Remote sensing image water body extraction method based on hydrological resource normalization monitoring

Abstract

The invention discloses a remote sensing image water body extraction method based on hydrological resource normalized monitoring, and relates to the technical field of water body monitoring extraction, and the method comprises the steps: carrying out the spatial resampling and time alignment of an optical remote sensing image, a radar remote sensing image and hydrological data, and carrying out the serialization of water body features extracted in a plurality of continuous monitoring periods, constructing a time stack, analyzing time stack data by using a time sequence prediction model, predicting a water body change trend in a next monitoring period, quantifying response coefficients of water body features in various dimensions, and performing time sequence prediction in the next monitoring period. And dynamically adjusting a spectral index calculation coefficient, a boundary detection threshold and fusion matrix weighting according to a predicted output response coefficient, and optimizing a water body feature extraction process. According to the method, the feature extraction of the water body is converted from a static fixed parameter into a dynamic self-adaptive strategy, and the rapid extraction and response of the water body features in a future period are realized, so that the monitoring efficiency is improved and the repeated calculation is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water body monitoring and extraction, and particularly relates to a remote sensing image water body extraction method based on hydrological resource normalization monitoring. BACKGROUND

[0002] In the process of normalizing monitoring of hydrological resources, long-term continuous observation of water body changes in a large range of watersheds or lakes is usually required to master the dynamic processes of seasonal fluctuations, sudden water rise and dry water of the water body. The commonly used automatic monitoring means is to perform image segmentation and classification processing on each periodically acquired remote sensing image to identify the current water body distribution and range. Most of the current automatic monitoring recognition methods mainly rely on post-processing comparison, that is, the difference analysis of two or more consecutive remote sensing images to determine the abnormal area. This is a post-detection method. Therefore, for monitoring periods with frequent weather changes, the recognition result of the water body change will lag behind the actual change process, which makes it difficult to reflect the water body dynamics in time and to realize the early identification and rapid response to water body abnormalities.

[0003] In addition, the water body extraction process of each monitoring period in the existing method is independent, and a fixed parameter configuration and processing flow are usually used. This independent execution method makes it difficult to adaptively optimize the attention to abnormal areas and the feature extraction range, affecting the continuous tracking ability of abnormal areas and the stability and accuracy of long-term monitoring results. SUMMARY

[0004] (I) Technical problems solved The present application provides a remote sensing image water body extraction method based on hydrological resource normalization monitoring, which can realize continuous tracking monitoring of abnormal water body areas and accurate extraction of water body features.

[0005] (II) Technical solutions To achieve the above purpose, the present application provides the following technical solutions: a remote sensing image water body extraction method based on hydrological resource normalization monitoring, comprising the following steps: Periodically acquiring multi-source remote sensing data and ground hydrological monitoring data in a selected hydrological area, the multi-source remote sensing data including optical remote sensing images and radar remote sensing images, and performing spatial resampling and time alignment on the multi-source remote sensing data in each monitoring period; In each of the monitoring periods, a reflectance spectral feature of the water body is extracted based on the optical remote sensing image, a spectral index is calculated according to a preset waveband combination coefficient and a normalized interval, a backscattering coefficient is extracted from the radar remote sensing image, and a mask map of the water body is generated through a feature fusion algorithm to determine a water body region; and a morphological feature of the water body is extracted based on the mask map. The water body features extracted in the continuous multiple monitoring periods are time sequenced to construct a time stack in a corresponding time period, and the time stack is used to reflect the evolution process of the spectral feature and the morphological feature. The water body feature sequence composed of dimensions in the time stack is taken as input, a time series prediction model is constructed, a water body change trend in a next monitoring period is predicted, and a response coefficient of each dimension water body feature is output, the response coefficient representing a contribution degree of the water body feature in the corresponding dimension in the water body change trend. In the next monitoring period, the response coefficient of each dimension water body feature output by the time series prediction model is used to adjust the extraction of the water body feature; wherein the response coefficient of the spectral feature dimension is used to adjust the waveband combination and the calculation weight of each waveband in the spectral index calculation, and the response coefficient of the morphological feature dimension is used to correct a scale parameter of the water body boundary extraction.

[0006] In an implementable embodiment, in each of the monitoring periods, the optical remote sensing image and the radar remote sensing image covering the selected hydrological region are subjected to spatial resampling and re-projection processing to align them under a unified spatial reference, and the aligned remote sensing images are represented in a unified rasterized pixel; the optical remote sensing image and the radar remote sensing image in the same monitoring period are subjected to time alignment processing according to a unified time reference to form a remote sensing data set with consistent acquisition time.

[0007] In an implementable embodiment, in each of the monitoring periods, the reflectance of each waveband is extracted from the optical remote sensing image after the spatial resampling and time alignment of the optical remote sensing image; then the reflectance of each waveband is subjected to spectral index calculation according to a preset waveband combination coefficient, and the calculation result is normalized according to a set interval to obtain a normalized index value reflecting the spectral reflectance characteristic for each pixel. In each monitoring period, the radar remote sensing image aligned with the optical remote sensing image in the unified rasterized pixel is used to calculate the backscattering coefficient of each pixel in the radar remote sensing image to obtain the radar water body response feature of each pixel; then the optical water body index and the radar response feature are fused according to a preset weighted fusion matrix to generate a mask map of the water body distribution for determining the spatial range of the water body; and the morphological feature of the water body is calculated based on the mask map according to a fixed boundary detection threshold and a geometric feature extraction scale parameter.

[0008] In an implementable embodiment, in each monitoring period, based on the generated mask map, the pixels marked as water bodies in the mask map are regionally identified, and the continuously connected pixels are aggregated into an independent water body unit; for the morphological feature identification of the water body, the following steps are performed: For each water body unit, the area of the water body is obtained by multiplying the number of pixels in the water body unit by the actual ground area corresponding to the pixel; The boundary pixels of each water body unit are traversed, and the distance between the boundary pixels and the adjacent water body units is accumulated to obtain the contour length of the water body; The center point coordinates of each water body unit are obtained by weighted averaging the pixel coordinates in the water body unit as the centroid position of the water body.

[0009] In an implementable embodiment, the data sets of a plurality of continuous monitoring periods are organized in time sequence to form a time stack of multi-source data; during the construction of the time stack, the data sets corresponding to each monitoring period are mapped onto the time axis in sequence with the monitoring period as the time dimension, and the correspondence between the pixels in different monitoring periods is maintained under a unified spatial grid.

[0010] In an implementable embodiment, when constructing the time series prediction model, based on the water body features extracted from the same water body object in the continuous monitoring periods in the time stack, a water body feature sequence arranged in time sequence is constructed, and the meteorological rainfall data corresponding to each monitoring period is aligned according to the same time reference and jointly organized with the water body feature sequence in the feature dimension to form multi-dimensional time series input data; the time series prediction model performs time series modeling by using the multi-dimensional time series input data to quantify the dynamic law of the evolution of water body features over time.

[0011] In an implementable embodiment, the constructed multi-dimensional time series input data is organized in time sequence, and a sliding time window is used to split the time series of the input data, so that each time window contains a fixed length of continuous historical water body features and corresponding external driving data to represent the time series evolution state of the water body features in adjacent monitoring periods.

[0012] In an implementable embodiment, in the next monitoring period, the response coefficients of the spectral features and the morphological features are calculated according to the prediction results of the water body feature sequences in each dimension in the current time stack by the time series prediction model; wherein, the output response coefficient of the spectral feature is mapped to the band combination coefficient and the normalization calculation interval of the optical remote sensing obtained in the monitoring period, and the band value of each pixel is weighted to adjust the contribution factor in the feature calculation formula in proportion to the response coefficient.

[0013] In a feasible embodiment, the response coefficients of the morphological features output by the time series prediction model are mapped to the threshold range of the water body boundary detection operator and the scale parameters of geometric feature extraction. The upper and lower bounds of the threshold range are linearly adjusted by the response coefficients, the pixel classification judgment conditions are adjusted accordingly, and the scale parameters of geometric feature extraction are corrected proportionally according to the response coefficients.

[0014] (III) Beneficial Effects: Compared with the prior art, this invention has the following beneficial effects: This invention models and predicts the changing trends of multidimensional water body characteristics through a time series prediction model. It can detect the possible direction and intensity of changes in water body spectrum, morphology and radar characteristics in advance before the monitoring period arrives, and form a forward-looking characteristic response assessment.

[0015] In the next monitoring cycle, the feature extraction of the water body reflectance spectrum is dynamically adjusted based on the predicted output response coefficient, enabling the optical feature extraction process to adapt to future changes in the optical response of the water body. Based on the predicted output response coefficient of morphological changes, the algorithm parameters for boundary and morphological detection of the water body area are corrected, improving the sensitivity to changes in water body morphology. This transforms the monitoring and feature extraction of the water body from static fixed parameters to a dynamic parameter adjustment mechanism based on the prediction results. In the new monitoring cycle, the focus and accuracy of water body feature extraction can be adaptively optimized, maintaining continuous perception and identification of changes in abnormal areas, and improving the stability and reliability of long-term monitoring results. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for extracting water bodies from remote sensing images based on routine monitoring of hydrological resources, provided in an embodiment of the present invention. Figure 2 The flowchart illustrates the process of extracting morphological features and reflectance spectral features of water bodies from multi-source data acquired within a monitoring cycle in a method for extracting water bodies from remote sensing images based on routine monitoring of hydrological resources, as provided in this embodiment of the invention. Figure 3 In a method for extracting water bodies from remote sensing images based on routine monitoring of hydrological resources provided in this embodiment of the invention, a flowchart is shown in which the reflectance spectrum and morphological feature sequence data of water bodies extracted from multiple consecutive monitoring cycles are used to construct a time stack corresponding to the time period. Figure 4 The flowchart of a method for extracting water bodies from remote sensing images based on routine monitoring of hydrological resources provided in this embodiment of the invention is shown below. It illustrates the process of adjusting the water body adjustment strategy for the next monitoring cycle based on the predicted output response coefficient from a time stack. Figure 5In the method for extracting water body in remote sensing image based on hydrological resource normalization monitoring provided by the embodiment of the present application, the water body mask diagram of the selected hydrological region is extracted. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0018] In addition, if the terms "first", "second", etc. are used only for differentiation and cannot be understood as indicating or implying relative importance.

[0019] It should be noted that the features in the embodiments of the present application can be combined with each other without conflict.

[0020] In the current hydrological monitoring technical system, remote sensing means has become the main way for water body identification and dynamic change analysis. Most methods rely on water body index extraction of optical images or scattering feature identification of radar images. Through classification and segmentation of each period of remote sensing image, the water body distribution range and area change are obtained to analyze the flood or drought trend.

[0021] However, the existing methods mostly perform independent analysis on single images, and fail to establish a stable time sequence relationship in the time dimension, so that the periodic law, trend evolution and abnormal fluctuation of water body change are difficult to identify. The monitoring results often show discreteness, which is difficult to support continuous tracking and dynamic early warning of hydrological processes.

[0022] In summary, in the time dimension of technical implementation, the existing technology stays in multiple independent detection, and does not realize continuous time sequence monitoring. In the information utilization of technical implementation, only single period data classification is relied on, and no prediction and feedback mechanism based on historical law is formed. Finally, in the extraction efficiency of technology, there is no quick update based on prediction results and key area priority extraction mechanism.

[0023] In order to solve the above problems, combined with the prior art as shown in the background of the present application, the embodiment of the present application provides a method for extracting water body in remote sensing image based on hydrological resource normalization monitoring. Figures 1 to 5 The method realizes prediction-driven dynamic monitoring and continuous feedback closed loop. The historical data, prediction model and new cycle fast feature extraction are closely linked, which can ensure the continuity of global monitoring, preferentially respond to key change areas, and provide reliable data support and decision basis for hydrological management, flood warning and water resource scheduling.

[0024] In the normalized hydrological monitoring system, a fixed time step is first set as a monitoring period to ensure uniformity in the time dimension and comparability of the monitoring results. It should be noted that the time step can be set according to the hydrological response characteristics of the target basin, the rate of meteorological change and the data acquisition frequency, and is usually several days to several weeks to balance timeliness and calculation stability.

[0025] First, S10: periodically acquiring multi-source remote sensing data and ground hydrological monitoring data in the selected hydrological region, the multi-source remote sensing data including optical remote sensing images and radar remote sensing images, and spatial resampling and time alignment are performed on the multi-source remote sensing data in each monitoring period. That is, in each monitoring period, the collection of multi-source data is performed on the selected hydrological region, including optical remote sensing images, radar remote sensing images and measured hydrological data.

[0026] Among them, the optical remote sensing image provides the information of the surface reflection characteristics, which is mainly used for the spectral difference identification of water body and non-water body, and the radar remote sensing image can obtain the surface scattering characteristics under the condition of cloud cover, which supplements the limitation of optical image under cloudy or night observation. It should be understood that since the above data has significant differences in spatial resolution, coordinate reference and sampling time, the multi-source data acquired at multiple time nodes in the same monitoring period need to be uniformly processed before entering the feature extraction step.

[0027] In the embodiment of the present application, a spatial resampling processing stage is first performed. Spatial resampling is a process of unifying the resolution differences of images acquired by different sensors. Through pixel-level interpolation methods such as bilinear interpolation or cubic convolution interpolation, low-resolution images are resampled to the same grid scale as the target resolution. At the same time, in this process, the coordinates of all data are converted to a unified geographic reference frame, such as WGS84 or UTM projection. After resampling, the spatial positions of various images and raster data are aligned, so that each pixel has a one-to-one correspondence between different data sources.

[0028] After completing the spatial resampling, the time alignment stage is entered. Since the acquisition times of multi-source data are not completely synchronized, there may be several hours to several days of sampling difference between data of different sensors in the same monitoring period. In order to ensure the consistency of the time dimension, alignment under a unified time reference is required. Time alignment maps each data source in the same period to the same time node by constructing a unified time reference step.

[0029] In some embodiments of the present application, for image data with inconsistent observation time (i.e. the above-mentioned optical remote sensing and radar remote sensing images), weighted average or time interpolation can be used for fusion. It can be understood here that the weighting coefficients are set according to the difference between the sampling time of each image and the central time. The closer the sampling time is to the central time, the higher the weight. For the measured hydrological data and meteorological data, the same time step weighted average method is also used to time extract the continuous observation records to make them synchronized under the same time reference as the image data.

[0030] For example, in a monitoring period of one week, if the optical remote sensing image of the selected watershed is obtained on the 2nd day of the period, the radar remote sensing image is obtained on the 4th day, and the measured hydrological data and rainfall data are daily scale continuous records, then the middle time point (3rd day) of the monitoring period is taken as the unified time reference. At this time, the weighted coefficients are calculated for the optical and radar images respectively And What is needed here is to perform time weighted average. Then continue to assume that the hydrological and meteorological data are weighted average values within the two days before and after, so as to form a time-consistent multi-source data set on the 3rd day time step. It can be understood that the data set processed in this way is completely aligned in spatial coordinates and time reference, and can be directly used for subsequent water body recognition and change analysis.

[0031] After sorting out the multi-source data in a monitoring period, S20 is performed: in each monitoring period, the reflectance spectral features of the water body are extracted based on the optical remote sensing image, the spectral index is calculated according to the preset waveband combination coefficient and the normalization interval, the backscattering coefficient is extracted from the radar remote sensing image, and the mask map of the water body is generated by the feature fusion algorithm to determine the water body area; the morphological features of the water body are extracted based on the mask map using boundary detection and geometric feature algorithm.

[0032] The execution process of the entire S20 flow can be divided into three main links: optical remote sensing image water body index calculation, radar remote sensing image feature extraction and feature fusion determination. Before explaining each link in detail, refer to the time period division of the monitoring period shown in Figures 2 to 4 It is assumed that the monitoring object is a seasonal lake located in a low latitude plain area. The area enters the wet season in summer (June-August), and the lake area changes significantly due to heavy rainfall and increased runoff. The monitoring period is set to one month, and the continuous three periods are: T1-T2: early summer (June), frequent rainfall, increased surface runoff; T2-T3: midsummer (July), lake expansion, surrounding vegetation flourishing; T3-T4: late summer (August), high evaporation, local contraction.

[0033] The three monitoring periods from T1 to T4 constitute a time stack in subsequent processing to predict the water body changes of T4-T5, and adjust the water body extraction strategy in the monitoring period of T4-T5 according to the prediction result, so as to achieve the extraction of the normal water body features in the previous period on the basis of the extraction of the normal water body features in the previous period, and realize the rapid feature extraction in the subsequent detection monitoring period.

[0034] The monitoring period (T3-T4) carried out at this stage, the process task of S20 step is to extract the real state of the current water body distribution, shape and spectral feature, as the input basis of time series analysis and next period prediction.

[0035] Specifically, for the water body index calculation of optical remote sensing image, it is used to represent the spectral reflectance characteristics of the pixel, which usually includes visible light and near-infrared waveband in optical remote sensing image, and the reflectance combination between different wavebands can effectively distinguish water body from other ground objects. Among them, in the water body pixel, the reflectivity of near-infrared waveband is low, and the reflectivity of blue or green waveband is relatively high. Based on this spectral difference, in the embodiment of the present application, the water body index value is calculated by performing spectral ratio operation on each pixel, that is, the calculation of water body index is realized by waveband ratio, difference or normalized ratio, and the calculation formula can be referred to as: ; Among them, is the water body index, is the reflectivity of green waveband, is the reflectivity of near-infrared waveband.

[0036] After calculation, each pixel in the generated water body index map is assigned a value, and the larger the value, the more likely the pixel is water body. This step converts the optical remote sensing image into a two-dimensional continuous index map reflecting the spectral characteristics of water body, providing basic input for subsequent judgment.

[0037] In the processing stage of radar remote sensing image, the structure and humidity characteristics of ground objects are identified by analyzing the backscattering coefficient and reflectivity. In the embodiment of the present application, considering that radar remote sensing uses microwave active emission signal for imaging, can penetrate cloud layer and is sensitive to surface roughness and water content, it can be complementary to optical image. At each pixel, the backscattering value (σ°) ) and reflectivity are calculated, which reflect the echo intensity of radar signal to the ground surface. In order to eliminate noise interference, local statistical processing is introduced in the calculation process, including but not limited to using window sliding mean, variance filtering or Lee filtering method to suppress speckle noise, and removing or smoothing the abnormal strong scattering pixels.

[0038] After processing, the radar image is transformed into a two-dimensional matrix with a stable scattering response, where each pixel represents the radar water body response characteristics of that surface unit. Generally, open water surfaces produce lower backscattering values ​​due to their smoothness and strong specular reflection, while land or vegetated areas have higher scattering values.

[0039] In the feature fusion stage, the optical water index and radar features are fused at the pixel level. It's important to understand that the core of this fusion is establishing a unified numerical space so that the two types of features, calculated above, jointly reflect the water body attributes under the same dimensions. Therefore, in this embodiment of the invention, the two types of features are weighted and integrated during the fusion stage to form a more stable water body identification result. The feature fusion function can be found as follows: ; in, A water index calculated from an optical remote sensing image of a water body region. This refers to the backscattering normalization value extracted from a radar remote sensing image within a water body area. and These are preset weighting coefficients, representing the trust ratio between optical and radar information. The selection of these coefficient values ​​is usually dynamically adjusted based on surface type, season, or sensor characteristics. No specific limitations are imposed here.

[0040] After the fusion calculation is completed, in some embodiments of the present invention, a set threshold is used. Mapping continuous values ​​to a binary mask, specifically when... When, it is marked as a water body (value 1); when When the value is 0, it is marked as a non-water body. In this way, the entire monitoring area is divided into two categories of pixels: water bodies and non-water bodies, generating a preliminary water body distribution mask map. (See reference below.) Figure 5 The white blocks represent water bodies, and the black blocks represent non-water bodies.

[0041] After feature fusion, to further determine the spatial morphology of the water body, boundary extraction processing needs to be performed on the binary mask. Specifically, boundary extraction involves identifying the spatial contour of the water body region. In some embodiments of this invention, a gradient operator (Canny algorithm) is used to detect the edge pixels of the mask. The detected edge pixels constitute the boundary lines of the water body region. Through vectorization processing, a boundary vector map can be generated, thereby clearly defining the contour, direction, and distribution pattern of the water body in space.

[0042] Still taking the above-mentioned current monitoring period as an example, assuming that the water body index calculated from the optical image is between 0.7 and 0.9, the pixel is regarded as a possible water body area, and the backscattering value in the radar image is lower than -18 dB, the pixel is displayed as a smooth reflection surface. After weighted fusion, the pixel area with a final comprehensive response value greater than a set threshold is determined to be a water body, and after Canny edge detection extraction, a continuous river boundary line can be obtained, which continues to refer to Figure 5 The boundary not only clearly defines the spatial range of the water body, but also provides a basic input for the calculation and dynamic change analysis of the subsequent morphological parameters.

[0043] After that, based on the generated water body distribution mask, the pixels marked as water bodies (value 1) are identified, and by performing connectivity analysis on these pixels, different water body units can be automatically distinguished. For example, in an area with a group of lakes or a dense water system, different lakes, river sections or depression water areas will be identified as independent water body units. This step usually uses a connectivity domain search algorithm, which can set 8-neighborhood or 4-neighborhood, and by traversing the surrounding pixels of each pixel and judging whether they belong to the same water body category, the complete water body area is gradually aggregated.

[0044] After identification, a series of morphological feature calculations are performed on each independent water body unit in turn. For the calculation of the water body area, according to the spatial resolution of the image, i.e. the actual ground area represented by each pixel, the number of pixels in the water body unit is counted, and the total area of the water body is calculated by multiplication. For the calculation of the perimeter of the water body, by traversing the pixels at the edge of the mask, the total length of the water body boundary is obtained by using the Euclidean distance between adjacent pixels. If the adjacent direction between pixels is horizontal or vertical, the edge distance is one unit of pixel resolution; if it is diagonal, the distance is times the pixel resolution. It can be understood that by this traversal accumulation method, a high-precision water body boundary length can be obtained, which provides a data basis for the subsequent calculation of the complexity of the shoreline.

[0045] In some embodiments, in addition to the morphological feature calculation, the centroid position of each water body unit is further extracted, which is specifically obtained by weighted averaging the geographic coordinates (latitude and longitude or projection coordinates) of all pixels in the water body unit, i.e. the center point position of the unit, and the calculation formula can be referred to as: ; wherein, is the total number of pixels in the water body unit, is the coordinates of each pixel in the water body area, and the obtained centroid can be used to represent the center of gravity of the overall spatial position of the water body.

[0046] In summary, it can be understood that by extracting and dynamically comparing these morphological parameters, we can not only quantitatively describe the evolution of water bodies over time, but also provide direct quantitative evidence for subsequent hydrological trend modeling, flood evolution analysis, or anomaly monitoring.

[0047] After acquiring multi-source data during the monitoring period, refer to Figure 3 As shown, S30 is performed: the water features extracted from multiple consecutive monitoring cycles are time-seriesd to construct a time stack structure within the corresponding time period. The time stack is used to reflect the evolution trajectory of spectral and morphological features.

[0048] In the process of hydrological monitoring, although data from a single point in time can reflect the characteristics of the water body at a certain moment, it cannot reflect the evolution pattern of the water body over time. Therefore, in the embodiments of the present invention, in order to achieve continuous observation and trend analysis of the dynamic changes of the water body, it is necessary to stack the multi-source data acquired in multiple monitoring cycles in chronological order to construct a time stack of multi-source data. In programming languages, this can be understood as constructing a #temporalstack#.

[0049] The significance of constructing a time stack lies in the fact that by organizing multi-source observation results at different times into a time series structure, spatial and temporal information can be preserved and correlated simultaneously, thereby providing data support for subsequent time series analysis, change detection, and trend prediction.

[0050] Specifically, in the embodiments of the present invention, firstly, after spatial resampling and temporal alignment are completed in each monitoring cycle, a set of multi-source datasets with consistent spatial resolution and temporal reference are obtained. As mentioned above, each set of datasets includes optical remote sensing images, radar remote sensing images, elevation and topographic data, meteorological and precipitation data, and measured hydrological data, etc. These data achieve pixel-level correspondence in space through resampling and temporal consistency in time through weighted averaging or interpolation correction.

[0051] Subsequently, these multi-source datasets, arranged chronologically, are stacked along the time dimension, maintaining a one-to-one correspondence between pixel locations in space. This ensures that each pixel in the time stack represents the continuous change trajectory of that spatial location over multiple monitoring periods. It's important to understand that constructing the time stack is equivalent to creating a time-series vector for each pixel or defined pixel region in a remote sensing image of a selected hydrological area. This time-series vector contains its multi-source feature parameters across different monitoring periods.

[0052] The structured data stack is constructed in this way, so that the subsequent constructed prediction model can directly calculate the time series change of the same spatial position in the analysis, such as calculating the pixel brightness change rate, vegetation coverage change rate, water body index change trend, etc. Then it can be understood that through this time continuity data organization form, seasonal water level fluctuation, sudden flood expansion, and long-term drought shrinkage and other dynamic processes can be effectively identified.

[0053] In addition to the above data application level of the constructed prediction model, the time stack not only improves the accuracy of hydrological change analysis, but also provides more rich sample dimensions for the training of the prediction model. For example, when predicting water body changes based on time series models (such as LSTM, GRU or one-dimensional convolution network), the time stack can be directly used as an input tensor, so that the model can automatically capture time-dependent features and spatial coupling relationships during the learning process. In addition, the time stack can also calculate the time correlation between pixels to construct dynamic water body boundary evolution maps or abnormal change heat maps.

[0054] Now assume that in the continuous monitoring of several months, the optical remote sensing and radar remote sensing images of each month are stacked to form a time stack, when analyzing a pixel point of a river, it can be found that the monthly decrease of reflectivity and the synchronous increase of radar scattering intensity correspond to the expansion trend of the water body. On the contrary, if the water body index gradually decreases and the backscattering coefficient increases in continuous periods, it indicates that the water body in this area is shrinking or drying out. In this way, the time stack not only realizes the unified expression of multi-source and multi-time data, but also provides a directly operable data basis for subsequent change detection, trend analysis and anomaly identification.

[0055] After constructing the time stack of multi-source data, S40 is performed: taking the water body feature sequence in the time stack as input, a time series prediction model is constructed, which is used to predict the water body change trend in the next monitoring period and the response coefficient of each dimension water body feature.

[0056] First, the sequence of water body features calculated in each monitoring period in the time stack and the corresponding period of meteorological rainfall data are uniformly processed and formatted. Through this formatting processing, the data of each period is kept consistent in the time dimension and the space dimension, so that each water body unit has complete historical parameter records and corresponding environmental conditions in the time series.

[0057] As for the architecture of the time series prediction model, some embodiments of the present application include, but are not limited to, using LSTM (to capture long and short term time dependencies) or a Transformer time series encoder (for adaptive learning of inter-feature weights). It is important to note here that the goal of this time series prediction model is not only to predict the future water area or the future water distribution, but also to predict the sensitivity of each feature dimension to the overall water body change in the future monitoring period (T4-T5). This means that the model needs to output two levels of results: one is the overall water change trend prediction, and the other is the response coefficient of each dimension.

[0058] Specifically, the time series prediction model first learns the time evolution rule of each dimension feature. Taking the above example, assume that the spectral feature NDWI rises significantly during T2-T3, and falls slightly during T3-T4; the radar backscattering fluctuates and increases during T3-T4; and the water area continues to grow during T1-T2 and T2-T3, and stabilizes during T3-T4. The model needs to learn these time-dependent information to reflect the seasonality of water expansion and the environmental response rhythm.

[0059] In the subsequent training phase, the time series prediction model further learns the mutual influence among spectral, radar, and morphological features through attention mechanism or correlation matrix calculation. For example, if the change trend of spectral and morphological features has high correlation, such as turbidity rising causing area extraction value to be too large, the model learns that the spectral feature has higher sensitivity to future area change. If the radar signal is decoupled from the morphological change (wind wave noise is large), the model reduces the sensitivity weight of the radar feature.

[0060] This cross-dimension relationship learning enables the model not only to predict future trends, but also to dynamically determine which type of feature dominates future changes.

[0061] In the prediction output layer of the time series prediction model, the response coefficient of each dimension is calculated by analyzing the contribution of the input feature to the future state. In some embodiments, it is usually obtained through attention weight, gradient attribution, or feature contribution rate statistics. Taking the above example, during the transition period from late summer to early autumn (T4-T5), due to changes in lighting conditions and weakened algal growth, the model predicts that the influence of spectral features on water boundary identification will weaken, so the output spectral feature response coefficient will be slightly lower, which means that the dynamic adjustment range of the normalization interval can be appropriately reduced when the band combination is performed in the next period.

[0062] Meanwhile, in the transition period from late summer to early autumn (T4-T5), due to the still high rainfall in late summer and the drastic change in surface roughness, the prediction model learns to judge the increasing stability of radar signal in identifying water bodies, so the output of the radar feature response coefficient is improved, which means that the weighting coefficient of the radar feature in the fusion matrix should be increased in the next period.

[0063] Finally, the output results of the time series model include the trend of water body change in the next period and the response coefficient of each water body feature. These two types of results together constitute the basis for adjusting the monitoring strategy in the next period. The trend prediction shows how the water body will change, while the response coefficient means which features are most critical to future changes, i.e. the response coefficient of each dimension represents the influence degree and direction of different water body features on the overall water quality change, which is used to quantify the driving effect of feature change on the comprehensive index of water body, so the calculation amount of subsequent feature recognition needs to be focused on the range involved by these response coefficients.

[0064] Finally, reference Figure 4 S50 is performed: in the next monitoring period, the response coefficients of each dimension of water body features output by the time series prediction model are used to adjust the extraction of water body features; among them, according to the response coefficient of spectral features, the calculation weight of band combination coefficient and normalization interval in optical remote sensing image is adjusted; according to the response coefficient of morphological features, the threshold range of water body boundary detection operator and the scale parameter of geometric feature extraction are corrected; in the process of optical and radar feature fusion, the response coefficient is introduced into the fusion function, and the feature fusion matrix is weighted and updated.

[0065] It should be noted here that after the start of the next monitoring period, the newly acquired multi-source remote sensing images also need to be preprocessed first. The specific process includes the above-mentioned spatial alignment of optical remote sensing images and radar remote sensing images, so that different sources and different resolution image data correspond to each pixel position under the unified geographical coordinates. At the same time, the spatial registration of the obtained elevation terrain data and meteorological rainfall data is also performed, so that all data establish a pixel-level correspondence relationship under the same spatial grid.

[0066] After spatial alignment, the above-mentioned predicted response coefficients are introduced into the water body extraction process of the current monitoring period, i.e. the feature extraction stage of the next monitoring period, and these response coefficients are fed back to the feature extraction process.

[0067] First, after obtaining the response coefficient, the feature extraction process is dynamically adjusted accordingly. For spectral feature dimension, the response coefficient is used to optimize the band combination method of optical remote sensing images, that is, in the calculation of multispectral index and normalization parameter, the wave band with high weight is improved in the combination coefficient, and the normalization interval is offset and scaled, so that the brightness response of sensitive wave band is more discriminative in the next cycle. In this way, the extraction of spectral features does not need to traverse all wave band combinations in the whole image, but only needs to perform weighted operation on high response coefficient wave band, reducing the amount of calculation. That is, by predicting which wave band and index is more critical in advance, the next cycle does not need to blindly scan all wave bands or iteratively adjust multiple times, but can quickly complete the generation of spectral index map by directly adjusting the calculation parameters.

[0068] In the morphological feature calculation, the threshold interval and geometric scale parameter of water body boundary detection operator are adaptively corrected according to the response coefficient output by the time series model. The morphological features (area, shoreline complexity) with high response coefficient indicate that the future water body morphological change amplitude is large, and the response interval of the boundary monitoring threshold is accordingly enlarged, and the boundary extraction scale is refined to improve the capture ability of subtle morphological changes.

[0069] In this way, only the predicted change area is detected with high precision, and the remaining area remains with the default parameters. In addition, the water body unit that may change can be locked in advance, and morphological feature data can be quickly generated. That is, the prediction response coefficient tells the algorithm to focus on which area and feature, and there is no need for full-image high-resolution processing, reducing the amount of calculation and achieving fast boundary extraction.

[0070] In the fusion process of optical and radar features, the response coefficient is further introduced into the fusion function to update the fusion matrix with weight. The weight value is used as an adjustment parameter of the fusion factor to determine the contribution proportion of different feature sources in the fusion space. Specifically, for the pixels with high response coefficient, the fusion weight is increased to make the mask more accurately cover the predicted change area. By predicting and locking the key area in advance, the fusion calculation does not need to process the whole image uniformly, thereby reducing the amount of calculation and processing time, making the mask more accurately cover the predicted change area, and generating the mask and water body boundary more efficiently.

[0071] Through the above process, the response coefficient is used to dynamically adjust the spectral, morphological and multi-source fusion levels, so that the feature extraction process in the next monitoring cycle no longer depends on fixed parameters, but is adaptively adjusted according to the prediction trend. Finally, without the need to rebuild the feature extraction model, the accurate extraction of water body features in the new cycle is quickly completed, the monitoring update efficiency and change response speed are improved, and the prediction-driven optimization mechanism of water body time series feature extraction is realized.

[0072] The above merely describes preferred embodiments of the present application, and is not intended to limit the present application. The scope of protection of the present application is defined by the claims, and equivalent changes made to the contents of the specification and drawings based on the claims are intended to be included in the scope of protection of the present application.

Claims

1. A method for extracting water bodies from remote sensing images based on routine hydrological resource monitoring, characterized in that, Includes the following steps: Periodically acquire multi-source remote sensing data and surface hydrological monitoring data within a selected hydrological area. The multi-source remote sensing data includes optical remote sensing images and radar remote sensing images. Spatial resampling and temporal alignment are performed on the multi-source remote sensing data within each monitoring period. Within each monitoring cycle, the reflectance spectral features of the water body are extracted based on the optical remote sensing image. The spectral index is calculated according to the preset band combination coefficient and normalization interval. Combined with the backscattering coefficient extracted from the radar remote sensing image, a mask map of the water body is generated through a feature fusion algorithm to determine the water body area. The morphological features of the water body are then extracted based on the mask map. The water features extracted from multiple consecutive monitoring periods are time-seriesd to construct a time stack for the corresponding time period. The time stack is used to reflect the evolution process of spectral and morphological features. Using the water body feature sequence composed of dimensions in the time stack as input, the water body change trend in the next monitoring period is predicted by the constructed time series prediction model, and the response coefficient of each dimension of water body feature is output. The response coefficient represents the contribution of the corresponding dimension of water body feature to the water body change trend. In the next monitoring cycle, the water body features are extracted by using the response coefficients of the water body features in each dimension output by the time series prediction model. Specifically, the band combination and the calculation weight of each band in the spectral index calculation are adjusted by using the response coefficients of the spectral feature dimension, and the scale parameters of the water body boundary extraction are corrected by using the response coefficients of the morphological feature dimension.

2. The method for extracting water bodies from remote sensing images based on routine hydrological resource monitoring according to claim 1, characterized in that, Within each monitoring cycle, the acquired optical and radar remote sensing images covering the selected hydrological area are spatially resampled and reprojected to align them under a unified spatial reference, and the aligned remote sensing images are represented by a unified rasterized pixel. Optical and radar remote sensing images from the same monitoring period are time-aligned according to a unified time reference to form a remote sensing dataset with consistent acquisition time reference.

3. The method for extracting water bodies from remote sensing images based on routine hydrological resource monitoring according to claim 2, characterized in that, Within each monitoring cycle, for the optical remote sensing image after spatial resampling and temporal alignment, the reflectance of each band is extracted from the optical remote sensing image. Then, according to the preset band combination coefficient, the spectral index of the reflectance corresponding to each band is calculated, and the calculation result is normalized according to the set interval so that each pixel obtains a normalized index value that reflects the spectral reflectance characteristics. Within each monitoring cycle, for the radar remote sensing image aligned with the optical remote sensing image in uniform rasterized pixels, the backscattering coefficient is calculated for each pixel in the radar remote sensing image to obtain the radar water body response characteristics of each pixel. The optical water index and radar response features are then fused according to a preset weighted fusion matrix to generate a mask map of water distribution, which is used to determine the spatial range of the water body. Based on the mask map, the morphological features of the water body are calculated according to a fixed boundary detection threshold and geometric feature extraction scale parameters.

4. The method for extracting water bodies from remote sensing images based on routine hydrological resource monitoring according to claim 3, characterized in that, Within each monitoring cycle, based on the generated mask image, regions are identified for pixels marked as water bodies within the mask image, and continuously connected pixels are aggregated into an independent water body unit; for the morphological feature identification of water bodies, the following steps are performed: For each water body unit, the area of ​​the water body is obtained by multiplying the number of pixels in the water body unit by the actual ground area corresponding to the pixel. Traverse the boundary cells of each water body unit and sum the distances between it and its adjacent water body units to obtain the outline length of the water body; By weighted averaging the pixel coordinates within a water body unit, the coordinates of the center point of each water body unit are obtained as the centroid position of that water body.

5. The method for extracting water bodies from remote sensing images based on routine hydrological resource monitoring according to claim 2, characterized in that, The datasets from multiple consecutive monitoring periods are organized in chronological order to form a time stack of multi-source data. During the construction of the time stack, the monitoring period is used as the time dimension, and the datasets corresponding to each monitoring period are mapped onto the time axis in sequence, while maintaining the correspondence between pixels in different monitoring periods under a unified spatial grid.

6. The method for extracting water bodies from remote sensing images based on routine hydrological resource monitoring according to claim 1, characterized in that, When constructing a time series prediction model, based on the water body features extracted from the same water body object within a continuous monitoring period in the time stack, a water body feature sequence arranged in chronological order is constructed. The meteorological and rainfall data corresponding to each monitoring period are aligned with the same time reference and jointly organized with the water body feature sequence in the feature dimension to form multi-dimensional time series input data. The time series prediction model performs time series modeling through the multi-dimensional time series input data to quantify the dynamic law of water body feature evolution over time.

7. The method for extracting water bodies from remote sensing images based on routine hydrological resource monitoring according to claim 6, characterized in that, The input data of the constructed multidimensional time series is organized in chronological order, and the time series of the input data is divided by a sliding time window method, so that each time window contains a fixed length of continuous historical water features and its corresponding external driving data, in order to characterize the temporal evolution of water features in adjacent monitoring periods.

8. The method for extracting water bodies from remote sensing images based on routine hydrological resource monitoring according to claim 1, characterized in that, In the next monitoring cycle, based on the prediction results of the water body feature sequences of each dimension in the current time stack by the time series prediction model, the response coefficients of spectral features and morphological features are calculated. Among them, the response coefficients of the output spectral features are mapped to the band combination coefficients and normalized calculation intervals of optical remote sensing obtained in this monitoring cycle. By performing weighted calculations on the band values ​​of each pixel, the contribution factor in the feature calculation formula is adjusted proportionally to the response coefficient.

9. The method for extracting water bodies from remote sensing images based on routine hydrological resource monitoring according to claim 8, characterized in that, The response coefficients of the morphological features output by the time series prediction model are mapped to the threshold range of the water body boundary detection operator and the scale parameters of geometric feature extraction. The upper and lower bounds of the threshold range are linearly adjusted by the response coefficients, the pixel classification judgment conditions are adjusted accordingly, and the scale parameters of geometric feature extraction are corrected proportionally according to the response coefficients.

Images