A lake water volume monitoring and prediction method and system based on multi-source remote sensing and machine learning

Abstract

The application discloses a kind of based on multi-source remote sensing and machine learning's lake water quantity monitoring and prediction method and system, comprising: acquisition multi-source remote sensing, meteorological climate data, after quality control repair and standardization, by the volume method of circular truncated cone, water level and area data are fused to construct lake historical month scale water quantity sequence;Again, meteorological driving factor is screened and interannual impact feature is constructed, and enhanced feature sequence is fused;Subsequently, a training cascade topology structure model is constructed, and the enhanced feature sequence is input, the long time sequence feature is extracted by Transform encoder Output benchmark prediction value, combined with meta-feature extraction, multi-branch correction fusion is completed Water quantity prediction and confidence interval quantization;Finally, after the deviation correction and ensemble average of CMIP6 data are input into the model, the data splicing jump is eliminated by soft landing smoothing, and the future multi-scenario lake water quantity evolution curve and uncertainty range are output.The application improves water quantity monitoring precision and prediction robustness.

Description

Technical Field

[0001] This invention relates to the field of remote sensing hydrology and climate change response technology, specifically to a method and system for monitoring and predicting lake water volume based on multi-source remote sensing and machine learning. Background Technology

[0002] As a crucial component of the terrestrial water cycle, lakes play a vital indicative role in regional climate response, water resource balance, and ecological environment maintenance through changes in their water volume. Particularly in the Tianshan region—a core water conservation area in the arid Central Asian region—lake hydrological dynamics not only affect local ecological security and agricultural and pastoral production but also pose significant challenges to transboundary water resource management and climate adaptation strategies. In recent years, under the combined influence of global climate change and human activities, lakes in this region have exhibited significant spatial heterogeneity and temporal volatility, with some lakes expanding and others shrinking, necessitating systematic monitoring and forecasting research.

[0003] Currently, monitoring of lake water volume changes mainly relies on traditional ground observation and remote sensing technologies. Among remote sensing methods, optical imagery (such as the Landsat series) and satellite altimetry data (such as Hydroweb) have been widely used to extract lake area and water level data. However, existing research still has significant limitations: First, due to cloud cover, sensor malfunctions (such as Landsat 7SLC-off), and limitations in data spatiotemporal resolution, it is difficult to obtain long-term, continuous, and highly consistent lake water level and area sequences. Second, most studies focus on single analyses of area or water level, lacking a systematic method to integrate the two for overall water volume estimation, resulting in insufficient accuracy in water volume estimation. Third, in terms of analyzing driving mechanisms, statistical correlation analysis is often relied upon, failing to deeply reveal the nonlinear relationship and dominant control mechanism between meteorological factors and lake water volume. Fourth, existing forecasting studies are mostly based on traditional hydrological models or linear extrapolation, making it difficult to handle the complex nonlinear processes in the climate-lake response, and lacking integrated forecasts and uncertainty assessments under multiple models and scenarios.

[0004] In complex, high-altitude regions like the Tianshan Mountains, existing lake water volume studies mostly focus on local water bodies within a single country, lacking comprehensive comparisons and mechanism exploration of the entire transboundary lake system. Because transboundary lake basins span different countries, significant technical barriers exist: firstly, there are severe data barriers, as the distribution of transboundary surface hydrological and meteorological stations is extremely uneven, and data sharing is limited, making it difficult to construct and calibrate traditional distributed hydrophysical models (such as SWAT models) that rely on a large number of measured geographical and soil parameters in transboundary basins; secondly, there are spatial heterogeneity barriers, as the microclimate and surface response mechanisms in transboundary basins are extremely complex, making it difficult for traditional statistical models to have cross-regional generalization capabilities.

[0005] Furthermore, existing research attempting to integrate multi-source meteorological reanalysis data (such as ERA5) with future multi-model climate scenarios (CMIP6) for lake water volume prediction faces extremely high challenges in model coupling and temporal series integration. These challenges primarily manifest in: difficulties in matching data distribution domain shifts and systematic errors; the CMIP6 climate model output typically has coarse spatial resolution and inherent systematic biases compared to real historical meteorological observation data (such as ERA5); and the direct input of climate model outputs as new parameters into the model, leading to predictions that significantly deviate from physical reality. The nonlinear impacts of greenhouse gas emission scenarios are difficult to capture. Different greenhouse gas emission scenarios (such as SSP1-2.6 and SSP5-8.5) are not simply adjustments to scalar parameters but represent a trend towards more frequent and highly destructive extreme climate events in the future. Traditional machine learning models (such as artificial neural networks) are prone to overfitting to historical stationary sequences. When faced with rapidly changing precipitation and temperature fluctuations under high-emission scenarios, they are highly susceptible to prediction distortion, collapse of generalization ability, and even outputting non-physical negative values. The problem of trend angles across boundaries arises when the model switches from inputting historical actual data to inputting future SSP climate scenario data. The change in the underlying distribution of the data source inevitably leads to a cliff-like jump in the prediction curve at the boundary and a trend angle that does not conform to the laws of hydrological evolution.

[0006] In summary, existing technologies for monitoring and predicting lake water volume in the Tianshan region suffer from problems such as poor data continuity, unclear driving mechanisms, weak model adaptability, and insufficient multi-scenario prediction capabilities. Therefore, there is an urgent need to develop a dynamic assessment method for lake water volume that can integrate multi-source remote sensing data, analyze multi-factor driving mechanisms, and possess multi-scenario prediction capabilities, in order to improve the scientific rigor, foresight, and adaptability of regional water resource management. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for monitoring and predicting lake water volume based on multi-source remote sensing and machine learning. This method effectively solves the problems of insufficient characterization of time-series dependence and limited generalization ability of single models in traditional lake water volume monitoring and prediction methods. It realizes long-term time-series reconstruction of lake water volume in arid areas and stable prediction under multiple climate scenarios, and can provide reliable technical support for regional water resource allocation, climate change risk assessment and lake ecological protection.

[0008] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for monitoring and predicting lake water volume based on multi-source remote sensing and machine learning, the method comprising:

[0010] S1. Collect historical monthly water level data, historical monthly lake area data, historical ERA5 atmospheric reanalysis data, and historical simulation data and future prediction data of CMIP6 climate model. After data quality control, repair and standardization, the historical monthly water level data and historical monthly lake area data are fused using the frustum volume method to calculate the monthly water volume change and construct the historical monthly water volume change sequence of the lake to be monitored.

[0011] S2 introduces various monthly-scale variables from historical ERA5 atmospheric reanalysis data as candidate driving factors, and selects meteorological driving factors that have a driving relationship with the water volume change of the lake to be monitored. The fluctuation of each meteorological driving factor from the global mean is calculated to construct the interannual impact characteristics. The standardized meteorological driving factor sequence is used as the independent variable, the historical monthly-scale water volume change sequence of the lake to be monitored is used as the dependent variable, and a training sample set is constructed according to the sequence sliding window.

[0012] S3. Construct and train a cascaded topology lake water volume monitoring model. This model integrates interannual impact characteristics into a standardized meteorological driving factor sequence, using this as the input to form an enhanced feature sequence. A Transformer encoder extracts long-term time-dependent features, outputting a baseline predicted value reflecting the basic evolution trend of lake water volume and simultaneously acquiring the hidden layer state at the end of the sequence. The hidden layer state is concatenated with the meteorological features of the current time step to construct and output a dimensionality-reduced meta-feature vector. Parallel XGBoost and LightGBM regressors receive the meta-feature vector and predict algebraic residuals, respectively. These residuals are then weighted and integrated according to preset jump and smoothing preference weights, outputting a biased value. The method involves: adjusting the fusion residual for difference compensation; using an independent XGBoost adjustment model to receive the meta-feature vector and predicting the fluctuation intensity factor reflecting the intensity of water volume fluctuations induced by extreme climate; using a LightGBM model configured with a quantile regression loss function to receive the meta-feature vector and predict the upper and lower bound residuals of the water volume distribution boundary, respectively, and introducing an interval coverage probability calibration mechanism to dynamically calculate the scaling factor based on the actual coverage effect of the validation set, and outputting the interval adjustment; adding and fusing the baseline predicted value and the fusion residual adjustment, and then performing a multiplicative scaling operation with the fluctuation intensity factor to obtain the predicted value of the target lake water volume; and simultaneously combining the interval adjustment to generate a statistically significant prediction confidence interval.

[0013] S4 uses historical simulation data from the CMIP6 climate model and historical ERA5 atmospheric reanalysis data from the same period to correct biases in the future prediction data of the CMIP6 climate model. After ensemble averaging of the corrected future meteorological data from all climate models under the same emission scenario, the data is input into the trained lake water monitoring model to obtain a preliminary monthly lake water prediction sequence and corresponding dynamic prediction interval. Through soft landing smoothing and cross-boundary moving average mechanisms, the abrupt transition problem between the historical actual water volume sequence and the soft landing smoothed monthly lake water prediction sequence is addressed. Finally, a smooth transition future multi-scenario lake water evolution curve and its uncertainty quantification range are output.

[0014] Step S1 further includes:

[0015] S11: Collect historical monthly water level data from Hydroweb satellite altimetry products, remove outliers through data quality control to ensure the continuity of the water level sequence, and generate historical monthly water level sequences for the lake to be monitored.

[0016] S12: Acquire Landsat 5 / 7 / 8 series remote sensing images, use interpolation to repair SLC fault stripes in Landsat 7, extract monthly lake area data using MNDWI index and Otsu adaptive thresholding method, and supplement missing data using interpolation.

[0017] S13 collects ERA5 atmospheric reanalysis data, historical simulation data and future forecast data of CMIP6 climate model, unifies the spatiotemporal resolution of the data and performs standardization processing.

[0018] S14. The historical monthly water level data and historical monthly lake area data were fused using the frustum volume method to calculate the monthly water volume change. :

[0019] ;

[0020] In the formula and The lake surface area for two consecutive months is shown. and By taking the water levels of two adjacent months as examples, a historical monthly water volume change sequence of the lake to be monitored is constructed.

[0021] The monthly water volume changes of the lakes to be monitored are standardized in a unified format, and the timestamp, lake name and data quality level of each data point are labeled to construct a historical monthly water volume change sequence of the lakes to be monitored.

[0022] Step S12, the process of extracting monthly lake area data using the MNDWI index and the Otsu adaptive thresholding method, includes the following steps:

[0023] Using the Google Earth Engine platform, an improved Normalized Difference Water Index (MNDWI) was calculated from preprocessed Landsat imagery.

[0024] ;

[0025] Green corresponds to the green light band of Landsat, and SWIR corresponds to the short-wave infrared band.

[0026] The Otsu adaptive threshold segmentation method is adopted to automatically determine the segmentation threshold by maximizing the inter-class variance between water bodies and non-water bodies, and to perform unsupervised classification of water bodies. Vector boundary extraction is performed on the classification results to remove small patches with an area smaller than the preset area threshold to eliminate noise interference, and the actual water surface area of ​​the lake is calculated every month.

[0027] The extracted monthly lake surface area is coupled with the water level data obtained from the Hydroweb platform for the same period. First, abnormal area values ​​that exceed the historical average or are greater than a preset percentage are verified for rationality by combining the water level data and meteorological conditions for the same period, and the abnormal values ​​are replaced by linear interpolation. Then, the consistency between the area and water level change trends is verified, and abnormal periods with contradictory water level and area change trends are removed and filled by linear interpolation. Finally, continuous historical monthly lake area data is formed and a correspondence is established with the historical monthly water level sequence.

[0028] Step S2 further includes:

[0029] Multiple monthly-scale variables, including 2m air temperature, total precipitation, 10m wind speed, and surface latent heat flux, were introduced as candidate driving factors. Historical ERA5 atmospheric reanalysis data were extracted as the historical actual observation data for each candidate driving factor. The extracted historical ERA5 atmospheric reanalysis data, monthly water level data, and monthly lake area data of the lake to be monitored were uniformly converted to monthly temporal resolution by linear interpolation. The candidate driving factors, water level, and area sequences were strictly aligned according to timestamps, and periods without synchronized data were removed. Finally, the Z-Score normalization method was used to eliminate the dimensional differences between variables, and a historical multi-source aligned dataset containing time stamps, lake identifiers, candidate driving factors, and corresponding water level and area labels was constructed.

[0030] A nonlinear mapping relationship between candidate driving factors and lake water volume changes was fitted using a historical multi-source aligned dataset;

[0031] Based on the principle of spatial heterogeneity of geographic detectors, the explanatory power and significance of each candidate driving factor are calculated. Meteorological driving factors that have a driving relationship with the water volume change of the lake to be monitored are selected, and an analytical framework for the driving relationship between meteorological driving factors and lake water volume change is constructed.

[0032] Furthermore, in step S3, the lake water volume monitoring model includes a main trunk prediction module, a meta-feature extraction module, a multi-branch correction and fusion module, and a cascade synthesis module;

[0033] The main prediction module adopts a Transformer encoder based on a multi-head self-attention mechanism. It receives the enhanced feature sequence as input, uses the self-attention mechanism to extract deep abstract features with long time-series dependencies, outputs a baseline prediction value that reflects the basic evolution trend of lake water volume, and simultaneously extracts the hidden layer state at the end of the sequence.

[0034] The meta-feature extraction module receives the hidden layer state, concatenates it with the meteorological features of the current time step, constructs and outputs the dimensionality-reduced meta-feature vector;

[0035] The multi-branch correction fusion module receives the meta-feature vector and assigns it to three parallel sub-network branches consisting of a two-branch residual fusion unit, an amplitude-sensitive prediction unit, and a confidence interval generation unit for independent inference and fine-tuning. Specifically, the two-branch residual fusion unit uses parallel XGBoost and LightGBM regressors to receive the meta-feature vector and predict algebraic residuals, then performs weighted integration according to preset jump and smoothing preference weights, outputting a fusion residual adjustment amount for bias compensation. The amplitude-sensitive prediction unit uses an independent XGBoost adjustment model to receive the meta-feature vector and predicts the fluctuation intensity factor reflecting the intensity of water volume fluctuations induced by extreme climate. The confidence interval generation unit uses a LightGBM model configured with a quantile regression loss function to receive the meta-feature vector, predicts the upper and lower bound residuals of the water volume distribution boundary, and introduces an interval coverage probability calibration mechanism to dynamically calculate the scaling factor based on the actual coverage effect of the validation set, outputting the interval adjustment amount.

[0036] The cascaded fusion module adds and fuses the baseline predicted value and the fusion residual adjustment, and then performs a multiplicative scaling operation with the fluctuation intensity factor to obtain the predicted value of the target lake water volume; and simultaneously combines the interval adjustment to generate a statistically significant prediction confidence interval.

[0037] Step S4 further includes:

[0038] For the future prediction data of the acquired CMIP6 climate model, the historical simulation data of the CMIP6 climate model is compared with the historical ERA5 atmospheric reanalysis data of the same period to extract the systematic bias of each climate model; the bias correction method is used to correct the future prediction data of the CMIP6 climate model using the systematic bias.

[0039] The ensemble averaging of future meteorological data from all climate models under the same emission scenario is used to generate a standardized sequence of future meteorological driving factors. This sequence is then input into a trained lake water volume monitoring model to output a preliminary monthly lake water volume prediction sequence and the corresponding dynamic prediction interval.

[0040] Extract the terminal true value of the historical actual water volume sequence of the lake to be monitored and the initial predicted value of the future monthly lake water volume prediction sequence, and calculate the jump difference; within a preset buffer period, introduce an asymptotic weight based on the cosine function to attenuate the jump difference and superimpose it on the prediction sequence to perform soft landing smoothing.

[0041] By splicing the historical actual water volume series with the predicted series after soft landing, and using a cross-boundary moving average mechanism at the splicing boundary to eliminate trend angles, the final output is a smooth transition future multi-scenario lake water volume evolution curve and its uncertainty quantification range.

[0042] Step S4 further includes:

[0043] Use a plotting library to create water volume change curves, including historical sequence reconstruction results and future multi-scenario prediction curves, and add uncertainty shading areas. Output monthly water volume data tables and comprehensive analysis reports.

[0044] Furthermore, the climate models include at least BCC-CSM2-MR, CESM2, CMCC-ESM2, EC-Earth3, and MPI-ESM1-2-HR; the emission scenarios include at least SSP1-2.6, SSP2-4.5, and SSP5-8.5.

[0045] Secondly, this invention discloses a lake water volume monitoring and prediction system based on multi-source remote sensing and machine learning, comprising:

[0046] The data acquisition module is used to collect historical monthly water level data, historical monthly lake area data, historical ERA5 atmospheric reanalysis data, and historical simulation data and future prediction data of the CMIP6 climate model, and to perform data quality control, repair and standardization processing.

[0047] The water volume sequence construction module is used to fuse historical monthly water level data and historical monthly lake area data using the frustum volume method, calculate the monthly water volume change, and construct the historical monthly scale water volume change sequence of the lake to be monitored.

[0048] The training set generation module is used to introduce various monthly-scale variables from historical ERA5 atmospheric reanalysis data as candidate driving factors, screen out meteorological driving factors that have a driving relationship with the water volume changes of the lake to be monitored, calculate the fluctuation of each meteorological driving factor from the global mean to construct interannual impact characteristics; use the standardized meteorological driving factor sequence as the independent variable, the historical monthly-scale water volume change sequence of the lake to be monitored as the dependent variable, and construct the training sample set according to the sequence sliding window;

[0049] The lake water volume monitoring module is used to construct and train a cascaded topology lake water volume monitoring model. This model integrates interannual impact characteristics into a standardized meteorological driving factor sequence, using this as the input to form an enhanced feature sequence. A Transformer encoder extracts long-term time-series dependent features, outputting a baseline predicted value reflecting the basic evolution trend of lake water volume and simultaneously acquiring the hidden layer state at the end of the sequence. The hidden layer state is concatenated with the meteorological features of the current time step to construct and output a dimensionality-reduced meta-feature vector. Parallel XGBoost and LightGBM regressors receive the meta-feature vector and predict algebraic residuals, respectively. These residuals are then weighted and integrated according to preset transition and smoothing preference weights, and the output is... The system outputs a fusion residual adjustment for bias compensation; it uses an independent XGBoost adjustment model to receive the meta-feature vector and predicts the fluctuation intensity factor reflecting the intensity of water volume fluctuations induced by extreme climate; it uses a LightGBM model configured with a quantile regression loss function to receive the meta-feature vector, predicts the upper and lower bound residuals of the water volume distribution boundary, and introduces an interval coverage probability calibration mechanism to dynamically calculate the scaling factor based on the actual coverage effect of the validation set, outputting the interval adjustment; it then performs an additive fusion of the baseline predicted value and the fusion residual adjustment, and performs a multiplicative scaling operation with the fluctuation intensity factor to obtain the predicted value of the target lake water volume; and simultaneously combines the interval adjustment to generate a statistically significant prediction confidence interval.

[0050] The future multi-scenario extrapolation and smoothing output module is used to correct biases in the future prediction data of the CMIP6 climate model by using historical simulation data from the CMIP6 climate model and historical ERA5 atmospheric reanalysis data from the same period. After ensemble averaging of the corrected future meteorological data from all climate models under the same emission scenario, the data is input into the trained lake water volume monitoring model to obtain a preliminary monthly lake water volume prediction sequence and the corresponding dynamic prediction interval. The module uses soft landing smoothing and cross-boundary moving average mechanisms to handle the abrupt transition problem between the historical actual water volume sequence and the soft landing smoothed monthly lake water volume prediction sequence. Finally, it outputs a smooth transition future multi-scenario lake water volume evolution curve and its uncertainty quantification range.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] First, the construction of long-term series is more accurate and continuous: By integrating Hydroweb satellite altimetry data and GEE platform Landsat imagery, and using the MNDWI index, Otsu threshold method, and SLC fault repair technology, monthly water volume series of six lakes from 2003 to 2023 were constructed. This solved the problems of insufficient spatiotemporal coverage and limited accuracy of traditional data, improved the accuracy and continuity of water volume series reconstruction, and provided a high-quality data foundation for the study of transboundary lakes in arid areas.

[0053] Second, the analysis of the driving mechanism is more systematic and comprehensive: the combined analysis of the geographic detector and atmospheric teleconnection not only identified the dominant local factors such as surface latent heat flux and air temperature, but also clarified the differentiated response patterns of different lakes. The explanatory power is improved by more than 30% compared with single factor analysis, filling the gap in the existing research on the coupled analysis of multi-scale driving mechanisms.

[0054] Third, the prediction model is more adaptable and reliable: This invention, through a collaborative mechanism of deep learning and machine learning, ensures that the prediction model is more adaptable and reliable in the context of climate change. Multi-scale adaptability in dynamic environments utilizes Transformer to extract long-term climate evolution trends and combines XGBoost / LightGBM to capture short-term nonlinear abrupt changes, enabling the model to adapt to different sequence patterns ranging from stable changes to extreme climate jumps. The introduction of "interannual shock features" allows the model to automatically identify and adapt to anomalous fluctuations in meteorological factors in different years (such as extreme drought years), thus maintaining robust prediction performance under different climate scenarios (such as SSP245 and SSP585). The system reliability under the full-chain mechanism is statistically reliable. Unlike traditional models that only provide a single prediction value, this scheme provides a mathematically validated 95% confidence interval through quantile regression and PICP (interval coverage probability) dynamic calibration technology, quantifying the uncertainty of the prediction and enhancing the scientific basis for decision-making. The introduction of the HuberLoss loss function and global offset correction reduces the impact of outliers on model convergence, ensuring the stability of the model output under extreme data conditions.

[0055] Fourth, multi-scenario prediction supports forward-looking decision-making: covering three carbon emission scenarios, SSP1-2.6, SSP2-4.5, and SSP5-8.5, it accurately identifies the risk of water volume decline in lakes such as Bosten and Zaysan under high emission scenarios, clarifies the fluctuation range and evolution path of lake water volume under different scenarios, and provides multi-dimensional quantitative references for cross-border water resource allocation, ecological protection, and climate adaptation strategy formulation, solving the problem of traditional research lacking systematic simulation of future trends.

[0056] Fifth, the technical framework is highly transferable: it establishes a standardized process of "multi-source data integration - sequence construction - mechanism analysis, machine learning modeling - multi-scenario prediction", which is adapted to the research needs of lakes in arid and semi-arid regions. It can be transferred to other regions such as the Altai Mountains and Kunlun Mountains, and can be integrated into existing remote sensing hydrological analysis systems, which has broad application prospects and promotion value. Attached Figure Description

[0057] Figure 1 A flowchart of a method and system for monitoring and predicting lake water volume based on multi-source remote sensing and machine learning;

[0058] Figure 2 A graph showing the monthly water volume changes of six lakes (Alakol, Bosten, Kapchagay, Sasykkol, Ulungur, and Zaysan) from 2003 to 2023.

[0059] Figure 3This is a schematic diagram showing the predicted water storage capacity of Lake Alakol under scenarios SSP1-2.6, SSP2-4.5, and SSP5-8.5.

[0060] Figure 4 This is a schematic diagram showing the predicted water storage capacity of Bosten Lake under scenarios SSP1-2.6, SSP2-4.5, and SSP5-8.5.

[0061] Figure 5 This is a schematic diagram showing the predicted water storage of Lake Kapchagay under scenarios SSP1-2.6, SSP2-4.5, and SSP5-8.5.

[0062] Figure 6 This is a schematic diagram showing the predicted water storage capacity of Lake Sasykkol under scenarios SSP1-2.6, SSP2-4.5, and SSP5-8.5.

[0063] Figure 7 A schematic diagram showing the predicted water storage capacity of Lake Ulungur under scenarios SSP1-2.6, SSP2-4.5, and SSP5-8.5.

[0064] Figure 8 This is a schematic diagram showing the predicted water storage capacity of Lake Zaysan under scenarios SSP1-2.6, SSP2-4.5, and SSP5-8.5. Detailed Implementation

[0065] The principles of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] This embodiment discloses a method for monitoring and predicting lake water volume based on multi-source remote sensing and machine learning. The method includes:

[0067] S1. Monthly water level data from Hydroweb satellite altimetry products for 2003-2023 were collected. Outliers were removed through data quality control to ensure the continuity of the water level series. Landsat 5 / 7 / 8 series remote sensing images were acquired based on the Google Earth Engine (GEE) platform. Interpolation was used to process SLC fault stripes in Landsat 7. Monthly lake area data were extracted using the MNDWI index and Otsu adaptive thresholding method. Missing data were supplemented using interpolation. ERA5 atmospheric reanalysis data (2m air temperature, total precipitation, 10m wind speed, and surface latent heat flux) and CMIP6 climate model data (including historical simulation data from 2003-2023 and future forecast data from 2024-2045 for BCC-CSM2-MR, CESM2, CMCC-ESM2, EC-Earth3, and MPI-ESM1-2-HR) were collected. The spatiotemporal resolution of the data was unified and standardized. The historical monthly water level data and historical monthly lake area data were fused using the frustum volume method to calculate the monthly water volume change and construct a historical monthly water volume change sequence for the lake to be monitored.

[0068] S2 introduces multiple monthly-scale variables from ERA5 reanalysis meteorological data as candidate driving factors, and selects meteorological driving factors that have a driving relationship with the water volume change of the lake to be monitored. The fluctuation of each meteorological driving factor from the global mean is calculated to construct the interannual impact characteristics. The standardized historical meteorological driving factor sequence is used as the independent variable, and the historical monthly-scale water volume change sequence of the lake to be monitored is used as the dependent variable. A training sample set is constructed according to the sequence sliding window.

[0069] Specifically, 2m air temperature, total precipitation, 10m wind speed, and surface latent heat flux were acquired from the ERA5 reanalysis data, covering the period from 2003 to 2023, with a spatial resolution of 0.25°×0.25°. Simultaneously, historical simulation data (2003-2023) for five climate models from CMIP6 (BCC-CSM2-MR, CESM2, CMCC-ESM2, EC-Earth3, and MPI-ESM1-2-HR) were acquired (for model calibration) and future forecast data (2024-2045) covering three emission scenarios: SSP1-2.6, SSP2-4.5, and SSP5-8.5. Bias corrections were applied to the historical simulation data and the ERA5 data to ensure consistency. Temporal interpolation, spatial resampling, and standardization were performed on all meteorological and climate data to eliminate dimensional differences.

[0070] S3. Construct and train a cascaded topology lake water volume monitoring model. The inference and training process of this model specifically includes:

[0071] Main trunk prediction stage: The main trunk prediction module adopts a Transformer encoder based on a multi-head self-attention mechanism, receives the enhanced feature sequence as input, uses the self-attention mechanism to extract deep abstract features with long time-series dependencies, outputs a baseline prediction value that reflects the basic evolution trend of lake water volume, and simultaneously extracts the hidden layer state at the end of the sequence.

[0072] Meta-feature extraction stage: The meta-feature extraction module receives the hidden layer state, concatenates it with the meteorological features of the current time step, constructs and outputs the dimensionality-reduced meta-feature vector.

[0073] Multi-branch parallel correction and fusion stage: The multi-branch correction and fusion module receives the meta-feature vector and assigns it to three parallel sub-network branches consisting of a two-branch residual fusion unit, an amplitude-sensitive prediction unit, and a confidence interval generation unit for independent inference and fine-tuning; wherein, the two-branch residual fusion unit: uses parallel XGBoost regressors and LightGBM regressors to receive the meta-feature vector and predict algebraic residuals respectively, performs weighted integration according to preset jump and smoothing preference weights, and outputs the fusion residual adjustment amount for bias compensation; the amplitude-sensitive prediction unit: uses an independent XGBoost adjustment model to receive the meta-feature vector and predicts the fluctuation intensity factor reflecting the intensity of water fluctuations induced by extreme climate; the confidence interval generation unit: uses a LightGBM model configured with quantile regression loss function to receive the meta-feature vector, predicts the upper and lower bound residuals of water distribution boundaries respectively, and introduces an interval coverage probability (PICP) calibration mechanism, dynamically calculates the scaling factor according to the actual coverage effect of the validation set, and outputs the interval adjustment amount.

[0074] Cascaded synthesis stage: The cascaded synthesis module adds and fuses the baseline predicted value and the fusion residual adjustment amount, and then performs a multiplicative scaling operation with the fluctuation intensity factor to obtain the predicted value of the target lake water volume; and simultaneously combines the interval adjustment amount to generate a statistically significant prediction confidence interval.

[0075] During model training, monthly ERA5 meteorological data (including 2m air temperature, total precipitation, 10m wind speed, and surface latent heat flux) from 2003 to 2023, along with actual monthly observation data of lake water volume during the same period, were input. First, a Transformer encoder based on a multi-head self-attention mechanism was used as the backbone network to learn features of the long-term temporal dependency between meteorological driving factors and water volume evolution. The Transformer encoder outputs two signals: the first output is the baseline predicted value reflecting the basic evolution trend of lake water volume; the second output is the hidden layer state at the end of the sequence. The hidden layer state not only encodes the current meteorological state but also aggregates long-term temporal dependency information within the historical time window through a self-attention mechanism, representing a high-order abstract representation of the physically driving data. Subsequently, a state-aware error correction module and an interval quantization module were constructed: using the hidden layer state as input, parallel LightGBM and XGBoost nonlinear regressors were constructed as error correction modules. Unlike traditional methods that directly predict water volume, this module predicts the prediction error (algebraic residual) that the backbone network may generate in the current high-order state, and fuses the residual prediction results of the two through an integration strategy. At the same time, a LightGBM-based interval quantization module based on quantile regression is introduced to receive the hidden layer state and predict the upper and lower bound residuals of the water volume distribution in that state. The above modules are integrated to obtain a complete lake water volume monitoring model, which is jointly trained using historical sample groups. Finally, the predicted lake water volume value is output as "baseline predicted value plus residual prediction result", and the dynamic confidence interval is output simultaneously.

[0076] S4, during forecasting, dynamic extrapolation of future multi-scenario predictions based on bias correction and soft landing smoothing mechanisms: For the CMIP6 future forecast data obtained in S1, historical simulation data from the CMIP6 multi-climate models are compared with ERA5 reanalysis data from the same period to extract the systematic biases of each climate model; the bias correction method is used to correct the CMIP6 future forecast data using the systematic biases; the corrected future meteorological data from all climate models under the same emission scenario are ensemble-averaged to generate a standardized future meteorological driving factor sequence; this sequence is then input into the trained lake water volume monitoring model. The system outputs a preliminary monthly lake water volume prediction sequence and corresponding dynamic prediction intervals. It extracts the terminal true value of the historical actual water volume sequence of the lake to be monitored and the initial predicted value of the monthly lake water volume prediction sequence, and calculates the jump difference. Within a preset buffer period, it introduces a progressive weight based on the cosine function to attenuate the jump difference and superimpose it onto the prediction sequence to perform soft landing smoothing. It further splices the historical actual water volume sequence with the soft landing smoothed prediction sequence, and uses a cross-boundary moving average mechanism at the splicing boundary to eliminate trend angles. Finally, it outputs a smooth transition future multi-scenario lake water volume evolution curve and its uncertainty quantification range.

[0077] Specifically, monthly data from the CMIP6 multi-climate model (including BCC-CSM2-MR, CESM2, CMCC-ESM2, EC-Earth3, and MPI-ESM1-2-HR) will be obtained, including historical simulation data from 2003 to 2023 and future forecast data from 2024 to 2045 (including 2m air temperature, total precipitation, 10m wind speed, and surface latent heat flux). First, a bias correction method is employed, comparing historical simulation data from the climate model with ERA5 reanalysis data from the same period to extract systematic biases, which are then used to correct future forecast data from various climate models. Subsequently, for specific greenhouse gas emission scenarios such as SSP1-2.6, SSP2-4.5, and SSP5-8.5, ensemble averaging is performed on the corrected future meteorological data from all corresponding climate models to generate a standardized sequence of future meteorological driving factors. Finally, this sequence of future meteorological driving factors is input into a trained lake water volume monitoring model, outputting a preliminary monthly-scale water volume prediction sequence and uncertainty range. At the boundary between the historical actual water volume final value and the initial value of the future prediction sequence, a progressive buffer soft landing mechanism based on a cosine function and a cross-boundary moving average mechanism are introduced to eliminate trend angles caused by time series switching, ultimately generating a future monthly lake water volume prediction curve that conforms to physical continuity and its corresponding 95% confidence interval.

[0078] Example

[0079] This example first obtains monthly water level data from 2003 to 2023 for six target lakes (Alakol, Bosten, Kapchagay, Sasykkol, Ulungur, and Zaysan) in the Tianshan region from the Hydroweb Global Land and Water Height Monitoring Platform. The data originates from a "research sequence" fused from multiple satellite altimetry products. Isolated outliers are identified using the 3-standard-deviation method and replaced by linear interpolation. For missing periods of three consecutive months or more, the data is repaired by combining the water level trends of neighboring lakes with regional meteorological data in segments to form a continuous monthly water level dataset.

[0080] Landsat 5 / 7 / 8 series imagery (30m spatial resolution, 16-day temporal resolution) was selected on the GEE platform. Local mean interpolation was used to repair SLC fault bands in Landsat 7. Atmospheric correction and cloud removal were performed on all imagery. The MNDWI index was calculated based on sensor characteristics (bands 2 and 5 for Landsat 5 / 7, and bands 3 and 6 for Landsat 8). Monthly lake boundaries were extracted and areas were calculated using the Otsu adaptive threshold segmentation method. Outliers exceeding ±50% of the historical average for the same period were removed to form a continuous area dataset. Meteorological and climate data were standardized, and ERA5 data processing was performed: monthly 2m temperature, total precipitation, 10m wind speed, and surface latent heat flux data (spatial resolution 0.25°x0.25°) for 2003-2023 were acquired. After resampling, the average values ​​for each lake basin were extracted as station-scale meteorological data. CMIP6 Data Processing: Five climate models, including BCC-CSM2-MR, were selected, and historical simulation data (2003-2023) and future forecast data (2024-2045) were collected. Four parameters—tas, pr, sfcwind, and hfls—were extracted and linearly interpolated to a monthly resolution. Systematic errors were corrected using bias correction, and meteorological sequences under three emission scenarios were extracted. Monthly SOI data from 2003-2023 were collected, smoothed using a 3-month moving average, and used for atmospheric teleconnection analysis. Water level, area, meteorological, and SOI data were aligned by year-month timestamps, and periods without synchronized data were removed. Z-score standardization was used to eliminate dimensional differences, and a standardized comprehensive dataset containing "time-lake name-water level-area-meteorological factors-SOI index" was finally constructed to provide statistical data support for subsequent analysis.

[0081] The Landsat 5TM, Landsat 7ETM+, and Landsat 8OL1 image datasets were used on the GEE platform, with a time range limited to 2003-2023 and a spatial resolution of 30m. For the missing bands in the Landsat 7 images after 2003 caused by SLC failure, local mean interpolation was used to fill in the missing pixels to restore image integrity. Atmospheric correction was uniformly applied to all images to remove cloud shadows and snow / ice interference, ensuring the accuracy of water body extraction. The local mean interpolation algorithm was used to fill in the missing pixels, restoring image integrity. The restored lake areas effectively eliminated data gaps and improved the continuity of the time series. This solved the problem of spatiotemporal discontinuity caused by sensor failure in remote sensing data, ensuring the availability and consistency of long-term series data from 2003-2023.

[0082] To enhance the distinction between water and non-water bodies, MNDWI indices were constructed based on the characteristics of different sensor bands. Landsat 5 / 7 used a green light band (2 bands) and a shortwave infrared band (5 bands), while Landsat 8 used a green light band (3 bands) and a shortwave infrared band (6 bands). The calculation formula is as follows:

[0083] ;

[0084] In this index, Green corresponds to the green band of Landsat (Band 2 for TM / ETM+, Band 3 for OLI), and SWIR corresponds to the shortwave infrared band (Band 5 for TM / ETM+, Band 6 for OLI). This index effectively reduces the interference of building shadows and vegetation on water body extraction, improving the accuracy of identifying turbid water bodies in arid areas. By comparing the differences between the Green band and the SWIR band, the distinguishability between water bodies and background features (such as vegetation and buildings) is effectively enhanced, solving the problem that traditional NDWI is easily interfered with by ice, snow, and shadows in arid areas. The results shown in the figure verify the high-precision water body identification capability of the MNDWI algorithm in the complex terrain of the Tianshan region, providing a reliable data foundation for subsequent area sequence construction. It overcomes the problems of inaccurate water body extraction and high susceptibility to environmental noise in optical remote sensing images, ensuring the continuity and accuracy of the monthly lake area sequence.

[0085] Based on the calculated MNDWI index imagery, the Otsu adaptive threshold segmentation method is employed. This method automatically determines the segmentation threshold by maximizing the inter-class variance between water and non-water bodies, achieving unsupervised water body classification. Vector boundaries are extracted from the classification results, and areas smaller than 0.1 km² are removed. 2 To eliminate noise interference, fragmented patches were extracted to calculate the actual monthly lake surface area. For anomalous area values ​​exceeding the historical average by ±50%, their rationality was verified by combining concurrent water level data and meteorological conditions. Outliers were replaced using linear interpolation to ensure the continuity of the area sequence. The extracted monthly lake area data was then subjected to response coupling analysis with concurrent water level data obtained from the Hydroweb platform to verify the consistency between area and water level trends (correlation coefficient ≥ 0.7). This resulted in a continuous monthly lake area sequence from 2003 to 2023, corresponding one-to-one with the water level sequence, providing matching data support for subsequent water volume calculations.

[0086] Based on this, the monthly area data of six target lakes (Alakol, Bosten, Kapchagay, Sasykkol, Ulungur, and Zaysan) extracted from 2003 to 2023 were precisely matched with the monthly water level data of the same period obtained from the Hydroweb platform according to the format of "lake name-year-month" to ensure that the water level (H) and area (A) data corresponded for the same period. The matched data were then subjected to consistency verification to remove abnormal periods where the water level and area change trends contradicted each other (such as a significant rise in water level but a significant reduction in area). A small number of missing synchronous data were filled in by linear interpolation. Figure 2 Monthly water volume variation curves for six lakes (Alakol, Bosten, Kapchagay, Sasykkol, Ulungur, Zaysan) from 2003 to 2023. Figure 2 It displays the comprehensive dynamics of water volume in various lakes (such as the rising-falling trend of Lake Alakol) and reveals the sensitivity of water volume to climate response (such as summer meltwater peaks). It solves the problems of separation of water level and area data and inaccurate water volume estimation, and provides a complete spatiotemporal series of lake water volume, supporting subsequent driving analysis and predictive modeling.

[0087] Based on the assumption that the lake's cross-section is approximately a frustum, the monthly water volume change is calculated using the formula for the difference in frustum volume. The formula is as follows:

[0088] ;

[0089] In the formula and The lake surface area for two consecutive months is shown. This formula represents the change in water level. A continuous monthly water volume change series from 2003 to 2023 was constructed using this formula. This formula transforms the dynamic changes in water level and area into quantified water volume changes, adapting to the volume calculation needs of irregular lakes. The water volume change series of the six lakes were standardized in format, clearly labeling each data point with its timestamp, lake name, water volume value, and data quality level (high quality / corrected / interpolated). This resulted in a standardized long-term lake water volume change series dataset from 2003 to 2023, providing core data support for subsequent analysis of the driving mechanism and model training.

[0090] Four core meteorological factors at the monthly scale were extracted from ERA5 atmospheric reanalysis data from 2003 to 2023, including 2m air temperature (t2m, unit: K), total precipitation (tp, unit: m), 10m wind speed (w10, unit: m / s), and surface latent heat flux (slhf, unit: J / m²). 2The data were spatiotemporally matched, and the average value of the watershed where each lake is located was extracted as the station-scale factor data. Z-score standardization was used to eliminate dimensional differences and outliers exceeding three standard deviations were removed to ensure data consistency and reliability. Based on the spatial heterogeneity principle of geographic detectors, a driving relationship analysis framework of "meteorological factors-lake water volume change" was constructed. Four standardized meteorological factors were used as independent variables, and the monthly lake water volume change series from 2003 to 2023 was used as the dependent variable. The sample groups were divided by year to retain seasonal and interannual variability characteristics.

[0091] Four meteorological driving factor sequences—temperature, precipitation, wind speed, and latent heat flux—were selected as the model input independent variables, while the historical monthly water volume variation sequence of the lake was used as the output dependent variable. To eliminate heteroscedasticity caused by extreme water volume values ​​and avoid model gradient explosion, a offset correction and natural logarithmic transformation were first applied to the lake water volume output sequence. Simultaneously, an interannual impact feature was constructed for the meteorological driving factor input sequence, calculating the fluctuation of each meteorological factor from its historical global mean in each year. This enhanced feature is specifically designed to assist the downstream model in accurately identifying the nonlinear physical impacts of extreme climate years (such as sudden droughts and torrential rains) on lake water volume.

[0092] A Transformer encoder based on a multi-head self-attention mechanism is constructed as the backbone network. Leveraging its long-term dependency modeling capabilities, deep abstract features are extracted from historical augmented sequences to output baseline predicted values ​​representing the basic evolution trend of lake water volume, while simultaneously extracting the hidden layer states at the end of the sequence. Receiving the hidden layer states output by the Transformer, on the one hand, the algebraic residuals generated by the backbone model are trained using parallel XGBoost and LightGBM nonlinear regressors, and fused according to preset weights for bias compensation; on the other hand, an amplitude-sensitive model is established, using independent XGBoost to learn and predict the fluctuation intensity factor for future years, and using it as a multiplicative operator to dynamically fine-tune the peak and trough values ​​of the predicted values ​​to accurately capture strong nonlinear jumps. A dynamic prediction interval is established using the quantile regression loss function of LightGBM (configured with Alpha=0.05 and 0.95), and an innovative interval coverage probability (PICP) calibration mechanism is introduced. Based on the actual coverage effect of the validation set, the multiplier is dynamically calculated and the prediction boundary is scaled to generate a 95% confidence interval with rigorous statistical significance. To assess the long-term impacts of different carbon emission intensities on lake water volume, five stable climate models from the CMIP6 ensemble (BCC-CSM2-MR, CESM2, CMCC-ESM2, EC-Earth3, and MPI-ESM1-2-HR) were selected. Monthly meteorological data for 2024-2045 under three greenhouse gas emission scenarios (SSP1-2.6, SSP2-4.5, and SSP5-8.5) were extracted for these five models. Historical simulation data were compared with ERA5 reanalysis data, and the inherent systematic bias of the climate models was corrected using a bias correction method. Subsequently, downscaling was performed to uniformly resample the data spatial resolution to 0.25°×0.25° to accurately match the watershed-scale requirements of the lakes to be monitored. For each SSP scenario, ensemble averaging was performed on the five corrected model data. The arithmetic mean method effectively reduced the uncertainty of a single climate model, generating robust and regionally adaptable future meteorological driving sequences. The standardized future meteorological driving sequence is input into the trained cascaded model for forward collaborative inference. Specifically, the backbone network outputs a baseline evolution curve, and a multi-branch correction model is used to superimpose and fuse residuals based on the current meteorological conditions, multiplying this by a fluctuation intensity factor to calculate a preliminary monthly prediction sequence of lake water volume changes and its uncertainty range. Building upon this, a progressive soft-landing smoothing mechanism based on cosine curve weights is introduced at the junction of the historical actual sequence and the future prediction sequence to eliminate rigid trend angles caused by data source switching.Finally, through the visualization rendering module, the system outputs the continuous evolution curves (including the uncertainty shaded area) of the historical sequence reconstruction results and future multi-scenario predictions, and simultaneously generates monthly water volume projection data reports and comprehensive comparative analysis reports, clarifying the differences in lake water resource evolution under different greenhouse gas emission scenarios, and providing quantitative decision support for watershed water resource allocation and disaster prevention and mitigation management.

[0093] Figures 3 to 8 This paper presents water volume projection curves for six lakes under SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios from 2024 to 2045, including median projections and uncertainty ranges (e.g., ±1 standard deviation). The vertical dashed lines in the figures represent the time boundary between historical and projection periods. The illustrations reveal water volume change trends under different emission pathways (e.g., a significant decrease in Lake Boston's water volume under SSP5-8.5), highlighting the risks of high emission scenarios. This approach addresses the lack of multi-scenario and multi-model integration in future projections, providing a visualized risk assessment tool to support adaptive water resource management decisions.

[0094] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0095] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for monitoring and predicting lake water volume based on multi-source remote sensing and machine learning, characterized in that, The method includes: S1. Collect historical monthly water level data, historical monthly lake area data, historical ERA5 atmospheric reanalysis data, and historical simulation data and future prediction data of CMIP6 climate model. After data quality control, repair and standardization, the historical monthly water level data and historical monthly lake area data are fused using the frustum volume method to calculate the monthly water volume change and construct the historical monthly water volume change sequence of the lake to be monitored. S2 introduces various monthly-scale variables from historical ERA5 atmospheric reanalysis data as candidate driving factors, and selects meteorological driving factors that have a driving relationship with the water volume change of the lake to be monitored. The fluctuation of each meteorological driving factor from the global mean is calculated to construct the interannual impact characteristics. The standardized meteorological driving factor sequence is used as the independent variable, the historical monthly-scale water volume change sequence of the lake to be monitored is used as the dependent variable, and a training sample set is constructed according to the sequence sliding window. S3. Construct and train a cascaded topology lake water volume monitoring model. This model integrates interannual impact characteristics into a standardized meteorological driving factor sequence, using this as the input to form an enhanced feature sequence. A Transformer encoder extracts long-term time-dependent features, outputting a baseline predicted value reflecting the basic evolution trend of lake water volume and simultaneously acquiring the hidden layer state at the end of the sequence. The hidden layer state is concatenated with the meteorological features of the current time step to construct and output a dimensionality-reduced meta-feature vector. Parallel XGBoost and LightGBM regressors receive the meta-feature vector and predict algebraic residuals, respectively. These residuals are then weighted and integrated according to preset jump and smoothing preference weights, outputting a biased value. The method involves: adjusting the fusion residual for difference compensation; using an independent XGBoost adjustment model to receive the meta-feature vector and predicting the fluctuation intensity factor reflecting the intensity of water volume fluctuations induced by extreme climate; using a LightGBM model configured with a quantile regression loss function to receive the meta-feature vector and predict the upper and lower bound residuals of the water volume distribution boundary, respectively, and introducing an interval coverage probability calibration mechanism to dynamically calculate the scaling factor based on the actual coverage effect of the validation set, and outputting the interval adjustment; adding and fusing the baseline predicted value and the fusion residual adjustment, and then performing a multiplicative scaling operation with the fluctuation intensity factor to obtain the predicted value of the target lake water volume; and simultaneously combining the interval adjustment to generate a statistically significant prediction confidence interval. S4 uses historical simulation data from the CMIP6 climate model and historical ERA5 atmospheric reanalysis data from the same period to correct biases in the future prediction data of the CMIP6 climate model. After ensemble averaging of the corrected future meteorological data from all climate models under the same emission scenario, the data is input into the trained lake water monitoring model to obtain a preliminary monthly lake water prediction sequence and corresponding dynamic prediction interval. Through soft landing smoothing and cross-boundary moving average mechanisms, the abrupt transition problem between the historical actual water volume sequence and the soft landing smoothed monthly lake water prediction sequence is addressed. Finally, a smooth transition future multi-scenario lake water evolution curve and its uncertainty quantification range are output.

2. The lake water volume monitoring and prediction method based on multi-source remote sensing and machine learning according to claim 1, characterized in that, Step S1 further includes: S11: Collect historical monthly water level data from Hydroweb satellite altimetry products, remove outliers through data quality control to ensure the continuity of the water level sequence, and generate historical monthly water level sequences for the lake to be monitored. S12: Acquire Landsat 5 / 7 / 8 series remote sensing images, use interpolation to repair SLC fault stripes in Landsat 7, extract monthly lake area data using MNDWI index and Otsu adaptive thresholding method, and supplement missing data using interpolation. S13 collects ERA5 atmospheric reanalysis data, historical simulation data and future forecast data of CMIP6 climate model, unifies the spatiotemporal resolution of the data and performs standardization processing. S14. The historical monthly water level data and historical monthly lake area data were fused using the frustum volume method to calculate the monthly water volume change. : ； In the formula and The lake surface area for two consecutive months is shown. and Historical monthly water volume variation sequences of the lake to be monitored were constructed by taking the water levels of two adjacent months as examples. The monthly water volume changes of the lakes to be monitored are standardized in a unified format, and the timestamp, lake name and data quality level of each data point are labeled to construct a historical monthly water volume change sequence of the lakes to be monitored.

3. The lake water quantity monitoring and prediction method based on multi-source remote sensing and machine learning according to claim 2, characterized in that, Step S12, the process of extracting monthly lake area data using the MNDWI index and the Otsu adaptive thresholding method, includes the following steps: Using the Google Earth Engine platform, an improved Normalized Difference Water Index (MNDWI) was calculated from preprocessed Landsat imagery. ； Green corresponds to the green light band of Landsat, and SWIR corresponds to the short-wave infrared band. The Otsu adaptive threshold segmentation method is adopted to automatically determine the segmentation threshold by maximizing the inter-class variance between water bodies and non-water bodies, and to perform unsupervised classification of water bodies. Vector boundary extraction is performed on the classification results to remove small patches with an area smaller than the preset area threshold to eliminate noise interference, and the actual water surface area of ​​the lake is calculated every month. The extracted monthly lake surface area is coupled with the water level data obtained from the Hydroweb platform for the same period. First, abnormal area values ​​that exceed the historical average or are greater than a preset percentage are verified for rationality by combining the water level data and meteorological conditions for the same period, and the abnormal values ​​are replaced by linear interpolation. Then, the consistency between the area and water level change trends is verified, and abnormal periods with contradictory water level and area change trends are removed and filled by linear interpolation. Finally, continuous historical monthly lake area data is formed and a correspondence is established with the historical monthly water level sequence.

4. The lake water volume monitoring and prediction method based on multi-source remote sensing and machine learning according to claim 1, characterized in that, Step S2 further includes: Multiple monthly-scale variables, including 2m air temperature, total precipitation, 10m wind speed, and surface latent heat flux, were introduced as candidate driving factors. Historical ERA5 atmospheric reanalysis data were extracted as the historical actual observation data for each candidate driving factor. The extracted historical ERA5 atmospheric reanalysis data, monthly water level data, and monthly lake area data of the lake to be monitored were uniformly converted to monthly temporal resolution by linear interpolation. The candidate driving factors, water level, and area sequences were strictly aligned according to timestamps, and periods without synchronized data were removed. Finally, the Z-Score normalization method was used to eliminate the dimensional differences between variables, and a historical multi-source aligned dataset containing time stamps, lake identifiers, candidate driving factors, and corresponding water level and area labels was constructed. A nonlinear mapping relationship between candidate driving factors and lake water volume changes was fitted using a historical multi-source aligned dataset; Based on the principle of spatial heterogeneity of geographic detectors, the explanatory power and significance of each candidate driving factor are calculated. Meteorological driving factors that have a driving relationship with the water volume change of the lake to be monitored are selected, and an analytical framework for the driving relationship between meteorological driving factors and lake water volume change is constructed.

5. The lake water volume monitoring and prediction method based on multi-source remote sensing and machine learning according to claim 1, characterized in that, In step S3, the lake water volume monitoring model includes a main trunk prediction module, a meta-feature extraction module, a multi-branch correction and fusion module, and a cascade synthesis module. The main prediction module adopts a Transformer encoder based on a multi-head self-attention mechanism. It receives the enhanced feature sequence as input, uses the self-attention mechanism to extract deep abstract features with long time-series dependencies, outputs a baseline prediction value that reflects the basic evolution trend of lake water volume, and simultaneously extracts the hidden layer state at the end of the sequence. The meta-feature extraction module receives the hidden layer state, concatenates it with the meteorological features of the current time step, constructs and outputs the dimensionality-reduced meta-feature vector; The multi-branch correction fusion module receives the meta-feature vector and assigns it to three parallel sub-network branches consisting of a two-branch residual fusion unit, an amplitude-sensitive prediction unit, and a confidence interval generation unit for independent inference and fine-tuning. Specifically, the two-branch residual fusion unit uses parallel XGBoost and LightGBM regressors to receive the meta-feature vector and predict algebraic residuals, then performs weighted integration according to preset jump and smoothing preference weights, outputting a fusion residual adjustment amount for bias compensation. The amplitude-sensitive prediction unit uses an independent XGBoost adjustment model to receive the meta-feature vector and predicts the fluctuation intensity factor reflecting the intensity of water volume fluctuations induced by extreme climate. The confidence interval generation unit uses a LightGBM model configured with a quantile regression loss function to receive the meta-feature vector, predicts the upper and lower bound residuals of the water volume distribution boundary, and introduces an interval coverage probability calibration mechanism to dynamically calculate the scaling factor based on the actual coverage effect of the validation set, outputting the interval adjustment amount. The cascaded fusion module adds and fuses the baseline predicted value and the fusion residual adjustment, and then performs a multiplicative scaling operation with the fluctuation intensity factor to obtain the predicted value of the target lake water volume; and simultaneously combines the interval adjustment to generate a statistically significant prediction confidence interval.

6. The lake water volume monitoring and prediction method based on multi-source remote sensing and machine learning according to claim 1, characterized in that, Step S4 further includes: For the future prediction data of the acquired CMIP6 climate model, the historical simulation data of the CMIP6 climate model is compared with the historical ERA5 atmospheric reanalysis data of the same period to extract the systematic bias of each climate model; the bias correction method is used to correct the future prediction data of the CMIP6 climate model using the systematic bias. The ensemble averaging of future meteorological data from all climate models under the same emission scenario is used to generate a standardized sequence of future meteorological driving factors. This sequence is then input into a trained lake water volume monitoring model to output a preliminary monthly lake water volume prediction sequence and the corresponding dynamic prediction interval. Extract the terminal true value of the historical actual water volume sequence of the lake to be monitored and the initial predicted value of the future monthly lake water volume prediction sequence, and calculate the jump difference; within a preset buffer period, introduce an asymptotic weight based on the cosine function to attenuate the jump difference and superimpose it on the prediction sequence to perform soft landing smoothing. By splicing the historical actual water volume series with the predicted series after soft landing, and using a cross-boundary moving average mechanism at the splicing boundary to eliminate trend angles, the final output is a smooth transition future multi-scenario lake water volume evolution curve and its uncertainty quantification range.

7. The lake water volume monitoring and prediction method based on multi-source remote sensing and machine learning according to claim 1, characterized in that, Step S4 further includes: Use a plotting library to create water volume change curves, including historical sequence reconstruction results and future multi-scenario prediction curves, and add uncertainty shading areas. Output monthly water volume data tables and comprehensive analysis reports.

8. The lake water volume monitoring and prediction method based on multi-source remote sensing and machine learning according to claim 1, characterized in that, The climate models include at least BCC-CSM2-MR, CESM2, CMCC-ESM2, EC-Earth3, and MPI-ESM1-2-HR; the emission scenarios include at least SSP1-2.6, SSP2-4.5, and SSP5-8.

5.

9. A lake water volume monitoring and prediction system based on multi-source remote sensing and machine learning, characterized in that, include: The data acquisition module is used to collect historical monthly water level data, historical monthly lake area data, historical ERA5 atmospheric reanalysis data, and historical simulation data and future prediction data of the CMIP6 climate model, and to perform data quality control, repair and standardization processing. The water volume sequence construction module is used to fuse historical monthly water level data and historical monthly lake area data using the frustum volume method, calculate the monthly water volume change, and construct the historical monthly scale water volume change sequence of the lake to be monitored. The training set generation module is used to introduce various monthly scale variables from historical ERA5 atmospheric reanalysis data as candidate driving factors, screen out meteorological driving factors that have a driving relationship with the changes in water volume of the lake to be monitored, and calculate the fluctuation of each meteorological driving factor from the global mean to construct interannual impact characteristics. The standardized meteorological driving factor sequence was used as the independent variable, and the historical monthly water volume change sequence of the lake to be monitored was used as the dependent variable. A training sample set was constructed according to the sequence sliding window. The lake water volume monitoring module is used to construct and train a cascaded topology lake water volume monitoring model. This model incorporates interannual impact characteristics into a standardized meteorological driving factor sequence, using this sequence as the model input to form an enhanced feature sequence. A Transformer encoder extracts long-term time-series dependent features, outputting a baseline predicted value reflecting the basic evolution trend of lake water volume and simultaneously acquiring the hidden layer state at the end of the sequence. The hidden layer state is concatenated with the meteorological features of the current time step to construct and output a dimensionality-reduced meta-feature vector. Parallel XGBoost and LightGBM regressors are used to receive the meta-feature vector and predict the algebraic residuals, according to preset transitions and... The smoothing preference weights are weighted and integrated to output a fusion residual adjustment for bias compensation. An independent XGBoost adjustment model is used to receive the meta-feature vector and predict the fluctuation intensity factor reflecting the intensity of water fluctuations induced by extreme climate. A LightGBM model configured with quantile regression loss function is used to receive the meta-feature vector and predict the upper and lower bound residuals of the water distribution boundary. An interval coverage probability calibration mechanism is introduced to dynamically calculate the scaling factor based on the actual coverage effect of the validation set and output the interval adjustment. The baseline prediction value and the fusion residual adjustment are added together and then multiplied and scaled with the fluctuation intensity factor to obtain the predicted water volume of the target lake. Simultaneously, the interval adjustment amount is combined to generate a statistically significant prediction confidence interval; The future multi-scenario extrapolation and smoothing output module is used to correct biases in the future prediction data of the CMIP6 climate model by using historical simulation data from the CMIP6 climate model and historical ERA5 atmospheric reanalysis data from the same period. After ensemble averaging of the corrected future meteorological data from all climate models under the same emission scenario, the data is input into the trained lake water volume monitoring model to obtain a preliminary monthly lake water volume prediction sequence and the corresponding dynamic prediction interval. Through soft landing smoothing and cross-boundary moving average mechanisms, the module addresses the jump problem in splicing the historical actual water volume sequence and the soft landing smoothed monthly lake water volume prediction sequence, and finally outputs a smooth transition future multi-scenario lake water volume evolution curve and its uncertainty quantification range.

Images