in Figure 5 using the method of this disclosure is as follows: Figure 7A As shown, the results obtained after water extraction and correction are as follows: Figure 7B As shown.

[0200] This disclosure combines remote sensing imagery for land cover classification, segmentation, super-resolution reconstruction, and digital elevation model (DEM) analysis of water flow direction. This enables high-resolution reconstruction of flood areas to more closely approximate reality. Compared to traditional methods that separately process image enhancement or hydrological analysis, this approach achieves a deep integration of computer vision technology and traditional physical models. This cross-disciplinary fusion not only improves the accuracy of flood area identification but also enhances the reliability of dynamic water flow analysis.

[0201] This disclosure also employs an object-level multi-scale fusion mechanism to jointly optimize data at different resolutions and their corresponding object information. This method improves image resolution while ensuring the coherence and accuracy of information, avoiding the distortion problems common in resolution enhancement processes. Furthermore, the ADMA algorithm is used to calculate various types of change indices in the image, effectively improving the accuracy of transformation parameters and thus ensuring high precision and stability of the analysis results.

[0202] The following are embodiments of the apparatus disclosed herein, which can be used to execute embodiments of the method disclosed herein.

[0203] Figure 8 A structural block diagram of a water area identification device according to an embodiment of the present disclosure is shown. This device can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Figure 8 As shown, the water area identification device includes:

[0204] The acquisition module 801 is configured to acquire the first-resolution remote sensing image of the preceding time step.

[0205] The generation module 802 is configured to upscale the first resolution remote sensing image to generate a multi-scale image sequence; the multi-scale image sequence includes multiple images of different scales.

[0206] The determination module 803 is configured to perform land cover classification and segmentation on the multi-scale image sequence, and determine the land cover classification results and land cover segmentation objects corresponding to images at different scales.

[0207] The demixing module 804 is configured to demix the second-resolution remote sensing image of the subsequent time into the initial remote sensing image corresponding to the images of different scales based on the land cover classification results, land cover segmentation objects, the first-resolution remote sensing image of the preceding time and the second-resolution remote sensing image of the preceding time.

[0208] The fusion module 805 is configured to perform incremental fusion of image residuals on the initial remote sensing images corresponding to the images at different scales to obtain predicted remote sensing images corresponding to images at different scales at subsequent time steps.

[0209] The identification module 806 is configured to identify water areas from the predicted remote sensing image.

[0210] The water area identification device in this embodiment corresponds to the water area identification method described above. For details, please refer to the description of the water area identification method above, which will not be repeated here.

[0211] Figure 9 This is a schematic diagram of the structure of an electronic device suitable for implementing a water area identification method according to an embodiment of the present disclosure.

[0212] like Figure 9 As shown, the electronic device 900 includes a processing unit 901, which can be implemented as a CPU, GPU, FPGA, NPU, or other processing unit. The processing unit 901 can execute various processes according to any of the methods described above in this disclosure, based on a program stored in the read-only memory (ROM) 902 or a program loaded from the storage portion 908 into the random access memory (RAM) 903. The RAM 903 also stores various programs and data required for the operation of the electronic device 900. The processing unit 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0213] The following components are connected to I / O interface 905: an input section 906 including a keyboard, mouse, etc.; an output section 907 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to I / O interface 905 as needed. A removable medium 911, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 910 as needed so that computer programs read from it can be installed into storage section 908 as needed.

[0214] In particular, according to embodiments of this disclosure, any of the methods described above in the embodiments of this disclosure can be implemented as a computer software program. For example, embodiments of this disclosure include a computer program product comprising a computer program tangibly embodied on a machine-readable medium, the computer program containing program code for performing any of the methods in the embodiments of this disclosure. In such an embodiment, the computer program can be downloaded and installed from a network via communication section 909, and / or installed from removable medium 911.

[0215] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0216] The units or modules described in the embodiments of this disclosure can be implemented in software or hardware. The described units or modules can also be located in a processor, and the names of these units or modules do not necessarily constitute a limitation on the unit or module itself.

[0217] In another aspect, this disclosure also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the apparatus described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores one or more programs that are used by one or more processors to perform the methods described in this disclosure.

[0218] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

Claims

1. A method for identifying water body areas, wherein, include: Acquire first-resolution remote sensing images from preceding time points; The first-resolution remote sensing image is upscaled to generate a multi-scale image sequence. The multi-scale image sequence includes multiple images at different scales; For each scale of the images, pixel clustering is performed separately; The normalized water index is used to constrain the clustering results to obtain the land cover classification results corresponding to the images at different scales; Image segmentation is performed on each scale image in the images at different scales to obtain the land feature segmentation objects corresponding to the images at different scales; For the images at different scales, identify the water bodies within the land feature segmentation objects; The water flow direction of each pixel within the water object is determined using a pre-built dynamic flow direction model; Based on the water flow direction, the boundary of the water body object is corrected; Based on the land cover classification results, land cover segmentation objects, first-resolution remote sensing images from previous time periods and second-resolution remote sensing images from previous time periods, the second-resolution remote sensing images from subsequent time periods are demixed to form the initial remote sensing images corresponding to the different scale images from subsequent time periods. The initial remote sensing images corresponding to the images at different scales are fused stepwise with image residuals to obtain the predicted remote sensing images corresponding to the images at different scales at subsequent time steps. In this process, the initial predicted images of the images at different scales are fused layerwise from object-based to pixel-based residuals to achieve layer-by-layer fusion and correction, thereby obtaining the predicted remote sensing images. Identify water areas from the predicted remote sensing images.

2. The method according to claim 1, wherein, Based on the land cover classification results, land cover segmentation objects, the first-resolution remote sensing image from the preceding time step, and the second-resolution remote sensing image from the preceding time step, the second-resolution remote sensing image from the subsequent time step is demixed to form the initial remote sensing image corresponding to the different scale images from the subsequent time step, including: Using the second-resolution remote sensing images from the preceding and subsequent time periods, the type change index within a pixel is calculated. By utilizing the intra-pixel type change index and the land cover segmentation objects corresponding to the images at different scales, the intra-image type change index of land cover segmentation objects of different land cover types in the images at different scales is optimized. Based on the in-image type change index and the multi-scale image sequence, the initial remote sensing images corresponding to the images at different scales at the subsequent time steps are predicted.

3. The method according to any one of claims 1-2, wherein, The initial remote sensing images corresponding to the images at different scales are fused stepwise with increasing image residuals to obtain predicted remote sensing images corresponding to images at different scales at subsequent time steps, including: The multi-scale image sequence is traversed in ascending order of scale, and the object-level residual of each object segmentation in the residual fusion image of the initial remote sensing image and the lower-level scale image corresponding to the current traversed scale image is calculated; wherein, the residual fusion image of the lower-level scale image of the smallest scale image in the multi-scale image sequence is the second resolution remote sensing image of the previous time step. The object-level residual is fused to the initial remote sensing image corresponding to the current traversal scale image to obtain the residual fused image corresponding to the current traversal scale image; Calculate the pixel-level residuals of the pixels in the residual fused image corresponding to the current traversed scale image and the residual fused image of the lower-level scale image; The pixel-level residuals are fused into the residual fused image corresponding to the image at the current traversal scale. After the traversal is completed, the residual fused image corresponding to the largest scale image in the multi-scale image sequence is determined as the predicted remote sensing image.

4. The method according to claim 3, wherein, Calculate the object-level residuals of each land cover segmentation object in the residual fusion image of the initial remote sensing image and the lower-level scale image corresponding to the current traversal scale image, including: Calculate the land cover type consistency weight between pixels in the image at the current traversal scale and other pixels within their respective land cover segmentation objects; Using the land cover type consistency weight and the first pixel residual of the pixels within the land cover segmentation object, the object-level residual of the land cover segmentation object in the current traversal scale image is calculated; wherein, the first pixel residual is calculated based on the residual fusion image of the initial remote sensing image and the lower-level scale image corresponding to the current traversal scale image.

5. The method according to claim 3, wherein, Calculating the pixel-level residuals of pixels in the residual fused image corresponding to the current traversed scale image and the residual fused image of the lower-level scale image includes: The contribution weight of the similar pixels is calculated using the spatial distance between the target pixel and similar pixels in the image at the current traversal scale; the similar pixels are pixels selected by setting conditions within a low-resolution window centered on the target pixel. Using the contribution weight and the second pixel residual of the target pixel, the pixel-level residual of the target pixel in the current traversal scale image is calculated; the second pixel residual is calculated based on the residual fusion image corresponding to the current traversal scale image and the residual fusion image of the lower-level scale image.

6. A water area identification device, wherein, include: The acquisition module is configured to acquire the first-resolution remote sensing imagery from the preceding time step. The generation module is configured to upscale the first-resolution remote sensing image to generate a multi-scale image sequence; the multi-scale image sequence includes multiple images at different scales. The determination module is configured to perform pixel clustering for each scale image in the different scale images; The clustering results are constrained using a normalized water index to obtain the land cover classification results corresponding to the images at different scales; image segmentation is performed on each scale image to obtain the land cover segmentation objects corresponding to the images at different scales; water bodies are identified among the land cover segmentation objects for the images at different scales; the water flow direction of each pixel within the water body object is determined using a pre-built dynamic flow direction model; and the boundaries of the water body objects are corrected based on the water flow direction. The demixing module is configured to demix the second-resolution remote sensing image of the subsequent time step into the initial remote sensing image corresponding to the images of different scales at the subsequent time step based on the land cover classification results, land cover segmentation objects, the first-resolution remote sensing image of the preceding time step and the second-resolution remote sensing image of the preceding time step. The fusion module is configured to perform incremental image residual fusion on the initial remote sensing images corresponding to the images at different scales to obtain the predicted remote sensing images corresponding to the images at different scales at subsequent time steps; wherein, the initial predicted images of the images at different scales are sequentially fused from object-based to pixel-based residual fusion to achieve layer-by-layer fusion and correction to obtain the predicted remote sensing images. The identification module is configured to identify water areas from the predicted remote sensing image.

7. A computer-readable storage medium having computer instructions stored thereon, wherein, When executed by a processor, the computer instructions implement the method described in any one of claims 1-5.

8. A computer program product comprising computer instructions, wherein, When executed by a processor, the computer instructions implement the method described in any one of claims 1-5.

Images