Deep learning-based poplar catkin flying concentration space-time diffusion forecasting method

Abstract

The invention discloses a poplar and willow catkin flying concentration space-time diffusion forecasting method based on deep learning, and relates to the technical field of environmental monitoring, and the method specifically comprises the steps: firstly collecting multi-source data such as meteorology, phenology, poplar and willow distribution, terrain and the like, carrying out the preprocessing, unifying the data to a grid with a preset size, then standardizing the data, quantifying the phenology characteristics, and carrying out the prediction of the poplar and willow catkin flying concentration. The method comprises the following steps: determining phenological parameters and effective weights, calculating initial flying concentration by combining area types, tree density and a temperature and humidity correction function, integrating the initial flying concentration into a multi-dimensional tensor, and finally, fusing meteorological and terrain correction factors through a three-section model of spatial-temporal feature extraction-phenological attention fusion-diffusion prediction, and performing phenological period weighted mean square error loss function training to obtain the phenological period. According to the method, the phenological period is quantified into a continuously changing dynamic factor by setting the phenological validity weight, so that the rationality of source estimation is improved.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring technology, specifically relating to a deep learning-based method for predicting the spatiotemporal diffusion concentration of poplar and willow catkins. Background Technology

[0002] As important tree species for urban greening in my country, poplar and willow trees play a vital role in purifying the air, beautifying the environment, and regulating the climate. However, the large-scale dispersal of poplar and willow catkins every spring causes a series of prominent problems: on the one hand, as allergens, these catkins can trigger respiratory diseases, skin itching, and other health issues, seriously affecting the daily lives of sensitive individuals; on the other hand, the catkins are lightweight and flammable, easily accumulating in building crevices and vegetation, posing a significant fire hazard. They also obstruct vision, disrupt traffic and the normal operation of public facilities, causing numerous inconveniences to urban management and residents' lives. Therefore, achieving accurate spatiotemporal dispersion forecasts of poplar and willow catkin concentrations to provide scientific support for prevention and control measures and emergency response has become an urgent need in the fields of environmental monitoring and urban management. However, current monitoring and forecasting technologies related to willow catkin dispersal still have many shortcomings and cannot meet the needs of practical applications. Existing technologies mostly treat phenological periods (the beginning, peak, and end of dispersal) as discrete time nodes, without considering their spatial heterogeneity and continuous dynamic changes in time. This makes it impossible to accurately depict the differences in dispersal potential in different regions and at different times, resulting in large deviations in the estimation of dispersal intensity at the source. Willow catkin dispersal is comprehensively affected by multiple factors such as meteorology (wind speed, wind direction, temperature, humidity, etc.), phenology, vegetation distribution, and topography. However, existing methods mostly consider only some factors or simply superimpose the effects of various factors, without establishing a deep coupling mechanism for multi-source data. This makes it difficult to accurately reflect the synergistic effects of various factors on dispersal and diffusion. Existing technologies mostly only output single concentration values ​​and lack visualization and targeted outputs such as phenological matching degree and high-concentration risk area boundaries, making it difficult to directly provide intuitive support for urban prevention and control deployment and key area management. Therefore, there is an urgent need for a deep learning-based method for predicting the spatiotemporal diffusion of poplar and willow catkin concentrations to solve the above problems. Summary of the Invention

[0003] The purpose of this invention is to provide a deep learning-based method for predicting the spatiotemporal diffusion of poplar and willow catkin concentration, which solves the technical problems of crude phenological quantification methods and unreasonable concentration estimation in existing technologies.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based method for predicting the spatiotemporal diffusion concentration of willow and poplar catkins includes: Step 1: Collect basic data related to poplar and willow catkins and standardize the data. By integrating ground observation, remote sensing inversion and historical data, key phenological periods and spatial distributions are obtained. Poplar and willow distribution areas are identified and density is calculated. Regional types are classified and digital elevation models are extracted to achieve the collection and preprocessing of multi-source data. Observation points were set up in layers according to the distribution density of poplar and willow trees. Photos were taken and observed at set intervals to record the flower bud swelling period, inflorescence appearance period, the beginning, peak, and end of the flowering period, as well as the spatial information of the observation points. High-resolution satellite images of GF-1 / GF-2 and MODIS satellite NDVI / EVI data were used to calculate the vegetation index time series curve and extract phenological nodes. The data were converted into a grid spatial distribution map and integrated with the ground observation data and remote sensing inversion data of the current year, as well as historical phenological data after outlier removal and time series correction, to form key phenological periods and spatial distribution data. Step 2: Unify the collected data to the preset grid range and standard time step, quantify the phenological period as a relative time feature, determine the phenological period-specific parameters through a nonlinear optimization algorithm, calculate the phenological effectiveness weight, combine the regional type weight, poplar and willow tree distribution density and temperature and humidity correction function, calculate the initial drift concentration of each grid at each time step through the formula, and integrate it into a structured multidimensional tensor. Step 3: A three-stage model of spatiotemporal feature extraction, phenological attention fusion, and diffusion prediction is adopted. By defining meteorological coupling and terrain attenuation correction factors, the final drift concentration prediction formula is constructed. The model is trained and optimized using the phenological period weighted mean square error as the loss function. Finally, the model outputs the future time-by-time grid concentration forecast, phenological matching degree heat map, and high concentration risk warning boundary.

[0005] Furthermore, the distribution areas of poplar and willow trees were identified and their density calculated. The area types were then classified, and digital elevation models were extracted. The specific methods are as follows: Based on high-resolution remote sensing images and GIS databases, the distribution areas of poplar and willow trees are identified through object-oriented classification or supervised classification algorithms. Combined with the green space boundary vector data in the GIS database for verification, the number of trees is counted by grid and the distribution density is calculated to extract the distribution density of poplar and willow trees. Spectral and spatial features were extracted using GF-1 / GF-2 high-resolution remote sensing images. The monitoring area was divided into different types using a support vector machine supervised classification algorithm. The classification of monitoring area types was achieved by overlaying and correcting the road network and green space planning vector data from the GIS database. Obtain the raw SRTM or ASTERGDEM data, and after preprocessing such as denoising, filling depressions, and smoothing, resample to a uniform grid resolution to extract the average elevation information of each grid.

[0006] Furthermore, the quantification of phenological periods as relative temporal characteristics is achieved through the following methods: Taking the starting point of the drifting of grid (i, j) as the reference starting point, denoted as Convert the current time step t to the cumulative duration relative to the beginning of the drifting period, and combine it with the duration of the peak period, denoted as . Using the formula It represents the relative time characteristic and is used to quantify the position of the current time step in the entire drift cycle.

[0007] Furthermore, the phenological effectiveness weight is calculated using the following method: Based on relative time characteristics and setting specific parameters for phenological periods, using formulas A phenological effectiveness weight is set to quantify the degree of matching between the current time step and the phenological period. This represents the phenological period type of grid (i, j) at time step t. This represents the phenological characteristic parameters of grid (i, j) at time step t. This represents the phenological potential weight of grid (i, j) at time step t. For the Sigmoid function, The slope parameter is derived from key phenological periods and drift concentration correlation data fused from ground-based fixed-point observations, remote sensing inversion, and historical data. It is obtained by fitting the Sigmoid function relationship between phenological effectiveness weights and relative time characteristics and minimizing the initial drift concentration estimation bias.

[0008] Furthermore, by combining the region type weights, willow tree distribution density, and temperature and humidity correction functions, the initial spore concentration at each time step for each grid is calculated using a formula. The specific method is as follows: Based on field observation data on the release potential and diffusion resistance of willow and poplar catkins, and combined with phenological characteristics and underlying surface influence mechanisms, the weights of different regional types were determined. The true value, combined with the distribution density of poplar and willow trees within the grid, the weight of the phenological effectiveness they represent, and the weight of the region type, is used using the formula Indicates the concentration of drift, where, This represents the initial drift concentration at time step t, in grid (i, j). The region type weight of grid (i, j) is represented. This represents the distribution density of poplar and willow trees in grid (i, j). The temperature and humidity correction function is obtained using the formula. Where T represents the average temperature of the grid and R represents the relative humidity of the grid. and These represent the temperature coefficient and humidity coefficient, respectively, obtained through quantitative calibration based on the meteorological response mechanism of poplar and willow catkin drift and historical observation data.

[0009] Furthermore, a three-stage model of spatiotemporal feature extraction, phenological attention fusion, and diffusion prediction is adopted. The specific method is as follows: A stacked structure of ConvLSTM and 3D convolution is adopted. The ConvLSTM layer captures the temporal dependence of concentration diffusion between time steps, while the 3D convolutional layer extracts the diffusion correlation features of the spatial neighborhood grid. At the same time, the spatial heterogeneity of terrain and region type is preserved, and an adaptive attention mechanism is introduced to weight phenological effectiveness. and the distribution density of poplar and willow trees For attention weights, use the formula Weighted enhancement of spatiotemporal features, among which, This represents the coordinates of the neighboring grids of the current grid (i, j). The original spatiotemporal features are output by the spatiotemporal feature extraction module.

[0010] Furthermore, the meteorological coupling factor and the topographic attenuation correction factor are defined, and the specific method is as follows: By integrating the combined effects of wind speed, wind direction, and humidity, a meteorological coupling correction factor is set. This embodiment utilizes the formula... It means that among them Represents the grid (i, j) at future time steps. wind speed, This indicates the angle between the wind direction and the grid diffusion direction. Represents the grid (i, j) at future time steps. The precipitation intensity, a1 and a2 are statistical fitting empirical coefficients based on historical observation data. By analyzing the correlation between historical wind speed and the diffusion increment of poplar and willow catkins, the wind speed term output range is matched with the actual diffusion intensity to obtain the value of a1. a2 is obtained by fitting the measured data of precipitation intensity and concentration decay. Considering the hindering effect of elevation on diffusion, a terrain attenuation correction factor is set, and the specific calculation formula is as follows: , Represents a grid Average elevation, The elevation attenuation coefficient was determined by collecting measured data on the diffusion concentration of poplar and willow catkins at different elevation gradients, analyzing the correlation between elevation and concentration attenuation, and using a nonlinear optimization algorithm to fit the coefficients with the goal of minimizing the error between the model-predicted concentration and the measured concentration.

[0011] Furthermore, the angle between the wind direction and the grid diffusion direction specifically includes: Grid diffusion refers to the direction of poplar and willow catkin migration and propagation between standardized grids. Taking any grid (i, j) as the core, its diffusion direction is the direction from that grid to the corresponding direction of the adjacent grid, which is quantified as azimuth angle. If the poplar and willow tree distribution density of the core grid (i, j) is higher than that of a certain neighboring grid... ,but The direction is the diffusion direction, targeting the core mesh. For each potential diffusion direction, a fixed grid diffusion direction azimuth angle is set. Combining the wind direction azimuth and the determined grid diffusion direction azimuth, the angle between the wind direction and the grid diffusion direction is calculated using the formula... express.

[0012] Furthermore, the feature is that a formula for predicting the final drift concentration is constructed, specifically using the following method: The model learns the mapping relationship from input features to the relative diffusion increment ratio, using the measured concentration change ratio as the label. The output dimension of the diffusion prediction head is consistent with the number of grid cells, and the activation function is Tanh. The output is the diffusion concentration increment. , The physical meaning of is the proportion of concentration change; Increment of diffusion concentration based on model output Combining initial concentration and multi-factor correction, using the formula Construct the final formula for predicting drift concentration, where, Indicates a step in the future time. The phenological effectiveness weight of multi-element coupling is equal to the cumulative product of the phenological effectiveness weight, the meteorological coupling correction factor, and the topographic attenuation correction factor.

[0013] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention transforms the phenological period into a continuous dynamic factor through relative time characteristics, and combines the Sigmoid function and nonlinear optimization algorithm to determine the phenological effectiveness weight, thereby achieving a smooth and gradual characterization of the scattering potential from the beginning to the peak to the end of the period. This solves the problem of source intensity estimation deviation caused by traditional discretized phenological processing, and improves the rationality and accuracy of the initial scattering concentration estimation. 2. This invention integrates multi-source data such as meteorology, phenology, poplar and willow tree distribution, and topography, unifies them into a standardized grid and time step, and constructs a structured multidimensional tensor input. This breaks through the limitations of traditional methods that consider a single factor or simply superimpose data, and deeply explores the synergistic effects of various factors on the drift and diffusion of poplar and willow catkins, so that the forecast can fully reflect the complex mechanisms of action in the actual environment. 3. This invention constructs a multi-dimensional correction system for the entire process by introducing regional type weights, temperature and humidity correction functions, meteorological coupling factors, and topographic attenuation correction factors. It not only considers the impact of underlying surface differences on the release of willow catkins, but also accurately depicts the regulatory effects of wind speed, wind direction, precipitation, and elevation on diffusion. This makes the prediction results more consistent with the physical laws of willow catkin diffusion, enhancing reliability. The use of a phenological period weighted mean square error loss function makes the model training more focused on the prediction accuracy of key release periods. The AdamW optimizer and early stopping method are combined to prevent overfitting. The parameters are calibrated by fusing historical data to ensure that the model can stably output reliable results under different regions and different climatic conditions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 The figure shows the steps of the spatiotemporal diffusion prediction method for poplar and willow catkin concentration based on deep learning according to the present invention; Figure 2 The diagram illustrates the steps of the method for comprehensively calculating the initial drift concentration of the grid according to the present invention. Figure 3 The diagram illustrates the steps of the method for operating the drift concentration prediction model of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1 , Figure 2 , Figure 3 The deep learning-based method for predicting the spatiotemporal diffusion concentration of poplar and willow catkins, as shown, includes the following steps: Step 1: Collect basic data related to poplar and willow catkins and standardize the data. Obtain key phenological periods and spatial distribution through ground observation, remote sensing inversion and fusion of historical data; identify the distribution areas of poplar and willow trees and calculate their density, classify the regional types, extract digital elevation models, and realize the collection and preprocessing of multi-source data. By using a network of meteorological observation stations, meteorological satellites, and IoT sensing devices, environmental impact data such as wind speed, wind direction, temperature, relative humidity, air pressure, and precipitation are collected in real time. This data is then integrated with ground-based fixed-point observation records, remote sensing vegetation index inversion data, and historical phenological databases to obtain key phenological periods (such as flower bud swelling period, inflorescence emergence period, beginning of flight, peak period, and end period) and their spatial distribution. Data collection on key phenological periods and their spatial distribution was conducted using a comprehensive method that combined ground-based fixed-point observations, remote sensing vegetation index inversion, and integration with historical databases. Within the study area, observation points were set up in layers according to the distribution density of poplar and willow trees (e.g., at least one per 10 km², and in key areas such as parks and green spaces, one per 2 km²). At each observation point, a standard tree was selected (in this embodiment, the standard tree is limited to trees aged 5-15 years and free from diseases). The observations were recorded by taking photos, and observations were conducted once every t hours. The specific dates of the following events were recorded: the flower bud swelling period (flower bud diameter ≥ 3 mm), the inflorescence appearance period (first inflorescence opens), the beginning of the fluffing period (first observation of fluff), the peak of the fluffing period (more than 80% of the standard trees continuously fluff), and the end of the fluffing period (no fluffing observation records for 3 consecutive days). At the same time, the spatial information of the observation point, such as latitude, longitude, and altitude, was recorded. The data was uploaded to the system in real time via a mobile APP.

[0018] High-resolution satellite imagery of GF-1 / GF-2 and NDVI / EVI data from MODIS satellite were used to calculate the time-series curve of vegetation index. Based on the curve characteristics, phenological nodes were extracted (the first significant rise in NDVI / EVI indicates the flower bud swelling stage, the stable period before the peak indicates the inflorescence appearance stage, the first decline indicates the beginning of the flowering stage, the rapid decline stage indicates the peak stage, and the drop to the lowest value of the year indicates the end stage), and the data were converted into a 1km×1km grid spatial distribution map. Collect recent garden monitoring archives, scientific research data, and historical phenological records, and remove abnormal data with a deviation of more than 3 days; integrate the current year's ground observation and remote sensing inversion data with historical data, and perform time-series correction through the linear trend method to finally form accurate key phenological periods and spatial distribution data.

[0019] Based on high-resolution remote sensing images and GIS databases, the distribution areas of poplar and willow trees are identified through object-oriented classification or supervised classification algorithms. Combined with the green space boundary vector data in the GIS database for verification, the number of trees is counted according to a 1km×1km grid, and the distribution density per square kilometer is calculated to extract the distribution density of poplar and willow trees. Spectral and spatial features were extracted using high-resolution remote sensing images (GF-1 / GF-2), and a support vector machine (SVM) supervised classification algorithm was used to divide the regions (e.g., parks, green spaces, roads, and built-up areas). The classification accuracy was ensured by overlaying and correcting the road network and green space planning vector data from the GIS database. Obtain the raw SRTM or ASTERGDEM data, perform noise reduction, depression filling, and smoothing preprocessing using GIS tools, resample to a uniform grid resolution, extract the average elevation information of each grid, and obtain the digital elevation model (DEM).

[0020] All spatial data are uniformly mapped to a 1km×1km grid system. Continuous data (such as elevation and wind speed) are processed using bilinear interpolation, while discrete data (such as region type and number of poplar and willow trees) are processed using nearest neighbor interpolation. This ensures that the grid coordinates of data from different sources are perfectly matched, conforming to the geographical spatial distribution patterns. Step 2: Unify the collected data to the preset grid range and standard time step, quantify the phenological period as a relative time feature, determine the phenological period-specific parameters through a nonlinear optimization algorithm, calculate the phenological effectiveness weight, combine the regional type weight, poplar and willow tree distribution density and temperature and humidity correction function, calculate the initial drift concentration of each grid at each time step through the formula, and integrate it into a structured multidimensional tensor. All data were standardized to a 1km × 1km regular geographic grid. Based on data characteristics (continuous / discrete), bilinear interpolation or nearest neighbor interpolation methods were used for resampling to ensure that all feature values ​​in each grid corresponded precisely in spatial location. A standard time step (e.g., 12 minutes / step) was set, and meteorological time series and phenological period information were transformed to this standardized scale. In particular, phenological periods (e.g., beginning and peak) were quantified as relative time features, and valid flyby periods were marked. Finally, the processed data were integrated into a structured multidimensional tensor (time step × row × column × feature). Missing values, outliers, and inconsistencies in the data were addressed through missing value imputation, outlier handling, and data consistency verification.

[0021] The specific method for quantifying phenological periods (such as the beginning and peak periods) as relative temporal characteristics is as follows: Taking the starting point of the drifting of grid (i, j) as the reference starting point (denoted as...) The current time step t is converted into the cumulative duration relative to the beginning of the drifting period, combined with the duration of the peak period (denoted as t). (Calculated from the peak start / end dates obtained through ground observation and remote sensing inversion in step one), ultimately forming a relative time proportion, which is used to quantify the position of the current time step in the entire drift cycle. The specific formula is shown below: ; in, This represents the relative temporal characteristics of grid (i, j) at time step t; Prepare a standardized historical dataset, including gridded historical phenological period data, corresponding relative time series, and observation / remote sensing inversion data of poplar and willow catkin concentration during the same period (training target). Initialize phenological period characteristic parameters based on prior phenological knowledge. Weight of flying potential ; With the goal of minimizing the error between the model-estimated initial concentration and the historical observed concentration, nonlinear optimization algorithms (such as Levenberg-Marquardt and Bayesian optimization) are used to automatically find the optimal solution on historical data. The output is the optimal set of parameters that minimizes the objective function, i.e., the phenological period-specific parameters.

[0022] Based on relative time characteristics and setting specific parameters for phenological periods, a phenological effectiveness weight is set to quantify the matching degree between the current time step and the phenological period. The specific formula is as follows: ; in, This indicates the phenological stage type of grid (i, j) at time step t (beginning stage = 1, peak stage = 2, end stage = 3, others = 0). Represents the phenological characteristic parameters of grid (i, j) at time step t (e.g.) , , , ), This represents the phenological potential weight of grid (i, j) at time step t. For the Sigmoid function, The slope parameter is derived from key phenological periods and drift concentration correlation data fused from ground-based fixed-point observations, remote sensing inversion, and historical data. It is obtained by fitting the relationship between phenological effectiveness weights and relative time characteristics using a Sigmoid function and minimizing the initial drift concentration estimation bias. This formula transforms phenological periods into continuous variables through relative time characteristics, and then maps them to smooth weights of 0 to 1 using a Sigmoid function, thus achieving a gradual characterization of drift potential from the beginning to the peak to the end of the period.

[0023] Based on field observation data on the release potential and diffusion resistance of willow and poplar catkins, and combined with phenological characteristics and underlying surface influence mechanisms, the weights of different regional types were determined. The true value, the region type weight in this embodiment The actual values ​​are set as follows: Parks / Green Spaces = 1.0, Roads = 0.7, Built-up Areas = 0.3, Others = 0.1; Combining the distribution density of poplar and willow trees within the grid, the weight of their phenological effectiveness, and the weight of their respective region types, a formula for calculating the concentration of drifting pollen is established, as shown below: ; in, This represents the initial drift concentration at time step t, in grid (i, j). The region type weight of grid (i, j) is represented. This represents the distribution density of poplar and willow trees in grid (i, j). This embodiment sets the temperature and humidity correction function (temperature promotes drift, humidity inhibits drift). , where T represents the average temperature of the grid and R represents the relative humidity of the grid.

[0024] Step 3: A three-stage model of spatiotemporal feature extraction, phenological attention fusion, and diffusion prediction is adopted. By defining meteorological coupling and terrain attenuation correction factors, the final drift concentration prediction formula is constructed. The model is trained and optimized using the phenological period weighted mean square error as the loss function. Finally, the model outputs the future time-by-time grid concentration forecast, phenological matching degree heat map, and high concentration risk warning boundary. Based on the structured multidimensional tensor (time step × row × column × feature), the core driving factors are expanded and integrated to form the model input feature set; The features described in this embodiment include: Core concentration-related characteristics: initial drift concentration phenological effectiveness weight Regional type weights Poplar and willow tree distribution density ; Meteorological dynamics: wind speed ,wind direction (Based on the meteorological observation / forecast data collected in step one, the wind direction is converted into azimuth, with due north as 0°, increasing clockwise, such as north wind = 0°, east wind = 90°, south wind = 180°, and west wind = 270°), temperature and humidity correction function. air pressure , ; Topography and Spatial Features: Digital Elevation Model Grid slope (Calculated from DEM), relative time characteristics ; Historical dependency characteristic: the first n time steps , ... (Capturing the temporal continuity of concentration).

[0025] The final input feature tensor dimension is [prediction duration × number of grid rows × number of grid columns × Na] (Na features), ensuring that the model simultaneously captures time series dependencies, spatial correlations, and multi-factor coupling effects.

[0026] The model adopts a three-stage structure of [spatiotemporal feature extraction - phenological attention fusion - diffusion prediction], incorporating the phenological specificity and spatiotemporal diffusion patterns of poplar and willow catkins: A stacked structure of ConvLSTM and 3D convolution is adopted. The ConvLSTM layer (64 hidden dimensions) captures the temporal dependence of concentration diffusion between time steps, and the 3D convolutional layer (kernel_size=3×3×3, stride 1) extracts the diffusion correlation features of the spatial neighborhood grid. At the same time, the spatial heterogeneity of terrain and region type is preserved, and an adaptive attention mechanism is introduced to weight phenological effectiveness. and the distribution density of poplar and willow trees The attention weights are used to weight and enhance the spatiotemporal features, as shown in the following formula: ; in, This represents the coordinates of the neighboring grids of the current grid (i, j). The original spatiotemporal features output by the spatiotemporal feature extraction module (Na-term features in the dimension of the input feature tensor); The model learns the mapping relationship from input features to the relative diffusion increment ratio, using the measured concentration change ratio as the label. The label calculation formula is as follows: The diffusion prediction head (two fully connected layers) outputs with the same dimension as the number of grid cells. The activation function is Tanh, and the output corresponds to the diffusion concentration increment based on the label dimension and physical meaning. , The physical meaning is the proportion of concentration change (a negative increment represents a decrease in concentration, and a positive increment represents an increase in diffusion). Increment of diffusion concentration based on model output , ( For the forecast time step, such as =1, 2, ..., 20, corresponding to minute-by-minute forecasts of future prediction durations), combined with initial concentrations and multi-factor corrections, to construct the final flotation concentration prediction formula: ; in, Indicates a step in the future time. The phenological effectiveness weight, which is coupled with multiple elements, is obtained by coupling the phenological effectiveness weight with meteorological coupling correction factors and topographic attenuation correction factors. In this embodiment, a meteorological coupling correction factor is set by integrating the combined effects of wind speed, wind direction, and humidity. The formula used is... It means that among them Represents the grid (i, j) at future time steps. wind speed, This indicates the angle between the wind direction and the grid diffusion direction. Represents the grid (i, j) at future time steps. The precipitation intensity is used to quantify the impact of future humidity through the predicted precipitation intensity value. a1 and a2 are empirical coefficients based on statistical fitting of historical observation data. By analyzing the correlation between historical wind speed and the increase in willow catkin diffusion, the wind speed term output range is fitted to match the actual diffusion intensity to obtain the a1 value. The measured data of precipitation intensity and concentration decay are fitted to make the precipitation term ( The inhibition effect can be reasonably characterized (e.g., the concentration decreases by about 40% when Pr=1mm / h and by about 92% when Pr=5mm / h, which is consistent with the actual law of willow catkins settling when they encounter water) to obtain a2.

[0027] Grid diffusion refers to the direction of migration and propagation of poplar and willow catkins within a 1km×1km standardized grid. Taking any grid (i, j) as the core, its diffusion direction is the direction from that grid to the eight adjacent grids corresponding to the eight directions (e.g., grid (i, j) pointing to (i+1, j) is due south). These directions are quantified as azimuth angles. If the poplar and willow tree distribution density of the core grid (i, j) is higher than that of a certain neighboring grid... ,but The direction is the diffusion direction (flocs migrate from high-density areas to low-density areas), targeting the core grid. For each potential diffusion direction, a fixed grid diffusion direction azimuth angle is set. such as the core grid Pointing to the neighborhood (North) → 0°, pointing (Northeast) → 45°, pointing (East) → 90°, and for other directions, increase by 45° clockwise. Combining the wind direction azimuth with the determined grid diffusion direction azimuth, the angle between the wind direction and the grid diffusion direction is calculated using the formula... express; By acquiring high-resolution numerical weather prediction (NWP) data (such as WRF, ECMWF, and GRAPES models), whose raw output includes future precipitation forecasts, low-resolution NWP precipitation data is mapped to a 1km×1km target grid using bilinear interpolation / machine learning downscaling (such as random forest and U-Net). This grid is tailored to regional topography and climate characteristics and corrected based on historical precipitation time series patterns. The final output grid (i, j) is then positioned at a future time step. The intensity of precipitation; Considering the hindering effect of elevation on diffusion, a terrain attenuation correction factor is set, and the specific calculation formula is as follows: , Represents a grid Average elevation, The elevation attenuation coefficient is determined by collecting measured data on the diffusion concentration of poplar and willow catkins at different elevation gradients, analyzing the correlation between elevation and concentration attenuation, and aiming to minimize the error between the model-predicted concentration and the measured concentration. The coefficient is fitted using a nonlinear optimization algorithm (such as Levenberg-Marquardt) to finally determine the optimal value that makes the simulation results fit the actual diffusion law.

[0028] It equals the cumulative product of the phenological effectiveness weight, the meteorological coupling correction factor, and the topographic attenuation correction factor; Historical gridded observation data (acquired using the method in step one) and corresponding meteorological / topographical data were used, divided into training, validation, and test sets in a 7:2:1 ratio. The loss function was set using phenological period weighted mean square error (WMSE), and the specific formula is shown below: ; in, This represents the weighted mean squared error loss value during model training. This represents the total number of training samples (the total number of samples across all time steps and all grids). For grid The phenological effectiveness weight at time step t, The relative increment ratio of the model prediction is represented by the AdamW optimizer (learning rate (1e-4), weight decay (1e-5)), and overfitting is prevented by early stopping (if the validation set loss does not decrease after 3 rounds). After the model training is complete, input the weather forecast data for the next 24 hours (wind speed, wind direction, temperature, humidity, etc.), poplar and willow tree distribution / topography data, and phenological extrapolation data. Using the model and formulas from steps two and three, output: Time scale: Time-by-time prediction of drift concentration over future duration; Spatial scale: Spatiotemporal distribution map of a 1km×1km grid; Auxiliary outputs include: a heat map of phenological period matching degree and early warning boundaries for high-concentration (in this embodiment, a floating concentration of ≥100 in the pre-set grid is recorded as a high-concentration grid) risk areas, providing support for prevention and control decisions.

[0029] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0030] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A deep learning-based method for predicting the spatiotemporal diffusion concentration of willow and poplar catkins, characterized in that, include: Step 1: Collect basic data related to poplar and willow catkins and standardize the data. By integrating ground observation, remote sensing inversion and historical data, key phenological periods and spatial distributions are obtained. Poplar and willow distribution areas are identified and density is calculated. Regional types are classified and digital elevation models are extracted to achieve the collection and preprocessing of multi-source data. Observation points were set up in layers according to the distribution density of poplar and willow trees. Photos were taken and observed at set intervals to record the flower bud swelling period, inflorescence appearance period, the beginning, peak, and end of the flowering period, as well as the spatial information of the observation points. The vegetation index time series curve was calculated and phenological nodes were extracted and converted into a grid spatial distribution map. The data were then integrated with the ground observation and remote sensing inversion data of the current year and the historical phenological data after outlier removal and time series correction to form key phenological periods and spatial distribution data. Step 2: Unify the collected data to the preset grid range and standard time step, quantify the phenological period as a relative time feature, determine the phenological period-specific parameters through a nonlinear optimization algorithm, calculate the phenological effectiveness weight, combine the regional type weight, poplar and willow tree distribution density and temperature and humidity correction function, calculate the initial drift concentration of each grid at each time step through the formula, and integrate it into a structured multidimensional tensor. Step 3: A three-stage model of spatiotemporal feature extraction, phenological attention fusion, and diffusion prediction is adopted. By defining meteorological coupling and terrain attenuation correction factors, the final drift concentration prediction formula is constructed. The model is trained and optimized using the phenological period weighted mean square error as the loss function. Finally, the model outputs the future time-by-time grid concentration forecast, phenological matching degree heat map, and high concentration risk warning boundary.

2. The method for predicting the spatiotemporal diffusion concentration of willow catkins based on deep learning according to claim 1, characterized in that, The specific methods for identifying the distribution areas and calculating the density of poplar and willow trees, classifying the area types, and extracting digital elevation models are as follows: Based on high-resolution remote sensing images and GIS databases, the distribution areas of poplar and willow trees are identified through object-oriented classification or supervised classification algorithms. Combined with the green space boundary vector data in the GIS database for verification, the number of trees is counted by grid and the distribution density is calculated to extract the distribution density of poplar and willow trees. Spectral and spatial features were extracted from remote sensing images, and a support vector machine supervised classification algorithm was used to divide the monitoring area into different types. The classification of monitoring area types was achieved by overlaying and correcting the road network and green space planning vector data from the GIS database. Obtain the raw SRTM or ASTERGDEM data, and after preprocessing such as denoising, filling depressions, and smoothing, resample to a uniform grid resolution to extract the average elevation information of each grid.

3. The method for predicting the spatiotemporal diffusion concentration of willow catkins based on deep learning according to claim 1, characterized in that, The quantification of phenological periods is a relative temporal characteristic, and the specific method is as follows: Taking the starting point of the drifting of grid (i, j) as the reference starting point, denoted as Convert the current time step t to the cumulative duration relative to the beginning of the drifting period, and combine it with the duration of the peak period, denoted as . Using the formula It represents the relative time characteristic and is used to quantify the position of the current time step in the entire drift cycle.

4. The method for predicting the spatiotemporal diffusion concentration of willow catkins based on deep learning according to claim 1, characterized in that, The specific method for calculating the phenological effectiveness weight is as follows: Based on relative time characteristics and setting specific parameters for phenological periods, using formulas A phenological effectiveness weight is set to quantify the degree of matching between the current time step and the phenological period. This represents the phenological period type of grid (i, j) at time step t. This represents the phenological characteristic parameters of grid (i, j) at time step t. This represents the phenological potential weight of grid (i, j) at time step t. For the Sigmoid function, The slope parameter is derived from key phenological periods and drift concentration correlation data fused from ground-based fixed-point observations, remote sensing inversion, and historical data. It is obtained by fitting the Sigmoid function relationship between phenological effectiveness weights and relative time characteristics and minimizing the initial drift concentration estimation bias.

5. The method for predicting the spatiotemporal diffusion concentration of willow catkins based on deep learning according to claim 1, characterized in that, Combining region type weights, poplar and willow tree distribution density, and temperature and humidity correction functions, the initial spore concentration at each time step for each grid is calculated using a formula. The specific method is as follows: Based on field observation data on the release potential and diffusion resistance of willow and poplar catkins, and combined with phenological characteristics and underlying surface influence mechanisms, the weights of different regional types were determined. The true value, combined with the distribution density of poplar and willow trees within the grid, the weight of the phenological effectiveness they represent, and the weight of the region type, is used using the formula Indicates the concentration of drift, where, This represents the initial drift concentration at time step t, in grid (i, j). The region type weight of grid (i, j) is represented. This represents the distribution density of poplar and willow trees in grid (i, j). The temperature and humidity correction function is obtained using the formula. Where T represents the average temperature of the grid and R represents the relative humidity of the grid. and These represent the temperature coefficient and humidity coefficient, respectively, obtained through quantitative calibration based on the meteorological response mechanism of poplar and willow catkin drift and historical observation data.

6. The method for predicting the spatiotemporal diffusion concentration of willow catkins based on deep learning according to claim 1, characterized in that, A three-stage model of spatiotemporal feature extraction, phenological attention fusion, and diffusion prediction is adopted. The specific method is as follows: A stacked structure of ConvLSTM and 3D convolution is adopted. The ConvLSTM layer captures the temporal dependence of concentration diffusion between time steps, while the 3D convolutional layer extracts the diffusion correlation features of the spatial neighborhood grid. At the same time, the spatial heterogeneity of terrain and region type is preserved, and an adaptive attention mechanism is introduced to weight phenological effectiveness. and the distribution density of poplar and willow trees For attention weights, use the formula Weighted enhancement of spatiotemporal features, among which, This represents the coordinates of the neighboring grids of the current grid (i, j). The original spatiotemporal features are output by the spatiotemporal feature extraction module.

7. The method for predicting the spatiotemporal diffusion concentration of willow catkins based on deep learning according to claim 1, characterized in that, The meteorological coupling factor and the topographic attenuation correction factor are defined using the following method: By integrating the combined effects of wind speed, wind direction, and humidity, a meteorological coupling correction factor is set. This embodiment utilizes the formula... It means that among them Represents the grid (i, j) at future time steps. wind speed, This indicates the angle between the wind direction and the grid diffusion direction. Represents the grid (i, j) at future time steps. The precipitation intensity, a1 and a2 are statistical fitting empirical coefficients based on historical observation data. By analyzing the correlation between historical wind speed and the diffusion increment of poplar and willow catkins, the wind speed term output range is matched with the actual diffusion intensity to obtain the value of a1. a2 is obtained by fitting the measured data of precipitation intensity and concentration decay. Considering the hindering effect of elevation on diffusion, a terrain attenuation correction factor is set, and the specific calculation formula is as follows: , Represents a grid Average elevation, The elevation attenuation coefficient was determined by collecting measured data on the diffusion concentration of poplar and willow catkins at different elevation gradients, analyzing the correlation between elevation and concentration attenuation, and using a nonlinear optimization algorithm to fit the coefficients with the goal of minimizing the error between the model-predicted concentration and the measured concentration.

8. The method for predicting the spatiotemporal diffusion concentration of willow catkins based on deep learning according to claim 7, characterized in that, The angle between the wind direction and the grid diffusion direction specifically includes: Grid diffusion refers to the direction of poplar and willow catkin migration and propagation between standardized grids. Taking any grid (i, j) as the core, its diffusion direction is the direction from that grid to the corresponding direction of the adjacent grid, which is quantified as azimuth angle. If the poplar and willow tree distribution density of the core grid (i, j) is higher than that of a certain neighboring grid... ,but The direction is the diffusion direction, targeting the core mesh. For each potential diffusion direction, a fixed grid diffusion direction azimuth angle is set. Combining the wind direction azimuth and the determined grid diffusion direction azimuth, the angle between the wind direction and the grid diffusion direction is calculated using the formula... express.

9. The method for predicting the spatiotemporal diffusion concentration of willow catkins based on deep learning according to claim 1, characterized in that, The final formula for predicting drift concentration is constructed using the following method: The model learns the mapping relationship from input features to the relative diffusion increment ratio, using the measured concentration change ratio as the label. The output dimension of the diffusion prediction head is consistent with the number of grid cells, and the activation function is Tanh. The output is the diffusion concentration increment. , The physical meaning of is the proportion of concentration change; Increment of diffusion concentration based on model output Combining initial concentration and multi-factor correction, using the formula Construct the final formula for predicting drift concentration, where, Indicates a step in the future time. The phenological effectiveness weight of multi-element coupling is equal to the cumulative product of the phenological effectiveness weight, the meteorological coupling correction factor, and the topographic attenuation correction factor.

Images