Method and system for simulating runoff driven by soil freeze-thaw state in cold region permafrost basin

Abstract

The application provides a runoff simulation method and system driven by a soil freeze-thaw state of a permafrost basin in a cold region, and relates to the technical field of permafrost freeze-thaw runoff simulation. In view of the problems that the spatial division of the permafrost response type is highly subjective, the quantitative relationship between the air temperature and the infiltration is lacked for each division, and the gradual change process of the infiltration in the freeze-thaw transition period is not described, the application obtains multi-source remote sensing images and permafrost samples, trains a deep learning semantic segmentation model to generate a permafrost response type division map, determines the starting and stable air temperature thresholds of each division according to historical air temperature and runoff data, establishes the air temperature-infiltration capacity relationship curve unique to each division, discretizes the basin into calculation units, calculates the surface runoff according to the real-time air temperature and the corresponding curve, and draws a thermal map. The application improves the objectivity of permafrost division identification and the simulation accuracy of runoff generation in the freeze-thaw transition period, and reduces the subjectivity of artificial interpretation and the uncertainty of runoff prediction.

Description

Technical Field

[0001] This invention relates to the field of frozen soil freeze-thaw runoff simulation technology, and in particular to a method and system for runoff simulation driven by the freeze-thaw state of soil in cold-region frozen soil watersheds. Background Technology

[0002] Hydrological processes in cold-region watersheds are jointly controlled by snowmelt and soil freeze-thaw cycles. Accurate simulation of runoff changes under permafrost conditions is of significant scientific importance and practical value for water resource management, spring flood disaster early warning, and climate change impact assessment in cold regions. With the cross-integration of high-resolution remote sensing, deep learning, and distributed hydrological modeling technologies, intelligently identifying the spatial heterogeneity of the permafrost underlying surface based on multi-source remote sensing images and driving runoff simulation that considers freeze-thaw conditions has become a cutting-edge development direction for improving the accuracy of hydrological forecasts in cold regions.

[0003] However, existing technologies still have the following shortcomings when simulating runoff in permafrost watersheds in cold regions: First, the spatial classification of permafrost response types relies on manual interpretation or traditional classification methods, which are inefficient and highly subjective, making it difficult to objectively capture the spatial differences in freezing characteristics caused by factors such as soil texture, organic matter content, and ice content. Second, existing models usually assume that there are uniform permafrost runoff parameters across the entire watershed. Even when using distributed parameter schemes, they are mostly based on coarse land cover types for assignment, failing to establish a quantitative response relationship between temperature and infiltration capacity unique to different permafrost response type areas. Third, the dynamic changes in infiltration capacity during the freeze-thaw transition period are often simplified to instantaneous abrupt changes or ignored, lacking a method to automatically extract the critical temperature thresholds for the onset and stabilization of freeze-thaw cycles in each area using long-term historical observation data. Fourth, existing simulation frameworks do not effectively couple remote sensing intelligent interpretation with permafrost runoff calculation, making it impossible to perform spatiotemporally continuous runoff calculations based on the real-time temperature of the calculation unit and the temperature-infiltration capacity relationship curve of its respective zone, resulting in limited accuracy in simulating runoff processes during the spring flood season.

[0004] Therefore, this invention is proposed. Summary of the Invention

[0005] To address these issues, this invention provides a method and system for simulating runoff driven by the freeze-thaw state of soil in cold-region permafrost watersheds. This method overcomes the problems of low efficiency and strong subjectivity in the spatial division of permafrost response types in existing technologies, lack of quantitative response relationships between temperature and infiltration capacity unique to different permafrost types, insufficient characterization of the gradual change process of infiltration capacity during the freeze-thaw transition period, and ineffective coupling between remote sensing intelligent interpretation and permafrost runoff calculation.

[0006] To achieve the above objectives, this invention provides a runoff simulation method driven by soil freeze-thaw state in cold-region permafrost watersheds, comprising: S1. Acquire multi-source remote sensing images, digital elevation models and permafrost field survey samples of the target watershed, and construct a permafrost response type dataset. S2 takes multi-source remote sensing images and digital elevation models as input and permafrost response type as output to train a semantic segmentation model based on deep learning. S3, use the trained semantic segmentation model to extrapolate the target watershed and generate a permafrost response type zoning map; S4. Based on historical temperature and runoff observation data of the target watershed, determine the initial temperature threshold that causes the runoff coefficient to rise and the stable temperature threshold that prevents the runoff coefficient from continuing to decrease in each frozen soil response type zone; establish the temperature-infiltration capacity relationship curves corresponding to each frozen soil response type zone. S5. Based on the frozen soil response type zoning map, the target watershed is discretized into calculation units, and each calculation unit is assigned a corresponding temperature-infiltration capacity relationship curve. S6. Obtain the real-time temperature of each calculation unit according to the time step, determine the infiltration capacity of the target time period according to the temperature-infiltration capacity relationship curve corresponding to the calculation unit, and calculate the surface runoff of each calculation unit based on the infiltration capacity. S7. Draw a heat map of runoff in the target watershed based on the surface runoff of each calculation unit.

[0007] Furthermore, the step of training a deep learning-based semantic segmentation model, using multi-source remote sensing images and digital elevation models as input and permafrost response type as output, includes: S21. Based on the permafrost field survey samples, combined with the multi-source remote sensing images and digital elevation model, polygonal regions corresponding to each permafrost response type are delineated and assigned type labels to construct a training sample set; wherein, the permafrost response types include at least fast freezing and thawing zones, slow freezing and slow thawing zones, and stagnant freezing and stagnant thawing zones. S22, Perform radiometric and atmospheric corrections on the multi-source remote sensing images, and extract a topographic factor layer from the digital elevation model, the topographic factor layer including slope, aspect and topographic humidity index; S23, Spatial registration and channel overlay of each spectral band of the corrected multi-source remote sensing image with the topographic factor layer are performed to form a multi-channel input feature map, and then cropped into training image blocks according to a fixed size; S24, Construct a deep learning semantic segmentation network based on an encoder-decoder structure. The encoder is used to extract multi-scale spatial and spectral features from the training image patch, and the decoder is used to restore the spectral features to the input resolution and output the permafrost response type probability of each pixel. S25, using the training image patch as input and the corresponding permafrost response type label as the expected output, backpropagation training is performed through a combined loss function until the result converges, thus obtaining the trained semantic segmentation model.

[0008] Furthermore, the step of using the trained semantic segmentation model to extrapolate the target watershed and generate a permafrost response type zoning map includes: S31. Perform radiometric correction, atmospheric correction, and topographic factor layer extraction on the full-frame multi-source remote sensing image of the target watershed, and crop it into image blocks to be extrapolated according to fixed pixel size and geographic coordinate grid. S32, Input the image blocks to be deduced one by one into the trained semantic segmentation model, and output the probability distribution map of the pixels belonging to each permafrost response type; S33, take the permafrost response type label corresponding to the maximum probability of each pixel for each probability distribution map, and obtain the initial discrete category prediction results for each pixel; S34, stitch together the initial prediction results of the discrete categories of each pixel into an initial raster map of the permafrost response type of the entire watershed according to the original geographic coordinate grid; S35, morphological post-processing of the initial raster map of the permafrost response type of the whole watershed to remove isolated fragments and smooth the patch boundaries; S36, the post-processed initial raster map of the whole watershed permafrost response type is converted into a vector polygon layer, and the permafrost response type attribute corresponding to each polygon is assigned to obtain the permafrost response type partition map.

[0009] Furthermore, based on historical temperature and runoff observation data of the target watershed, the initial temperature threshold that causes the runoff coefficient to rise and the stable temperature threshold that prevents the runoff coefficient from continuously decreasing are determined for each frozen soil response type zone; temperature-infiltration capacity relationship curves are established for each frozen soil response type zone, including: S41, extract the daily average temperature sequence and corresponding daily runoff sequence during the freeze-thaw period from the historical observation data of the target watershed, and calculate the daily runoff coefficient for each frozen soil response type zone; S42, the daily average temperature sequence is divided into intervals according to a predetermined temperature interval, and the average value of the daily runoff coefficient corresponding to each interval is taken to obtain the temperature-average runoff coefficient scatter sequence. S43, based on the scatter plot of the temperature-average runoff coefficient, the critical temperature corresponding to the continuous monotonic increase of the average runoff coefficient from a low value is determined as the initial temperature threshold; the critical temperature corresponding to the continuous decrease of the average runoff coefficient with the increase of temperature and then tending to stabilize is determined as the stable temperature threshold. S44, taking the initial temperature threshold and the stable temperature threshold as boundaries, when the temperature is lower than or equal to the initial temperature threshold, the infiltration capacity takes the minimum value; when the temperature is higher than or equal to the stable temperature threshold, the infiltration capacity takes the maximum value; when the temperature is between the two, the infiltration capacity smoothly transitions with the temperature. S45. Based on the correspondence between air temperature and infiltration capacity, establish the air temperature-infiltration capacity relationship curve for each frozen soil response type zone.

[0010] Furthermore, the discretization of the target watershed into computational units based on the permafrost response type zoning map, and the assignment of a corresponding temperature-infiltration capacity relationship curve to each computational unit, includes: S51, the frozen soil response type zoning map is superimposed on the target watershed boundary to generate a regular grid covering the entire watershed with a predetermined spatial resolution, which serves as the computing unit; S52, for each calculation unit, take the frozen soil response type partition into which the geometric center point falls, and assign the frozen soil response type label of the corresponding partition to the calculation unit; if the geometric center point of the calculation unit happens to fall on the partition boundary line, then assign the type label of the partition with the largest area ratio among the adjacent partitions to the calculation unit. S53, based on the mapping relationship between each frozen soil response type zone and the temperature-infiltration capacity relationship curve, match the temperature-infiltration capacity relationship curve corresponding to the frozen soil response type to each calculation unit; S54 packages and stores the spatial location, permafrost response type label, and corresponding temperature-infiltration capacity relationship curve parameters of each calculation unit, generating an attribute table for each calculation unit.

[0011] Furthermore, the process of acquiring the real-time air temperature of each calculation unit according to the time step, determining the infiltration capacity for the target time period based on the air temperature-infiltration capacity relationship curve corresponding to the calculation unit, and calculating the surface runoff of each calculation unit based on the infiltration capacity includes: S61, acquire meteorological data of the target watershed, and obtain the temperature and precipitation of each computing unit through spatial interpolation; S62, read the temperature-infiltration capacity relationship curve corresponding to each calculation unit, and determine the infiltration capacity value for the target time period using temperature as input; S63, the infiltration capacity value is converted into actual infiltrate water volume and compared with the precipitation in the target time period: if the precipitation is less than or equal to the actual infiltrate water volume, all the precipitation will infiltrate and the surface runoff will be zero; if the precipitation is greater than the actual infiltrate water volume, the excess will form surface runoff. S64. Repeat S62 to S63 to obtain the surface runoff sequence for each calculation unit in the target time period.

[0012] Furthermore, the initial temperature threshold is the critical temperature at which, during the freeze-thaw period, as the temperature decreases, the soil infiltration capacity within the frozen soil response type zone begins to change from its maximum value to its minimum value, and the runoff coefficient begins to rise monotonically from a low value. The initial temperature threshold is related to the soil texture, organic matter content, and ice content of the frozen soil response type zone, and the initial temperature threshold is different for different frozen soil response type zones.

[0013] Furthermore, the stable temperature threshold is the critical temperature at which, during the freeze-thaw period, the soil infiltration capacity of the frozen soil response type zone recovers to its maximum value and remains stable as the temperature rises, and the runoff coefficient continues to decrease and then tends to stabilize as the temperature rises. The stable temperature threshold is related to the soil texture, organic matter content and ice content of the frozen soil response type zone, and the stable temperature threshold is different for different frozen soil response type zones.

[0014] This invention also provides a runoff simulation system driven by soil freeze-thaw state in cold-region permafrost watersheds, characterized in that the system is used to implement any of the aforementioned runoff simulation methods driven by soil freeze-thaw state in cold-region permafrost watersheds, and the system includes: The sample set construction module is used to acquire multi-source remote sensing images, digital elevation models and permafrost field survey samples of the target watershed, and construct permafrost response type datasets. The segmentation model training module, connected to the sample set construction module, is used to train a semantic segmentation model based on deep learning, taking multi-source remote sensing images and digital elevation models as inputs and permafrost response types as outputs. The permafrost type differentiation module is connected to the segmentation model training module and is used to extrapolate the target watershed using the trained semantic segmentation model to generate a permafrost response type partition map. The infiltration capacity coupling module, connected to the permafrost type differentiation module, is used to determine the initial temperature threshold that causes the runoff coefficient to rise and the stable temperature threshold that prevents the runoff coefficient from continuing to decrease in each permafrost response type zone based on historical temperature and runoff observation data of the target watershed; and to establish the temperature-infiltration capacity relationship curves corresponding to each permafrost response type zone. The watershed discretization module, connected to the infiltration capacity coupling module, is used to discretize the target watershed into calculation units according to the frozen soil response type zoning map, and assign a corresponding temperature-infiltration capacity relationship curve to each calculation unit. The runoff calculation module, connected to the watershed discretization module, is used to obtain the real-time temperature of each calculation unit according to the time step, determine the infiltration capacity of the target time period according to the temperature-infiltration capacity relationship curve corresponding to the calculation unit, and calculate the surface runoff of each calculation unit based on the infiltration capacity. The runoff heat map output module is connected to the runoff calculation module and is used to draw a runoff heat map of the target watershed based on the surface runoff of each calculation unit.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: Firstly, this invention constructs a dataset by acquiring multi-source remote sensing images, digital elevation models, and permafrost field survey samples of the target watershed. Using multi-source remote sensing images and digital elevation models as input and permafrost response type as output, a semantic segmentation model based on deep learning is trained. The trained model is then used to extrapolate the target watershed to generate a permafrost response type zoning map. This solves the problems in existing technologies where the spatial division of permafrost response types relies on manual interpretation or traditional classification methods, resulting in low efficiency and strong subjectivity. It also makes it difficult to objectively capture the spatial differences in freezing characteristics caused by factors such as soil texture, organic matter content, and ice content. This invention improves the automation, objectivity, and spatial resolution of permafrost response type zoning. Secondly, this invention automatically determines the initial temperature threshold that causes the runoff coefficient to rise and the stable temperature threshold that prevents the runoff coefficient from continuously decreasing in each frozen soil response type zone based on historical temperature and runoff observation data of the target watershed. It establishes a unique temperature-infiltration capacity relationship curve for each frozen soil response type zone, and discretizes the target watershed into calculation units according to the frozen soil response type zone map. It obtains the real-time temperature of each calculation unit according to the time step, determines the infiltration capacity of the target time period according to the temperature-infiltration capacity relationship curve corresponding to the calculation unit, and then calculates the surface runoff of each calculation unit. This solves the problem that existing models usually assume that the entire watershed has uniform frozen soil runoff parameters and lack a unique temperature-infiltration capacity quantitative response relationship for different frozen soil response type zones. It improves the ability to express the spatial heterogeneity of runoff parameters in different frozen soil type zones and the accuracy of the continuous characterization of runoff process during the freeze-thaw transition period. Thirdly, this invention effectively couples intelligent interpretation of permafrost response types based on a deep learning semantic segmentation model with distributed runoff calculation based on the temperature-infiltration capacity relationship curve. Finally, it draws a runoff heat map of the target watershed based on the surface runoff of each calculation unit. This solves the problem that existing simulation frameworks do not effectively couple remote sensing intelligent interpretation with permafrost runoff calculation and cannot perform spatiotemporally continuous runoff calculation based on the real-time temperature of the calculation unit and the temperature-infiltration capacity relationship curve of its respective zone. It realizes the full-chain automated processing from permafrost spatial heterogeneity identification to runoff dynamic simulation, improves the spatiotemporal continuity and prediction accuracy of runoff process simulation during the spring flood season in cold permafrost watersheds, enhances the visualization ability of simulation results and the decision support ability for water resource management and spring flood disaster early warning in cold regions. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A flowchart of a method for simulating runoff driven by the freeze-thaw state of soil in a cold region permafrost watershed, as provided in an embodiment of the present invention. Figure 2 The structural block diagram of the runoff simulation system driven by the freeze-thaw state of soil in the cold region watershed provided in the embodiment of the present invention. Detailed Implementation

[0018] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0019] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0020] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0021] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0022] Example 1 like Figure 1 As shown, this invention provides a runoff simulation method driven by soil freeze-thaw state in cold-region permafrost watersheds, comprising: S1. Acquire multi-source remote sensing images, digital elevation models and permafrost field survey samples of the target watershed, and construct a permafrost response type dataset. In one possible implementation, taking a typical permafrost watershed in Northeast my country as an example, Sentinel-2 multispectral remote sensing images (spatial resolution 10m) and Landsat-8 thermal infrared remote sensing images (spatial resolution 100m) covering the watershed are acquired and resampled to 30m. An ASTER GDEM digital elevation model (spatial resolution 30m) is downloaded from a geospatial data cloud platform. Simultaneously, permafrost field survey samples are collected from 42 borehole sites within and around the watershed. Sample records include information such as geographical location, soil type, organic matter content level, volumetric ice content, active layer thickness, and observational descriptions of freeze-thaw processes.

[0023] Forty-two field survey samples were randomly divided into training and validation sets in a 7:3 ratio. Based on three indicators—soil type, organic matter content, and ice content—the corresponding geographical locations of each sample point were labeled as three permafrost response types: rapid freezing-rapid thawing, slow freezing-slow thawing, or prolonged freezing-slow thawing. This formed an initial dataset containing geographical coordinates and permafrost response type labels. For example, sample points with sandy soil, low organic matter content, and low ice content were labeled as rapid freezing-rapid thawing; sample points with clay soil, high organic matter content, and high ice content were labeled as prolonged freezing-slow thawing; and other intermediate types were labeled as slow freezing-slow thawing. Based on this, the permafrost response type dataset was constructed.

[0024] S2, taking multi-source remote sensing imagery and a digital elevation model as input, and permafrost response type as output, trains a deep learning-based semantic segmentation model, including: S21. Based on the permafrost field survey samples, combined with the multi-source remote sensing images and digital elevation model, polygonal regions corresponding to each permafrost response type are delineated and assigned type labels to construct a training sample set; wherein, the permafrost response types include at least fast freezing and thawing zones, slow freezing and slow thawing zones, and stagnant freezing and stagnant thawing zones. S22, Perform radiometric and atmospheric corrections on the multi-source remote sensing images, and extract a topographic factor layer from the digital elevation model, the topographic factor layer including slope, aspect and topographic humidity index; S23, Spatial registration and channel overlay of each spectral band of the corrected multi-source remote sensing image with the topographic factor layer are performed to form a multi-channel input feature map, and then cropped into training image blocks according to a fixed size; S24, Construct a deep learning semantic segmentation network based on an encoder-decoder structure. The encoder is used to extract multi-scale spatial and spectral features from the training image patch, and the decoder is used to restore the spectral features to the input resolution and output the permafrost response type probability of each pixel. S25, using the training image patch as input and the corresponding permafrost response type label as the expected output, backpropagation training is performed through a combined loss function until the result converges, thus obtaining the trained semantic segmentation model.

[0025] In one possible implementation, Centered on the training set sample points labeled in S1, and referencing the visual features and elevation changes of Sentinel-2 imagery, polygonal boundaries of homogeneous areas surrounding each sample point were manually drawn, with each polygon having an area of ​​no less than 0.5 km². Each polygon was assigned a permafrost response type label corresponding to its central sample point. Several polygons of each of the three types—rapid freezing and thawing zones, slow freezing and thawing zones, and stagnant freezing and stagnant thawing zones—were drawn, forming a raster-level training sample set containing multispectral imagery, thermal infrared imagery, digital elevation model, and permafrost response type labels.

[0026] The Sen2Cor atmospheric correction algorithm was used for Sentinel-2 imagery, and a single-channel algorithm was used to retrieve land surface temperature from Landsat-8 thermal infrared imagery. Both were converted into land surface reflectance / land surface temperature products. Based on ASTER GDEM, slope and aspect rasters were calculated using ArcGIS or the open-source GDAL tool, and topographic humidity index rasters were calculated using the D8 flow direction algorithm or multi-flow direction algorithm, forming three topographic factor layers: slope, aspect, and topographic humidity index. The spatial resolution was uniformly resampled to 10m.

[0027] Using a uniform spatial resolution of 10m, the Sentinel-2 blue, green, red, and near-infrared bands, the Landsat-8 land surface temperature band, and three topographic factor layers (slope, aspect, and topographic humidity index), totaling eight channels, were spatially registered pixel-by-pixel and superimposed into a single multi-channel input feature map. A sliding window method was employed, with a window size of 256×256 pixels and a stride of 128 pixels, to crop the multi-channel input feature map into several training image patches, each retaining its original geographic reference information.

[0028] U-Net is used as the backbone network. The encoder consists of five convolutional stages, each containing two 3×3 convolutional layers and one 2×2 max-pooling layer, extracting high-level semantic features stage by stage. The decoder consists of four upsampling stages, each restoring the feature map resolution through transposed convolution and fusing it with the feature map of the corresponding stage in the encoder via skip connections. Finally, a 1×1 convolutional layer and a Softmax activation function output the probability of each pixel belonging to one of three categories: fast-freeze / fast-thaw, slow-freeze / slow-thaw, or stagnant-freeze / stagnant. The input layer has 8 channels, along with multi-channel features. Figure 1 To.

[0029] The training sample set constructed in S21 was divided into a training set and a validation set. The Adam optimizer was used, with an initial learning rate of 0.0001 and a batch size of 8. The weighted sum of cross-entropy loss and Dice coefficient loss was used as the combined loss function. During training, the average intersection-over-union ratio (IoU) was monitored on the validation set. Training was stopped when the IoU no longer improved after several consecutive rounds. The model weights with the optimal IoU on the validation set were saved as the trained semantic segmentation model.

[0030] This invention obtains true surface reflectance data by performing radiometric and atmospheric correction on remote sensing images, and extracts topographic factors such as slope, aspect, and topographic humidity index from digital elevation models. This solves the problems of atmospheric interference and lack of topographic information in the original images, and improves the ability of the semantic segmentation model's input features to represent environmental conditions related to permafrost response types. By spatially registering and channel-overlaying multispectral bands, surface temperature bands, and various topographic factor layers to form a unified multi-channel input feature map, this invention solves the problems of inconsistent spatial benchmarks and isolated features from different data sources, and improves the ability of deep learning models to simultaneously use multi-source heterogeneous information such as spectral, thermal infrared, and topographic data to distinguish permafrost types. By constructing a semantic segmentation network based on an encoder-decoder structure, the encoder extracts multi-scale spatial and spectral features, and the decoder restores the resolution and outputs the probability of permafrost response type for each pixel. This solves the shortcomings of traditional pixel-by-pixel classification methods that ignore spatial context information, and improves the boundary fineness and classification consistency of permafrost response type partitioning.

[0031] S3, using the trained semantic segmentation model to extrapolate the target watershed, generating a permafrost response type zoning map, including: S31. Perform radiometric correction, atmospheric correction, and topographic factor layer extraction on the full-frame multi-source remote sensing image of the target watershed, and crop it into image blocks to be extrapolated according to fixed pixel size and geographic coordinate grid. S32, Input the image blocks to be deduced one by one into the trained semantic segmentation model, and output the probability distribution map of the pixels belonging to each permafrost response type; S33, take the permafrost response type label corresponding to the maximum probability of each pixel for each probability distribution map, and obtain the initial discrete category prediction results for each pixel; S34, stitch together the initial prediction results of the discrete categories of each pixel into an initial raster map of the permafrost response type of the entire watershed according to the original geographic coordinate grid; S35, morphological post-processing of the initial raster map of the permafrost response type of the whole watershed to remove isolated fragments and smooth the patch boundaries; S36, the post-processed initial raster map of the whole watershed permafrost response type is converted into a vector polygon layer, and the permafrost response type attribute corresponding to each polygon is assigned to obtain the permafrost response type partition map.

[0032] In one possible implementation, Sen2Cor atmospheric correction is applied to Sentinel-2 images covering the entire target watershed to obtain surface reflectance. A single-channel algorithm is used to invert surface temperature in the Landsat-8 thermal infrared band. Three topographic factor layers—slope, aspect, and topographic humidity index—are calculated based on ASTER GDEM. All layers are uniformly resampled to a 10m spatial resolution. After overlaying these eight channels into a multi-channel input feature map of the entire watershed, a 256×256 pixel window size and 128 pixel step size, consistent with S23, are used for regular cropping along the geographic coordinate grid, with no or partial overlap, generating several image patches to be extrapolated. The generated image patches to be inferred are sequentially input into the U-Net semantic segmentation model. The model performs forward inference on each image patch, outputting a probability distribution map of size 256×256×3. The three channels correspond to the predicted probabilities of rapid freezing and thawing zones, slow freezing and slow thawing zones, and stagnant freezing and stagnant thawing zones, respectively, with the sum of the three class probabilities for each pixel being 1. For the output probability distribution map, the Argmax operation is used to compare the probability values ​​of the three channels pixel by pixel. The class channel corresponding to the highest probability value is encoded as 1, 2, or 3, representing rapid freezing and rapid thawing zones, slow freezing and slow thawing zones, and stagnant freezing and stagnant thawing zones, respectively, generating a single-band category label map with the same size as the input image patch. Based on the geographical coordinates of the upper left corner recorded for each image patch, the generated image patch category label maps are mosaicked in situ one by one to form a single-band raster file covering the entire watershed, with pixel values ​​of 1, 2, or 3, i.e., the initial raster map of the permafrost response type for the entire watershed.

[0033] Morphological opening is employed to first erode the initial raster image, eliminating isolated fragments with areas smaller than a preset threshold. Then, an equal-scale dilation operation is performed to restore the original boundaries of the remaining patches. Finally, majority filtering is used to smooth the patch boundaries, eliminating jagged classification jumps that may occur at the splicing edges. The processed raster image is converted into a vector polygon layer, where adjacent cells with the same label type are automatically aggregated into a polygon. A permafrost response type field is added to the attribute table for each polygon, with values ​​for fast-freezing / fast-thawing zone, slow-freezing / slow-thawing zone, and prolonged freezing / prolonged thawing zone, ultimately generating a permafrost response type zoning map for the target watershed.

[0034] This invention addresses the inconsistency in feature distribution between input and training data during the extrapolation phase by processing full-basin images and cropping them into image patches to be extrapolated using the same radiometric correction, atmospheric correction, and topographic factor extraction methods as in the training phase. This improves the predictive stability of the semantic segmentation model during full-area extrapolation. By inputting each image patch to be extrapolated into the semantic segmentation model and outputting a pixel-level probability distribution map, this invention solves the problem that traditional manual interpretation cannot efficiently cover the entire basin, thus improving the automation of spatial mapping of permafrost response types. By taking the maximum probability for each pixel in the probability distribution map to obtain the initial prediction results for discrete categories, this invention solves the problem that probability output cannot be directly used for spatial partitioning, thus improving the certainty and operability of the classification results. By stitching together the prediction results of each image patch into an initial raster map of the entire watershed according to the original geographic coordinate grid, the problem that the patch extrapolation results cannot form a complete watershed coverage is solved, improving the global integrity and spatial continuity of the zoning results. By performing morphological post-processing on the initial raster map to remove isolated fragments and smooth patch boundaries, the common problems of salt-and-pepper noise and jagged edges in deep learning prediction are solved, improving the cleanliness of the zoning map and the regularity of patch boundaries. By converting the raster map into a vector polygon layer and assigning a permafrost response type attribute, the problem that the raster zoning results are not convenient for subsequent calculation unit matching and model parameter assignment is solved, improving the applicability and connection efficiency of the permafrost response type zoning map to runoff simulation steps.

[0035] S4. Based on historical temperature and runoff observation data of the target watershed, determine the initial temperature threshold that causes the runoff coefficient to rise and the stable temperature threshold that prevents the runoff coefficient from continuously decreasing in each frozen soil response type zone; establish the temperature-infiltration capacity relationship curves corresponding to each frozen soil response type zone, including: S41, extract the daily average temperature sequence and corresponding daily runoff sequence during the freeze-thaw period from the historical observation data of the target watershed, and calculate the daily runoff coefficient for each frozen soil response type zone; S42, the daily average temperature sequence is divided into intervals according to a predetermined temperature interval, and the average value of the daily runoff coefficient corresponding to each interval is taken to obtain the temperature-average runoff coefficient scatter sequence. S43, based on the scatter plot of the temperature-average runoff coefficient, the critical temperature corresponding to the continuous monotonic increase of the average runoff coefficient from a low value is determined as the initial temperature threshold; the critical temperature corresponding to the continuous decrease of the average runoff coefficient with the increase of temperature and then tending to stabilize is determined as the stable temperature threshold. S44, taking the initial temperature threshold and the stable temperature threshold as boundaries, when the temperature is lower than or equal to the initial temperature threshold, the infiltration capacity takes the minimum value; when the temperature is higher than or equal to the stable temperature threshold, the infiltration capacity takes the maximum value; when the temperature is between the two, the infiltration capacity smoothly transitions with the temperature. S45. Based on the correspondence between air temperature and infiltration capacity, establish the air temperature-infiltration capacity relationship curve for each frozen soil response type zone.

[0036] The initial temperature threshold is the critical temperature at which, during the freeze-thaw period, the soil infiltration capacity of the frozen soil response type zone begins to change from its maximum value to its minimum value, and the runoff coefficient begins to rise monotonically from a low value. The initial temperature threshold is related to the soil texture, organic matter content, and ice content of the frozen soil response type zone, and the initial temperature threshold is different for different frozen soil response type zones.

[0037] The stable temperature threshold is the critical temperature at which, during the freeze-thaw period, the soil infiltration capacity of the frozen soil response type zone recovers to its maximum value and remains stable as the temperature rises, and the runoff coefficient continues to decrease and then tends to stabilize as the temperature rises. The stable temperature threshold is related to the soil texture, organic matter content and ice content of the frozen soil response type zone, and the stable temperature threshold is different for different frozen soil response type zones.

[0038] In one possible implementation, based on 11 years of hydrological and meteorological observation data of the target watershed from 2010 to 2020, the period from March 1st to May 31st of each year is extracted as the freeze-thaw period. The daily average temperature of the watershed surface and the measured daily runoff at the watershed outlet hydrological station are obtained during this period. According to the generated permafrost response type zoning map, the area of ​​rapid freezing and thawing zones accounts for approximately 35% of the watershed, slow freezing and slow thawing zones account for approximately 40%, and areas of prolonged freezing and prolonged thawing account for approximately 25%. The daily runoff coefficient is obtained by dividing the daily runoff by the average precipitation of the watershed surface on that day. Days with precipitation less than 0.5 mm are excluded to avoid abnormally high runoff coefficients due to light rainfall. The arithmetic mean of the daily runoff coefficients for each zone is taken as the representative runoff coefficient for that zone on that calendar day.

[0039] Using 1℃ as a predetermined temperature interval, several temperature intervals were divided within the main temperature variation range of the freeze-thaw period, i.e., -10℃ to 15℃. The daily runoff coefficients of all days during the freeze-thaw period within each frozen soil response type zone that fall within the same interval were arithmetically averaged to obtain the temperature-average runoff coefficient scatter plot sequence for that zone. For example, in the rapid freezing and thawing zone, the average runoff coefficient is 0.31 in the [-2℃, -1℃) interval, 0.38 in the [0℃, 1℃) interval, and 0.12 in the [3℃, 4℃) interval; in the delayed freezing and thawing zone, the average runoff coefficient is 0.28 in the [-2℃, -1℃) interval, 0.33 in the [0℃, 1℃) interval, 0.25 in the [3℃, 4℃) interval, and 0.11 in the [8℃, 9℃) interval.

[0040] The scatter sequence of the rapid freezing and rapid thawing zone is analyzed. When the average runoff coefficient is below -1°C, it is stably maintained at a low level of about 0.08. Starting from -1°C, it shows a continuous monotonic upward trend as the air temperature rises, and the starting air temperature threshold is determined to be -1°C; the average runoff coefficient first rises and then falls in the range of 0°C to 3°C, and continues to decrease as the air temperature rises starting from 3°C. After reaching 4°C, it stabilizes at about 0.06 and basically does not change any more, and the stable air temperature threshold is determined to be 4°C. For the slow freezing and slow thawing zone, the starting air temperature threshold is determined to be 0.5°C, and the stable air temperature threshold is determined to be 7°C; for the delayed freezing and delayed thawing zone, the starting air temperature threshold is determined to be 1.5°C, and the stable air temperature threshold is determined to be 10°C.

[0041] Smooth transition is achieved by a linear interpolation method.

[0042] In this embodiment, the starting air temperature threshold of the rapid freezing and rapid thawing zone is -1°C, and the stable air temperature threshold is 4°C. The maximum value of infiltration capacity is defined as 1.0, and the minimum value is defined as 0.1. Then when the air temperature T ≤ -1°C, the infiltration capacity of this zone is 0.1; when the air temperature T ≥ 4°C, the infiltration capacity is 1.0; when -1°C < T < 4°C, the infiltration capacity is linearly interpolated between 0.1 and 1.0. The slow freezing and slow thawing zone and the delayed freezing and delayed thawing zone are processed by the same rule with their respective thresholds respectively.

[0043] With air temperature as the abscissa and infiltration capacity as the ordinate, the thresholds of each zone and the interpolation rules determined in S44 are drawn into curves or a lookup table is established. The curve characteristics of the rapid freezing and rapid thawing zone are as follows: the infiltration capacity is constantly 0.1 below -1°C, rises linearly in the interval from -1°C to 4°C, and is constantly 1.0 above 4°C. The curves of the slow freezing and slow thawing zone and the delayed freezing and delayed thawing zone have similar shapes, but the thresholds are shifted to the right, and the transition interval is wider and gentler. The parameters of the three curves are stored in a lookup table format for calling in subsequent steps. Each lookup table records five parameters: the corresponding frozen soil response type, the starting air temperature threshold, the stable air temperature threshold, the minimum infiltration capacity and the maximum infiltration capacity.

[0044] The present invention automatically identifies the critical air temperature at which the runoff coefficient continues to monotonically rise from a low value and tends to be stable after continuous decline based on the air temperature-average runoff coefficient scatter sequence, solves the problem that traditional methods rely on experience to manually set freeze-thaw thresholds and lack objective quantitative criteria, and improves the accuracy of the starting air temperature threshold and the stable air temperature threshold; by taking the starting air temperature threshold and the stable air temperature threshold as boundaries, assigning the minimum and maximum infiltration capacity to the two end intervals respectively, and performing smooth transition in the transition interval, it solves the problem that traditional frozen soil runoff generation models simplify the freeze-thaw infiltration capacity as binary mutation or ignore the gradual change process, and improves the accuracy and smoothness of the continuous change process of infiltration capacity during the frozen soil thawing period.

[0045] S5, based on the permafrost response type zoning map, the target watershed is discretized into computational units, and each computational unit is assigned a corresponding temperature-infiltration capacity relationship curve, including: S51, the permafrost response type zoning map is superimposed on the target watershed boundary to generate a regular grid covering the entire watershed with a predetermined spatial resolution, which serves as the computing unit; S52, for each calculation unit, take the frozen soil response type partition into which the geometric center point falls, and assign the frozen soil response type label of the corresponding partition to the calculation unit; if the geometric center point of the calculation unit happens to fall on the partition boundary line, then assign the type label of the partition with the largest area ratio among the adjacent partitions to the calculation unit. S53, based on the mapping relationship between each frozen soil response type zone and the temperature-infiltration capacity relationship curve, match the temperature-infiltration capacity relationship curve corresponding to the frozen soil response type of each calculation unit; S54 packages and stores the spatial location, permafrost response type label, and corresponding temperature-infiltration capacity relationship curve parameters of each calculation unit, generating an attribute table for each calculation unit.

[0046] In one possible implementation, the generated permafrost response type zoning map is loaded into a GIS platform and overlaid with the watershed boundary vector file. Using a 1km × 1km spatial resolution, regular square grids are generated with the watershed boundary's bounding rectangle as the area. All grids whose center points are located within the watershed boundary are retained, while grids completely outside the watershed boundary are discarded, ultimately resulting in several regular grids covering the entire watershed as computational units. The watershed area is approximately 4200 km², generating approximately 4200 1km² computational units.

[0047] Using spatial connectivity tools, the coordinates of the geometric center point of each computational unit are extracted. It is then determined whether the point falls within the polygonal area of ​​a fast-freezing / fast-thawing zone, a slow-freezing / slow-thawing zone, or a stagnant / stagnant zone. The type label for the zone is assigned to the computational unit. For the rare case where the center point happens to fall on the boundary line of two or more zones, the overlap area between the computational unit and each adjacent zone is automatically calculated, and the zone label with the largest overlap area percentage is assigned to the computational unit. For example, if the center point of a computational unit falls on the boundary line between a fast-freezing / fast-thawing zone and a slow-freezing / slow-thawing zone, and the overlap area between the unit and the fast-freezing / fast-thawing zone is 70%, and the overlap area with the slow-freezing / slow-thawing zone is 30%, then the unit is labeled as a fast-freezing / fast-thawing zone.

[0048] Each calculation unit is assigned a unique number, and each record consists of seven fields: the X-coordinate and Y-coordinate of the calculation unit's center point, the frozen soil response type label, the initial temperature threshold, the stable temperature threshold, the minimum infiltration capacity, and the maximum infiltration capacity. Approximately 4200 records from the entire watershed are aggregated into an attribute table and stored in CSV format. For example, the attribute record for calculation unit number 1024 is: (512300, 4837200, slow freezing and slow thawing zone, 0.5, 7, 0.1, 1.0).

[0049] This invention solves the problem that traditional sub-basin division methods do not fully match the boundaries of frozen soil zoning and cannot accurately express the consistency of freeze-thaw characteristics within the unit by overlaying the frozen soil response type zoning map with the boundary of the target watershed and generating a regular grid covering the entire watershed with a predetermined spatial resolution. This improves the accuracy of the computing unit in characterizing the spatial heterogeneity of frozen soil and the spatial resolution of runoff generation calculation.

[0050] S6, acquire the real-time temperature of each calculation unit according to the time step, determine the infiltration capacity for the target time period based on the temperature-infiltration capacity relationship curve corresponding to the calculation unit, and calculate the surface runoff of each calculation unit based on the infiltration capacity, including: S61, acquire meteorological data of the target watershed, and obtain the temperature and precipitation of each computing unit through spatial interpolation; S62, read the temperature-infiltration capacity relationship curve corresponding to each calculation unit, and determine the infiltration capacity value for the target time period using temperature as input; S63, the infiltration capacity value is converted into actual infiltrate water volume and compared with the precipitation in the target time period: if the precipitation is less than or equal to the actual infiltrate water volume, all the precipitation will infiltrate and the surface runoff will be zero; if the precipitation is greater than the actual infiltrate water volume, the excess will form surface runoff. S64. Repeat S62 to S63 to obtain the surface runoff sequence for each calculation unit in the target time period.

[0051] In one possible implementation, numerical weather forecasts for the next three days are obtained at a time resolution of 3 hours, with forecast elements including near-surface temperature and precipitation. The 3-hour forecast data are aggregated into daily average temperature and daily precipitation. Bilinear interpolation is then used to interpolate the forecast data to the same computational unit center point as S5, generating daily temperature and precipitation forecast sequences for each computational unit for the next three days. Taking computational unit 1728 as an example, unit 1728 belongs to the frozen-thaw zone, with an initial temperature threshold of 1.5℃, a stable temperature threshold of 10℃, a minimum infiltration capacity of 0.1, and a maximum of 1.0. On April 21st, the forecast temperature is 5.5℃, falling into the transition range, and the infiltration capacity value is calculated to be 0.52 through linear interpolation. On April 22nd, the forecast temperature is 7.2℃, and the infiltration capacity value rises to 0.70. On April 23rd, the forecast temperature is 9.8℃, approaching the stable temperature threshold, and the infiltration capacity value rises to 0.98, indicating that soil infiltration capacity has basically recovered. Calculation unit 0356 belongs to the rapid freezing and thawing zone. The forecast temperature has exceeded its stable temperature threshold of 4℃, and the infiltration capacity value has remained at its maximum value of 1.0 for three consecutive days.

[0052] The maximum daily infiltration rate for calculation unit 1728 is 12 mm. On April 21st, the infiltration capacity value is 0.52, the actual infiltrate is 0.52 × 12 = 6.24 mm, the forecast precipitation is 14.3 mm, the excess is 14.3 - 6.24 = 8.06 mm, and the predicted runoff for this unit on that day is 8.06 mm. On April 22nd, the infiltration capacity value is 0.70, the actual infiltrate is 8.40 mm, the forecast precipitation is 0 mm, and the predicted surface runoff is zero. On April 23rd, the infiltration capacity value is 0.98, the actual infiltrate is 11.76 mm, the forecast precipitation is 8.7 mm, all precipitation has infiltrated, and the predicted surface runoff is also zero. The predicted runoff for this calculation unit over three days is arranged chronologically as [8.06, 0, 0] mm. For approximately 4200 computational units across the entire watershed, steps S62 and S63 are executed daily to complete the daily runoff forecast for the next three days. The final result is a 72-hour surface runoff forecast sequence for each computational unit, stored in a matrix format with the computational unit number as the row and the forecast date as the column.

[0053] This invention solves the problem that traditional permafrost runoff simulation relies solely on historical observations and cannot predict future runoff by acquiring gridded meteorological forecast data of the target watershed and spatially interpolating it to each computing unit, thereby improving the short-term forecasting capability of surface runoff in permafrost watersheds in cold regions. By converting infiltration capacity values ​​into actual infiltrative water volume and comparing them with forecasted precipitation to determine excess infiltration runoff, this invention solves the problem that future runoff cannot be accurately calculated when precipitation and infiltration capacity do not match in permafrost regions, thereby improving the accuracy of future surface runoff prediction for each computing unit.

[0054] S7. Draw a heat map of runoff in the target watershed based on the surface runoff of each calculation unit.

[0055] In one possible implementation, the predicted surface runoff data for 4200 calculation units across the entire watershed as of April 21, 2025, is loaded into a GIS platform. Using the geometric center coordinates of each calculation unit as the spatial positioning basis, the runoff values ​​are divided into five levels using a natural discontinuity grading method. In the zero-runoff zone, the surface runoff is 0 mm. In low-yield areas, surface runoff is 0.1-3.0 mm; In the medium-yield runoff area, the surface runoff is 3.1-8.0 mm; In high-yield runoff areas, surface runoff is 8.1-15.0 mm; In areas with extremely high runoff, surface runoff is >15.0 mm. Each calculation unit is assigned a color gradient from light yellow to dark red, with light yellow for zero runoff areas and dark red for extremely high runoff areas. The watershed boundary, main river channels, and permafrost response type zoning boundaries are overlaid as base map elements. A legend, scale bar, and north arrow are added, and a runoff heatmap at 300 dpi resolution is output. Runoff heatmaps for April 22nd and 23rd are generated simultaneously on the same day. The three-day heatmaps are displayed in chronological order, visually illustrating the spatial evolution of the watershed's runoff area shrinking from widespread distribution to scattered gully areas as temperatures rise, providing a spatial basis for spring flood risk early warning.

[0056] Example 2 like Figure 2 As shown, the present invention also provides a runoff simulation system driven by the freeze-thaw state of soil in a cold region permafrost watershed, characterized in that the system is used to implement the runoff simulation method driven by the freeze-thaw state of soil in a cold region permafrost watershed as described in any of Embodiment 1, and the system includes: The sample set construction module is used to acquire multi-source remote sensing images, digital elevation models and permafrost field survey samples of the target watershed, and construct permafrost response type datasets. The segmentation model training module, connected to the sample set construction module, is used to train a semantic segmentation model based on deep learning, taking multi-source remote sensing images and digital elevation models as inputs and permafrost response types as outputs. The permafrost type differentiation module is connected to the segmentation model training module and is used to extrapolate the target watershed using the trained semantic segmentation model to generate a permafrost response type partition map. The infiltration capacity coupling module, connected to the permafrost type differentiation module, is used to determine the initial temperature threshold that causes the runoff coefficient to rise and the stable temperature threshold that prevents the runoff coefficient from continuing to decrease in each permafrost response type zone based on historical temperature and runoff observation data of the target watershed; and to establish the temperature-infiltration capacity relationship curves corresponding to each permafrost response type zone. The watershed discretization module, connected to the infiltration capacity coupling module, is used to discretize the target watershed into calculation units according to the frozen soil response type zoning map, and assign a corresponding temperature-infiltration capacity relationship curve to each calculation unit. The runoff calculation module, connected to the watershed discretization module, is used to obtain the real-time temperature of each calculation unit according to the time step, determine the infiltration capacity of the target time period according to the temperature-infiltration capacity relationship curve corresponding to the calculation unit, and calculate the surface runoff of each calculation unit based on the infiltration capacity. The runoff heat map output module is connected to the runoff calculation module and is used to draw a runoff heat map of the target watershed based on the surface runoff of each calculation unit.

[0057] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for simulating runoff driven by soil freeze-thaw state in frozen soil watersheds of cold regions, characterized in that, include: S1. Acquire multi-source remote sensing images, digital elevation models and permafrost field survey samples of the target watershed, and construct a permafrost response type dataset. S2, taking multi-source remote sensing images and digital elevation models as input and permafrost response types as output, trains a semantic segmentation model based on deep learning; wherein, the permafrost response types include fast freezing and fast thawing zones, slow freezing and slow thawing zones, and stagnant freezing and stagnant thawing zones. S3, use the trained semantic segmentation model to extrapolate the target watershed and generate a permafrost response type zoning map; S4. Based on historical temperature and runoff observation data of the target watershed, determine the initial temperature threshold that causes the runoff coefficient to increase and the stable temperature threshold that prevents the runoff coefficient from continuously decreasing in each frozen soil response type zone. Using the initial temperature threshold and the stable temperature threshold as boundaries, when the temperature is lower than or equal to the initial temperature threshold, the infiltration capacity takes the minimum value; when the temperature is higher than or equal to the stable temperature threshold, the infiltration capacity takes the maximum value; when the temperature is between the two, the infiltration capacity smoothly transitions with the temperature. Based on the correspondence between temperature and infiltration capacity, establish the temperature-infiltration capacity relationship curve for each frozen soil response type zone. S5. Based on the frozen soil response type zoning map, the target watershed is discretized into calculation units, and each calculation unit is assigned a corresponding temperature-infiltration capacity relationship curve. S6. Obtain the real-time temperature of each calculation unit according to the time step, determine the infiltration capacity of the target time period according to the temperature-infiltration capacity relationship curve corresponding to the calculation unit, and calculate the surface runoff of each calculation unit based on the infiltration capacity. S7. Draw a heat map of runoff in the target watershed based on the surface runoff of each calculation unit.

2. The method for simulating runoff driven by soil freeze-thaw state in cold-region permafrost watersheds according to claim 1, characterized in that, The process of training a deep learning-based semantic segmentation model, using multi-source remote sensing imagery and digital elevation models as input and permafrost response type as output, includes: S21. Based on the permafrost field survey samples, combined with the multi-source remote sensing images and digital elevation model, polygonal regions corresponding to each permafrost response type are delineated and assigned type labels to construct a training sample set. S22, Perform radiometric and atmospheric corrections on the multi-source remote sensing images, and extract a topographic factor layer from the digital elevation model, the topographic factor layer including slope, aspect and topographic humidity index; S23, Spatial registration and channel overlay of each spectral band of the corrected multi-source remote sensing image with the topographic factor layer are performed to form a multi-channel input feature map, and then cropped into training image blocks according to a fixed size. S24, Construct a deep learning semantic segmentation network based on an encoder-decoder structure. The encoder is used to extract multi-scale spatial and spectral features from the training image patch, and the decoder is used to restore the spectral features to the input resolution and output the permafrost response type probability of each pixel. S25, using the training image patch as input and the corresponding permafrost response type label as the expected output, backpropagation training is performed through a combined loss function until the result converges, thus obtaining the trained semantic segmentation model.

3. The method for simulating runoff driven by soil freeze-thaw state in cold-region frozen soil watersheds according to claim 1, characterized in that, The step of using the trained semantic segmentation model to extrapolate the target watershed and generate a permafrost response type zoning map includes: S31. Perform radiometric correction, atmospheric correction, and topographic factor layer extraction on the full-frame multi-source remote sensing image of the target watershed, and crop it into image blocks to be extrapolated according to fixed pixel size and geographic coordinate grid. S32, Input the image blocks to be deduced one by one into the trained semantic segmentation model, and output the probability distribution map of the pixels belonging to each permafrost response type; S33, take the permafrost response type label corresponding to the maximum probability of each pixel for each probability distribution map, and obtain the initial discrete category prediction results for each pixel; S34, stitch together the initial prediction results of the discrete categories of each pixel into an initial raster map of the permafrost response type of the entire watershed according to the original geographic coordinate grid; S35, morphological post-processing of the initial raster map of the permafrost response type of the whole watershed to remove isolated fragments and smooth the patch boundaries; S36, the post-processed initial raster map of the whole watershed permafrost response type is converted into a vector polygon layer, and the permafrost response type attribute corresponding to each polygon is assigned to obtain the permafrost response type partition map.

4. The method for simulating runoff driven by soil freeze-thaw state in cold-region frozen soil watersheds according to claim 1, characterized in that, Based on historical temperature and runoff observation data of the target watershed, the initial temperature threshold that causes the runoff coefficient to rise and the stable temperature threshold that prevents the runoff coefficient from continuing to decrease are determined in each frozen soil response type zone. Establish the temperature-infiltration capacity relationship curves for each frozen soil response type zone, including: S41, extract the daily average temperature sequence and corresponding daily runoff sequence during the freeze-thaw period from the historical observation data of the target watershed, and calculate the daily runoff coefficient for each frozen soil response type zone; S42, the daily average temperature sequence is divided into intervals according to a predetermined temperature interval, and the average value of the daily runoff coefficient corresponding to each interval is taken to obtain the temperature-average runoff coefficient scatter sequence. S43, based on the scatter plot of the temperature-average runoff coefficient, the critical temperature corresponding to the continuous monotonic increase of the average runoff coefficient from a low value is determined as the initial temperature threshold; the critical temperature corresponding to the continuous decrease of the average runoff coefficient with the increase of temperature and then tending to stabilize is determined as the stable temperature threshold. S44, taking the initial temperature threshold and the stable temperature threshold as boundaries, when the temperature is lower than or equal to the initial temperature threshold, the infiltration capacity takes the minimum value; when the temperature is higher than or equal to the stable temperature threshold, the infiltration capacity takes the maximum value; when the temperature is between the two, the infiltration capacity smoothly transitions with the temperature. S45. Based on the correspondence between air temperature and infiltration capacity, establish the air temperature-infiltration capacity relationship curve for each frozen soil response type zone.

5. The method for simulating runoff driven by soil freeze-thaw state in cold-region frozen soil watersheds according to claim 1, characterized in that, The target watershed is discretized into computational units based on the permafrost response type zoning map, and each computational unit is assigned a corresponding temperature-infiltration capacity relationship curve, including: S51, the permafrost response type zoning map is superimposed on the target watershed boundary to generate a regular grid covering the entire watershed with a predetermined spatial resolution, which serves as the computing unit; S52, for each calculation unit, take the frozen soil response type partition into which the geometric center point falls, and assign the frozen soil response type label of the corresponding partition to the calculation unit; if the geometric center point of the calculation unit happens to fall on the partition boundary line, then assign the type label of the partition with the largest area ratio among the adjacent partitions to the calculation unit. S53, based on the mapping relationship between each frozen soil response type zone and the temperature-infiltration capacity relationship curve, match the temperature-infiltration capacity relationship curve corresponding to the frozen soil response type to each calculation unit; S54 packages and stores the spatial location, permafrost response type label, and corresponding temperature-infiltration capacity relationship curve parameters of each calculation unit, generating an attribute table for each calculation unit.

6. The method for simulating runoff driven by soil freeze-thaw state in cold-region permafrost watersheds according to claim 1, characterized in that, The process involves acquiring the real-time air temperature of each calculation unit according to a time step, determining the infiltration capacity for the target time period based on the air temperature-infiltration capacity relationship curve corresponding to the calculation unit, and calculating the surface runoff of each calculation unit based on the infiltration capacity, including: S61, acquire meteorological data of the target watershed, and obtain the temperature and precipitation of each computing unit through spatial interpolation; S62, read the temperature-infiltration capacity relationship curve corresponding to each calculation unit, and determine the infiltration capacity value for the target time period using temperature as input; S63, the infiltration capacity value is converted into actual infiltrate water volume and compared with the precipitation in the target time period: if the precipitation is less than or equal to the actual infiltrate water volume, all the precipitation will infiltrate and the surface runoff will be zero; if the precipitation is greater than the actual infiltrate water volume, the excess will form surface runoff. S64. Repeat S62 to S63 to obtain the surface runoff sequence for each calculation unit in the target time period.

7. The method for simulating runoff driven by soil freeze-thaw state in cold-region permafrost watersheds according to claim 1, characterized in that, The initial temperature threshold is the critical temperature at which, during the freeze-thaw period, the soil infiltration capacity of the frozen soil response type zone begins to change from its maximum value to its minimum value, and the runoff coefficient begins to rise monotonically from a low value. The initial temperature threshold is related to the soil texture, organic matter content, and ice content of the frozen soil response type zone, and the initial temperature threshold is different for different frozen soil response type zones.

8. The method for simulating runoff driven by soil freeze-thaw state in cold-region frozen soil watersheds according to claim 1, characterized in that, The stable temperature threshold is the critical temperature at which, during the freeze-thaw period, the soil infiltration capacity of the frozen soil response type zone recovers to its maximum value and remains stable as the temperature rises, and the runoff coefficient continues to decrease and then tends to stabilize as the temperature rises. The stable temperature threshold is related to the soil texture, organic matter content and ice content of the frozen soil response type zone, and the stable temperature threshold is different for different frozen soil response type zones.

9. A runoff simulation system driven by soil freeze-thaw state in a cold region permafrost watershed, characterized in that, The system is used to implement the runoff simulation method driven by the soil freeze-thaw state in cold-region permafrost watersheds as described in any one of claims 1 to 8, and the system comprises: The sample set construction module is used to acquire multi-source remote sensing images, digital elevation models and permafrost field survey samples of the target watershed, and construct permafrost response type datasets. The segmentation model training module, connected to the sample set construction module, is used to train a semantic segmentation model based on deep learning, taking multi-source remote sensing images and digital elevation models as inputs and permafrost response types as outputs. The permafrost type differentiation module is connected to the segmentation model training module and is used to extrapolate the target watershed using the trained semantic segmentation model to generate a permafrost response type partition map. The infiltration capacity coupling module, connected to the permafrost type differentiation module, is used to determine the initial temperature threshold that causes the runoff coefficient to rise and the stable temperature threshold that prevents the runoff coefficient from continuing to decrease in each permafrost response type zone based on historical temperature and runoff observation data of the target watershed; and to establish the temperature-infiltration capacity relationship curves corresponding to each permafrost response type zone. The watershed discretization module, connected to the infiltration capacity coupling module, is used to discretize the target watershed into calculation units according to the frozen soil response type zoning map, and assign a corresponding temperature-infiltration capacity relationship curve to each calculation unit. The runoff calculation module, connected to the watershed discretization module, is used to obtain the real-time temperature of each calculation unit according to the time step, determine the infiltration capacity of the target time period according to the temperature-infiltration capacity relationship curve corresponding to the calculation unit, and calculate the surface runoff of each calculation unit based on the infiltration capacity. The runoff heat map output module is connected to the runoff calculation module and is used to draw a runoff heat map of the target watershed based on the surface runoff of each calculation unit.

Images