A river dynamic ice condition extraction method combining spectral information and temperature time series

Abstract

This invention discloses a method for extracting dynamic river ice conditions by combining spectral information and temperature time series. The method includes: delineating the river channel range based on multi-source imagery, utilizing spectral indices and topographic features, and combining a random forest model with the SNIC segmentation method; within the extracted river mask, retrieving surface temperature using thermal infrared bands, employing a statistical single-window model and harmonic analysis, generating a daily-scale LST time series, and determining the freezing and thawing periods with 0°C as a threshold to calculate the river ice duration; during the river ice period, identifying ice, snow, water, and cloud types using multispectral features and an optimized random forest classifier; and using a temperature-constrained ice classification and denoising model, integrating LST constraints and classification results to interpolate cloud pollution pixels, generating river ice and snow cover products with a spatial resolution of 30 meters. This invention achieves continuous time series river ice monitoring, significantly improving the spatiotemporal characterization capability of dynamic river ice changes.

Description

Technical Field

[0001] This invention relates to a method for extracting dynamic ice conditions in rivers, and more particularly to a method for extracting dynamic ice conditions in rivers that combines spectral information and temperature time series. Background Technology

[0002] River ice phenomena encompass the entire process from formation, transport, deposition to melting. During this period, remote sensing technology has demonstrated its ability to monitor aspects such as the duration of river ice, the dates of freezing and thawing, the type of river ice, the area and distribution characteristics of the ice cover, the thickness of the river ice, and ice jam floods.

[0003] Optical satellites (such as MODIS, Landsat, and the Sentinel-2 series) are equipped with multispectral sensors covering the visible to near-infrared bands. These sensors offer wide imaging ranges and short revisit times, making them widely used for snow and ice monitoring. Different types of river ice cover (such as ice, snow-covered ice, broken ice, and dark ice) exhibit different reflection, absorption, and transmission characteristics to electromagnetic waves, resulting in significant spectral features. River ice identification methods based on spectral characteristics can be categorized into four types: 1) Visual interpretation of river ice distribution patterns using the visible and panchromatic bands of optical sensors such as MODIS, Landsat, and AVHRR. However, this interpretation process is complex and difficult to automate; 2) Ice / snow indices rely on the reflectivity difference between the green and near-infrared bands in the visible spectrum. These indices, including the Normalized Difference in Snow Cover Index (NDSI) and the Relative River Ice Difference Index (RDRI), are used to extract river ice extent and freezing duration. Their calculation is simple and easily implemented over long time series and large scales. However, the threshold range for specific indices is difficult to select, and undifferentiated index thresholds may lead to errors in river ice extraction; 3) Machine learning methods (such as artificial neural networks) can effectively fuse spectral and texture features to achieve automatic feature extraction that is beneficial for river ice identification, which helps to achieve fine classification in the process of river ice formation and melting. However, due to limitations in sample size and computing power, such methods are difficult to apply to global and interdecadal river ice research; 4) Based on the quality control bands of the MODIS and Landsat datasets, cloud and shadow mask classification results (Fmask) are used to classify pixels as snow / ice. In the dynamic analysis of river ice changes, river mask pixels marked as snow / ice can be directly used as the river ice coverage area. This method is convenient for long-term rapid monitoring of river ice due to its time-saving efficiency and high spatial resolution, and has been adopted by many recent studies. However, "dark ice" existing during the melting period may be unintentionally ignored, leading to inaccurate results.

[0004] Despite these advances, current research still faces limitations, leading to uncertainties and insufficient spatiotemporal resolution in river ice monitoring, thus restricting our understanding of the driving mechanisms of river ice formation and melting at the watershed scale. For example, optical satellite imagery such as Landsat is constrained by weather conditions and cannot meet the temporal resolution required for effective river ice monitoring. River ice conditions can change significantly within a single day, thus requiring diurnal observation data. Furthermore, spatial datasets of river freezing phenology remain insufficient in terms of continuity and coverage, hindering quantitative analysis of the relationship between channel morphology and ice dynamics, thereby limiting our understanding of how channel geometry regulates river freezing processes. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to propose a method for extracting dynamic ice conditions of rivers by combining spectral information and temperature time series. Based on machine learning, an analysis framework is constructed that integrates temperature time series and ice and snow spectral characteristics to achieve long-term, high-spatiotemporal continuous monitoring of river ice periods with a spatial resolution of 30 meters and daily time intervals.

[0006] Technical solution: This invention includes the following steps:

[0007] S1. Delineation of river channel: Based on multi-source imagery, using spectral indices and topographic features, combined with a random forest model and the SNIC segmentation method, the river channel is delineated.

[0008] S2. Generation of River Ice Phenology Dataset: Within the extracted river mask, the surface temperature is retrieved by thermal infrared band, and the statistical single-window model and harmonic analysis method are used to generate a daily-scale LST time series. The freezing and melting periods are determined with 0°C as the threshold, and the duration of river ice is calculated.

[0009] S3. River Ice Type Identification: During the river ice period, the types of ice, snow, water, and clouds are identified using multispectral features and an optimized random forest classifier.

[0010] S4. River Ice Dataset Generation: Using a temperature-constrained ice classification and denoising model, and combining LST constraints and classification results, cloud-polluted pixels are interpolated to generate river ice and snow cover products.

[0011] S1 specifically includes:

[0012] Optical images of Landsat 8 and Sentinel-2, as well as global DEM data, were selected from non-winter months.

[0013] Satellite imagery with cloud cover below 20% was selected, and cloud masking was applied to both sets of imagery.

[0014] Median composite method was used to generate non-winter river channel images;

[0015] Calculate the improved spectral index and extract terrain features;

[0016] The samples were divided according to a set ratio, and the river water area was extracted using the RF model.

[0017] The non-winter river channel imagery was visually interpreted to label multiple river pixels and multiple non-river pixels within the study area.

[0018] The spectral indices include: Normalized Difference Water Index (MNDWI), Normalized Difference Vegetation Index (NDVI), Enhanced Difference Vegetation Index (EVI), and Total Greenness of Tassels (TCGW), calculated using the following formulas:

[0019]

[0020]

[0021]

[0022]

[0023] Wherein, B, G, R, NIR, SWIR1 and SWIR2 represent the reflectance of blue, green, red, near-infrared and two short-wave infrared bands, respectively; C1 and C2 are aerosol resistance coefficients; and L is the canopy background adjustment term.

[0024] The samples were segmented using the object-oriented SNIC segmentation method on NDVI imagery to segment the river region.

[0025] The formula for calculating the duration of the river ice is:

[0026]

[0027] in, Represents the total number of days in the winter season studied; This represents the latest date on which the surface temperature of cell i is greater than or equal to a threshold during the cooling phase, indicating the start date of freezing. s ; This represents the earliest date during the warming phase of the following year when the surface temperature is less than or equal to the threshold, indicating the end date of the freezing period. e .

[0028] S3 specifically includes:

[0029] Sentinel-2 images were registered with Landsat images using half-pixel offset and then resampled to a spatial resolution of 30 meters.

[0030] Use annual river range mask y As a spatial constraint, median synthesis was used to reconstruct the image set at multi-day intervals;

[0031] Based on the above reconstructed image set, an RGB composite image is generated using the visible light band;

[0032] We used a labeled dataset as the training set to train and test a parameter-optimized random forest classifier RF-O, and tuned key hyperparameters: the number of trees nt, the minimum number of samples required to split internal nodes mss, and the maximum depth of the trees mdt.

[0033] The image set includes 3–5 original spectral bands, modified Normalized Difference Water Index (MNDWI), Normalized Difference Snow Index (NDSI), and Normalized Difference Glacier Index (NDGI). The formulas for calculating the Normalized Difference Snow Index (NDSI) and the Normalized Difference Glacier Index (NDGI) are as follows:

[0034]

[0035]

[0036] In this context, G, R, SWIR1, and SWIR2 represent the reconstructed bands, respectively.

[0037] The specific process of the temperature-constrained ice classification and denoising model is as follows:

[0038] 1) Time series pixel labeling: For any pixel r(t) within the river channel, surface temperature is used as a constraint condition;

[0039] 2) Water body weight allocation rule: If r(t) is classified as a water body, and date t falls within the freezing / thawing period [doy s doy e If the water body is within the specified range, then the classification will be retained as a water body.

[0040] 3) Missing value imputation: If r(t) is classified as ice or snow, the classification remains unchanged.

[0041] When r(t) is missing due to image unavailability, and t ∈ [doy s doy e When [the following rules apply], the following rules shall apply:

[0042] If r(t−1) = r(t+1), then assign the value r(t) = r(t−1);

[0043] If r(t−1) ≠ r(t+1) or r(t-1) is missing, then assign r(t) = r(t−1), assuming it has short-term persistence.

[0044] Beneficial effects: This invention has the following advantages:

[0045] (1) Improve the spatiotemporal resolution of river ice monitoring: By using multi-source remote sensing data fusion and introducing surface temperature time series interpolation and harmonic analysis technology, the river ice monitoring has been expanded from single-phase images to continuous time series. The product has a spatial resolution of 30 meters and a temporal resolution of daily, which greatly improves the spatiotemporal characterization of river ice dynamic changes.

[0046] (2) Effective removal of cloud and rain weather effects: By combining optical images with thermal infrared data and using median synthesis, cloud masking and temperature-constrained classification denoising algorithms, the impact of adverse meteorological conditions such as clouds, shadows and precipitation on image classification accuracy is significantly reduced, and stable and reliable river ice identification results are achieved.

[0047] (3) Quantify and refine the characteristics of ice and snow cover: A fine classification system for river ice based on surface temperature threshold and multispectral characteristics was constructed, which can distinguish various surface types such as ice, snow and open water, quantify the duration of river ice, spatial distribution of ice and snow and phenological change process, and provide high-precision data support for the analysis of river ice evolution law and climate response research. Attached Figure Description

[0048] Figure 1 This is a flowchart of the present invention;

[0049] Figure 2 This is a graph showing the annual river water area results extracted in this embodiment of the invention;

[0050] Figure 3 This is a graph showing the duration of river ice in an embodiment of the present invention;

[0051] Figure 4 (a) Figure 4 (b) and Figure 4 (c) Parameter optimization plots for nt, mdt, and mss, respectively;

[0052] Figure 5 This is a graph showing the quantitative percentage results of river ice and snow cover types in an embodiment of the present invention;

[0053] Figure 6 This is a comparison chart of the duration of river ice and hydrological station records in an embodiment of the present invention. Detailed Implementation

[0054] The invention will now be further described with reference to the accompanying drawings.

[0055] Example 1

[0056] like Figure 1As shown in the figure, the river dynamic ice condition extraction method combining spectral information and temperature time series in this embodiment achieves high-precision automatic extraction and time-series consistency of river ice information, providing technical support for ice condition monitoring and river ice phenology research. Specifically, it includes the following steps:

[0057] S1. River Channel Delineation: Based on multi-source imagery, utilizing spectral indices and topographic features, and combining a random forest model with the SNIC segmentation method, the river channel is accurately delineated. Specifically, this includes:

[0058] S11. Select Landsat 8 Collection 2 and Sentinel-1 / 2 images from the Google Earth Engine (GEE) platform as data sources. Select Landsat 8 and Sentinel-2 optical images and global DEM data from May to July (August and September are excluded because they are at the peak of the flood season, during which the water area is significantly affected by floods caused by precipitation) in non-winter months.

[0059] S12. Select satellite images with cloud cover below 20% and perform cloud masking on the two sets of images.

[0060] S13. High-quality river extent imagery for non-winter periods was generated using median composite method. Through visual interpretation, 200 river pixels and 200 non-river pixels were labeled within the study area. Non-river pixels included features such as bridges, roads, farmland, bare land, and buildings.

[0061] S14. Calculate the improved Normalized Difference Water Index (MNDWI), Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Total Green Cap (TCGW) spectral indices, as shown in the formula. Furthermore, terrain features, including slope (SLO) and elevation (Ele), were extracted.

[0062]

[0063]

[0064]

[0065]

[0066] Wherein, B, G, R, NIR, SWIR1, and SWIR2 represent the reflectance of the blue, green, red, near-infrared, and two short-wave infrared bands, respectively. C1 and C2 are aerosol resistance coefficients (usually C1=6.0, C2=7.5), and L is the canopy background adjustment term (usually set to 1).

[0067] S15. The samples were split in a 7:3 ratio, and a Random Forest (RF) model was used to train and extract river water areas using the following features. Simultaneously, the object-oriented SNIC segmentation method was applied to segment the river areas on the NDVI image, with segmentation parameters compactness, connectivity, and neighborhood size set to 0.1, 8, and 250, respectively. This method maximized the accuracy of river channel delineation. Based on the segmentation results, the RF model was further used to extract river water areas using specified features.

[0068] S2. Generation of River Ice Phenology Dataset: Within the extracted river mask, the surface temperature is retrieved using the Landsat thermal infrared band. The statistical single-window model and harmonic analysis method are used to generate a daily-scale LST time series. The freezing and melting periods are determined with 0°C as the threshold, and the duration of river ice is calculated.

[0069] A refined river ice region was constrained using extracted water body masks. The period from October 1st to May 1st of the following year was defined as the winter freezing and melting period. Furthermore, the duration of river ice and ice cover type during this period were extracted using brightness temperature and optical images, respectively, thus generating a river ice cover category dataset with a spatial resolution of 30 meters. The river freezing process is mainly controlled by air temperature; it begins when the temperature drops below 0°C, and surface ice gradually forms. When the temperature rises above 0°C, the ice begins to melt.

[0070] To capture this process, this invention leverages the potential of Landsat series data to provide land surface temperature (LST) data. Landsat's thermal infrared (TIR) ​​bands can calculate the top of the atmosphere (TOA) brightness temperature. These TOA values ​​are resampled to a 30-meter spatial resolution and organized into an image ensemble on the GEE platform for efficient time-series analysis. Total atmospheric water vapor content (TCWV) values ​​from Landsat satellite scenes are obtained from NCEP / NCAR reanalysis data and integrated into the GEE. Surface emissivity (ε) is derived from the ASTER GEDv3 dataset and adjusted to match the spectral response of the Landsat TIR bands. To account for annual and interannual variations in surface emissivity, adjustments are made for snow cover variations using NDSI derived from Landsat, combined with emissivity and average NDSI products from ASTER GEDv3. LST retrieval employs a statistical single-window (SMW) model, a single-channel linear regression model that considers atmospheric water vapor content and surface emissivity. The algorithm coefficients are generated and calibrated using the RTTOV model, based on a large atmospheric database containing various surface and atmospheric conditions.

[0071] For the river mask y For any Landsat pixel i within the imagery, the LST time series from October 1st to May 1st of the following year was retrieved using LST inversion results integrated into the GEE platform. Due to the spatiotemporal variability of image acquisition, the number of available LST observations varies between pixels, increasing the difficulty of direct comparison. To address this issue, linear interpolation was applied to resample the LST time series to a daily scale. Subsequently, the interpolation results were smoothed using the Temporal Harmonic Analysis (HANTS) algorithm, which minimizes the impact of outliers while preserving the overall seasonal trend. Finally, a pixel-level daily land surface temperature time series (LST) from October 1st to May 1st of the following year was obtained. i Using T=0°C as the threshold condition, according to the formula... Calculate the duration of the ice age (D) i ):

[0072]

[0073] in, Represents the total number of days in the winter season studied; This represents the latest date on which the surface temperature of cell i is greater than or equal to a threshold during the cooling phase, indicating the start date of freezing. s ; This represents the earliest date during the warming phase of the following year when the surface temperature is less than or equal to the threshold, indicating the end date of the freezing period. e .

[0074] S3. River Ice Type Identification: During the river ice period, the types of ice, snow, water, and clouds are identified using multispectral features and an optimized random forest classifier.

[0075] During river ice cover periods, climate variability leads to a mixture of land cover types, including ice, snow, and open water. To ensure comprehensive classification within the river extent mask, five land cover categories were defined: water, snow, ice, thin clouds, and thick clouds. First, Sentinel-2 imagery was registered with Landsat imagery by applying half-pixel migration and resampled to 30-meter spatial resolution. An annual river extent mask was used. y As a spatial constraint, median composite was used to reconstruct the image set at 10-day intervals. It includes 3–5 original spectral bands—namely blue (B), green (G), red (R), near-infrared (NIR), and shortwave infrared (SWIR) bands—and three indices commonly used to distinguish between water bodies and ice: the Modified Normalized Water Index (MNDWI), the Normalized Snow Index (NDSI) (formula...). ) and Normalized Difference Glacier Index (NDGI) (formula) ).

[0076]

[0077]

[0078] In this context, G, R, SWIR1, and SWIR2 represent the reconstructed bands, respectively.

[0079] RGB composite images were generated using the visible light band based on a 10-day image set (from October 1st to May 1st of the following year). A total of 350 reference points were selected along the river through visual interpretation. For each 10-day image from 2019, these points were manually labeled with their target land surface categories. A total of 2097 samples were labeled for each of the following categories: water, snow, ice, and thin clouds. For thick clouds, 65 samples were labeled. Using this labeled dataset as the training set, a parameter-optimized Random Forest classifier (RF-O) was trained and tested. Three key hyperparameters—the number of trees (nt), the minimum number of samples required to split internal nodes (mss), and the maximum tree depth (mdt)—were tuned to ensure optimal performance of the RF classifier.

[0080] S4. River Ice Dataset Generation: Using a temperature-constrained ice classification and denoising model, and combining LST constraints and classification results, cloud-polluted pixels are interpolated to generate river ice and snow cover products with a spatial resolution of 30 meters.

[0081] Further post-processing is needed in the river ice type identification dataset to generate a standardized, complete, and analyzable dataset of Songhua River ice and snow cover. This embodiment proposes an ensemble method called Tem-IcD (Temperature-Constrained Ice Classification and Denoising Model), which combines the surface temperature dataset and classification results to generate river ice and snow cover products with a spatial resolution of 30 meters and a temporal resolution of 10 days.

[0082] The Tem-IcD model operates as follows: 1) Time-series cell labeling: For any cell r(t) within the river channel, where t ranges from October 1st of each year to May 1st of the following year (in 10-day intervals, labeled t=1,2,...,22), surface temperature is used as a constraint. When LST is below 0°C, the cell may be ice or snow. However, the possibility of open water bodies during freezing or thawing periods is not excluded. 2) Water body weight allocation rules: If r(t) is classified as a water body, and date t falls within the freezing / thawing period... s doy e If r(t) is classified as water, then the classification is retained. This takes into account the possibility of melting during warm periods. 3) Missing value imputation: If r(t) is classified as ice or snow, the classification remains unchanged. When r(t) is missing due to image unavailability, and t ∈ [doy s doy e When [the following rules apply], the following rules shall apply:

[0083] If r(t−1) = r(t+1), then assign r(t) = r(t−1).

[0084] If r(t−1) ≠ r(t+1) or r(t-1) is missing, then r(t) = r(t−1) is assigned, assuming short-term persistence. After completing the above steps, each pixel is assigned a final classification value for each of the 22 time intervals in each winter. This process produces a consistent, smooth, and temporally complete SHR-IS (Songhua River Ice and Snow) dataset. The final product supports further spatiotemporal analysis of river ice dynamics and phenology.

[0085] This embodiment achieves high spatiotemporal resolution monitoring of river ice processes by integrating Landsat 8, Sentinel-1 / 2, multi-source remote sensing imagery, and land surface temperature (LST) time series. Improving the spatiotemporal resolution of river ice monitoring and compensating for the lack of dense area-scale observation data enables the capture of dynamic changes in the formation and dissipation of river ice. By incorporating time-series land surface temperature data and performing cloud processing during ice condition identification, classification errors caused by weather conditions such as clouds, rain, and snow are effectively removed, resulting in more complete and continuous temporal characteristics of river ice. By constructing an improved random forest classifier (RF-O) and introducing a temperature-constrained ice classification and denoising model (Tem-IcD), refined differentiation and time series reconstruction of ice and snow types (bare ice, snow cover, water bodies, etc.) are achieved, significantly improving the quantitative accuracy of ice and snow cover information and facilitating subsequent assessments of the watershed's ecosystem health and carbon cycle.

[0086] Example 2

[0087] This embodiment takes the Neijiang section of the Songhua River in Songyuan City as an example, and constructs a framework for river ice monitoring that combines surface temperature time series and optical reflectance characteristics to extract river ice information for five winters from 2018 to 2023. Specifically, it includes the following steps:

[0088] (1) Delineation of river channel boundaries: Based on the GEE platform, Landsat 8 Collection 2 and Sentinel-1 / 2 multi-source remote sensing images were selected, combined with global DEM data. To avoid interference from floods in water body identification, non-flood season images from May to July were selected as river channel extraction samples, and scene images with cloud cover below 20% were selected. High-quality annual water body images were generated using cloud masking and median synthesis. River and non-river samples were selected within the Songyuan City section of the river through visual interpretation, and the improved normalized water index (MNDWI), normalized vegetation index (NDVI), enhanced vegetation index (EVI), and tassel greenness (TCGW) were calculated. Slope and elevation features were extracted at the same time. The random forest (RF) model was used for river water body classification, and the spatial boundary identification accuracy was improved by combining SNIC object-oriented segmentation (compactness 0.1, neighborhood 250), thereby generating accurate river channel masks and calculating the interannual variation of river water area, such as Figure 2 As shown.

[0089] (2) Extraction of River Ice Phenological Information: The research scope was defined by using the extracted riverbed mask, and the period from October 1st to May 1st of the following year was defined as the freezing-thawing period. Land surface temperature (LST) was retrieved based on the Landsat 8 thermal infrared band, and combined with NCEP / NCAR reanalysis water vapor content (TCWV) and ASTER GEDv3 surface emissivity (ε), a statistical single-window model (SMW) was used for LST retrieval. All images were resampled to 30-meter resolution and organized into a time series set. To ensure temporal continuity, the LST time series was linearly interpolated and smoothed by harmonic analysis (HANTS) to obtain a continuous daily-scale LST sequence. The freezing start time (doy) was calculated with 0°C as the threshold. s ) and the end of ablation (doy e The time frame is used to determine the duration of the ice period for each pixel, thus enabling the quantitative extraction of river ice phenology. Figure 3 As shown.

[0090] (3) Snow and Ice Type Classification and Dataset Construction: During the freezing period (October to May of the following year), a multi-temporal dataset with 10-day intervals was constructed based on the fused images of Sentinel-2 and Landsat (resampled to 30 meters). Blue, green, red, near-infrared (NIR), and shortwave infrared (SWIR) bands were selected, and spectral indices such as MNDWI, NDSI, and NDGI were calculated to enhance the difference between water bodies and snow and ice. Combined with the annual river channel mask, a 10-day synthetic image sequence was generated using the median synthesis method. Reference points were selected through visual interpretation, and five types of land surfaces—water, snow, ice, thin clouds, and thick clouds—were labeled to construct a sample set and train a parameter-optimized random forest classifier (RF-O). The optimized tree depth ( Figure 4 (a) and minimum sample parameter ( Figure 4 (b) and the maximum number of leaf nodes ( Figure 4 (c)) ensures that the classification results are consistent and highly accurate across different years.

[0091] To mitigate classification discontinuities caused by cloud cover and missing observations, a temperature-constrained ice classification and denoising model (Tem-IcD) is introduced. This model uses the LST time series as a temperature constraint, correcting the river channel pixel r(t) at every 10-day interval: when the LST is below 0°C, the pixel may be ice or snow; if it is in a frozen state but classified as water, this state is preserved to reflect the melting process; if the image is missing, continuous reconstruction is achieved through temporal proximity interpolation. Ultimately, a temporally consistent and spatially continuous dataset of river ice and snow cover is obtained, and river ice cover changes are calculated, such as... Figure 5 As shown. Finally, the accuracy of the extracted results was verified using observation data from river ice phenology stations recorded at hydrological stations, such as... Figure 6As shown, this framework enables high-precision, automated, and temporal extraction of river ice information, effectively overcoming the interference of cloud and rain weather on optical imagery, significantly improving the spatiotemporal resolution of river ice monitoring, and providing data support for the analysis of ice condition evolution and climate response in the Songhua River Basin.

Claims

1. A method for extracting dynamic ice conditions in rivers by combining spectral information and temperature time series, characterized in that, Includes the following steps: S1. Delineation of river channel: Based on multi-source imagery, using spectral indices and topographic features, combined with a random forest model and the SNIC segmentation method, the river channel is delineated. S2. Generation of River Ice Phenology Dataset: Within the extracted river mask, surface temperature is retrieved using thermal infrared bands. A statistical single-window model and harmonic analysis are employed, combined with the generation of a diurnal LST time series. Freezing and thawing periods are determined using 0°C as a threshold, and the duration of river ice is calculated. The formula for calculating the duration of river ice is as follows: in, Represents the total number of days in the winter season studied; This represents the latest date on which the surface temperature of cell i is greater than or equal to a threshold during the cooling phase, indicating the start date of freezing. s ; This represents the earliest date during the warming phase of the following year when the surface temperature is less than or equal to the threshold, indicating the end date of the freezing period. e ; S3. River Ice Type Identification: During the river ice period, the types of ice, snow, water, and clouds are identified using multispectral features and an optimized random forest classifier. S4. River Ice Dataset Generation: Using a temperature-constrained ice classification and denoising model, and combining LST constraints and classification results, interpolation of cloud-polluted pixels is performed to generate river ice and snow cover products; the temperature-constrained ice classification and denoising model includes: 1) Time series pixel labeling: For any pixel r(t) within the river channel, surface temperature is used as a constraint condition; 2) Water body weight allocation rule: If r(t) is classified as a water body, and date t falls within the freezing / thawing period [doy s doy e If the water body is within the specified range, then the classification will be retained as a water body. 3) Missing value imputation: If r(t) is classified as ice or snow, the classification remains unchanged.

2. The method for extracting river dynamic ice conditions by combining spectral information and temperature time series according to claim 1, characterized in that, S1 specifically includes: Optical images of Landsat 8 and Sentinel-2, as well as global DEM data, were selected from non-winter months. Satellite imagery with cloud cover below 20% was selected, and cloud masking was applied to both sets of imagery. Median composite method was used to generate non-winter river channel images; Calculate the improved spectral index and extract terrain features; The samples were divided according to a set ratio, and the river water area was extracted using the RF model.

3. The method for extracting river dynamic ice conditions by combining spectral information and temperature time series according to claim 2, characterized in that, The non-winter river channel imagery was visually interpreted to label multiple river pixels and multiple non-river pixels within the study area.

4. The method for extracting river dynamic ice conditions by combining spectral information and temperature time series according to claim 2, characterized in that, The spectral indices include: Normalized Difference Water Index (MNDWI), Normalized Difference Vegetation Index (NDVI), Enhanced Difference Vegetation Index (EVI), and Total Greenness of Tassels (TCGW), calculated using the following formulas: Wherein, B, G, R, NIR, SWIR1 and SWIR2 represent the reflectance of blue, green, red, near-infrared and two short-wave infrared bands, respectively; C1 and C2 are aerosol resistance coefficients; and L is the canopy background adjustment term.

5. The method for extracting river dynamic ice conditions by combining spectral information and temperature time series according to claim 2, characterized in that, The samples were segmented using the object-oriented SNIC segmentation method on NDVI imagery to segment the river region.

6. The method for extracting river dynamic ice conditions by combining spectral information and temperature time series according to claim 1, characterized in that, S3 specifically includes: Sentinel-2 images were registered with Landsat images using half-pixel offset and then resampled to a spatial resolution of 30 meters. Use annual river range mask y As a spatial constraint, median synthesis was used to reconstruct the image set at multi-day intervals; Based on the above reconstructed image set, an RGB composite image is generated using the visible light band; We used a labeled dataset as the training set to train and test a parameter-optimized random forest classifier RF-O, and tuned key hyperparameters: the number of trees nt, the minimum number of samples required to split internal nodes mss, and the maximum depth of the trees mdt.

7. The method for extracting river dynamic ice conditions by combining spectral information and temperature time series according to claim 6, characterized in that, The image set includes 3–5 original spectral bands, modified Normalized Difference Water Index (MNDWI), Normalized Difference Snow Index (NDSI), and Normalized Difference Glacier Index (NDGI). The formulas for calculating the Normalized Difference Snow Index (NDSI) and the Normalized Difference Glacier Index (NDGI) are as follows: Wherein, G, R, and SWIR1 represent the reflectivity of the reconstructed bands, respectively.

8. The method for extracting river dynamic ice conditions by combining spectral information and temperature time series according to claim 1, characterized in that, When r(t) is missing due to image unavailability, and t ∈ [doy s doy e When [the number of cases is], the following rules apply: If r(t-1) = r(t+1), then assign the value r(t) = r(t-1); If r(t-1) ≠ r(t+1), then assign r(t) = r(t-1), assuming it has short-term persistence.

Images