A city isoprene concentration prediction method based on deep transfer learning

Abstract

The present application belongs to the technical field of ecological environment monitoring, and particularly relates to a city isoprene concentration prediction method based on deep migration learning, comprising: acquiring multi-source environmental data; performing principal component analysis on the vegetation data; constructing a deep learning model with physical constraints; pre-training the deep learning model using isoprene concentration observation data of multiple source regions and corresponding model input variables, to obtain a pre-trained model; fine-tuning the pre-trained model using part of the isoprene concentration observation data of a target city region, and outputting isoprene concentration prediction results of the target city region. The present application solves the technical problems of insufficient prediction accuracy of city area isoprene concentration, poor physical consistency of prediction results, and weak model cross-region generalization ability in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of ecological environment monitoring technology, and in particular relates to a method for predicting urban isoprene concentration based on deep transfer learning. Background Technology

[0002] Isoprene is one of the most abundant biogenic volatile organic compounds emitted into the atmosphere. It participates in the formation of ozone and secondary organic aerosols in atmospheric chemical processes, and has a significant impact on regional air quality and climate change. Current estimates of isoprene concentration or emissions usually rely on biogenic emission models such as MEGAN or chemical transport models, but these methods have significant limitations in urban-scale characterization.

[0003] First, the vegetation type data that traditional models rely on is updated infrequently, making it difficult to reflect the dynamic changes in urban greening and land use. Second, due to the complexity of urban surface types, medium-resolution satellite products struggle to accurately identify vegetation distribution at the block scale, leading to an underestimation of emissions.

[0004] Furthermore, chemical transport models are computationally expensive and highly sensitive to the quality of input data, making it difficult to achieve high-precision predictions in data-scarce regions. Although machine learning methods have been applied in recent years, most models rely on data from only a single region, resulting in limited generalization ability. Moreover, due to the lack of physical or mechanistic constraints, they are prone to producing predictions that contradict actual emission mechanisms. Summary of the Invention

[0005] To address the technical problems of insufficient accuracy in predicting isoprene concentration in urban areas, poor physical consistency of prediction results, and weak cross-regional generalization ability of the model in the existing technologies, this invention proposes an urban isoprene concentration prediction method based on deep transfer learning.

[0006] This invention is achieved through the following technical solution:

[0007] A method for predicting urban isoprene concentration based on deep transfer learning includes the following steps:

[0008] S1. Acquire multi-source environmental data, including vegetation data retrieved from satellite remote sensing, emission inventory data, and meteorological reanalysis data, and perform spatial and temporal resolution alignment processing on the multi-source environmental data to obtain model input variables.

[0009] S2. Perform principal component analysis on the vegetation data to extract the principal components that characterize the regional vegetation change features as vegetation feature variables; for the emission inventory data, use the time allocation coefficient to convert it into emission input data with a target time resolution.

[0010] S3. Construct a deep learning model with physical constraints. The deep learning model includes a feature extraction layer and an output prediction layer. Add a monotonicity constraint term to the loss function of the deep learning model so that the first-order partial derivative of the isoprene concentration output by the model with respect to the vegetation feature variables and the emission input data remains positive.

[0011] S4. The deep learning model is pre-trained using isoprene concentration observation data from multiple source regions and the corresponding model input variables to obtain a pre-trained model.

[0012] S5. Fine-tune the pre-trained model using partial isoprene concentration observation data of the target city area, input the model input variables of the target city area into the fine-tuned model, and output the isoprene concentration prediction result of the target city area.

[0013] Furthermore, the vegetation data includes leaf area index and normalized difference vegetation index; the emission inventory data includes black carbon emission data from transportation sources; and the meteorological reanalysis data includes air temperature, solar radiation, and wind speed.

[0014] Alignment processing of multi-source environmental data with spatial and temporal resolution includes:

[0015] The above multi-source environmental data are resampled, spatially reprojected, and region extracted to ensure that different data maintain consistency in temporal and spatial resolution.

[0016] Furthermore, principal component analysis is performed on the vegetation data to extract principal components characterizing regional vegetation change features as vegetation feature variables, including:

[0017] Calculate the covariance matrix of the leaf area index and the normalized vegetation index and perform eigenvalue decomposition.

[0018] Principal components that can characterize vegetation change are extracted, and the principal component with the largest contribution rate to explaining variance is selected as the vegetation characteristic variable.

[0019] Furthermore, the step of converting the emission inventory data into emission input data with a target time resolution using a time allocation factor includes:

[0020] Obtain the monthly traffic source allocation coefficients obtained from statistics of different countries or regions;

[0021] The annual black carbon emissions from traffic sources are time-weighted and redistributed using the monthly traffic source allocation coefficient, and then converted into monthly black carbon emissions from traffic sources to obtain the emission input data with higher time resolution.

[0022] Furthermore, the deep learning model employs an artificial neural network with a residual connection structure;

[0023] The feature extraction layer is used to learn the potential feature relationship between the model input variables and the isoprene concentration, and the output prediction layer is used to generate the isoprene concentration prediction result.

[0024] Furthermore, a monotonicity constraint term is added to the loss function of the deep learning model to ensure that the first-order partial derivatives of the isoprene concentration output by the model with respect to the vegetation characteristic variables and the emission input data remain positive. This is achieved through the following formula:

[0025] ;

[0026] ;

[0027] in, This indicates the predicted concentration of isoprene. Indicates vegetation-related input variables, Indicates emission-related input variables;

[0028] The loss function includes a data fitting loss term, a monotonicity constraint term, and a structure regularization loss term.

[0029] Furthermore, the step of pre-training the deep learning model using isoprene concentration observation data from multiple source regions and the corresponding model input variables to obtain a pre-trained model includes:

[0030] The model weights and biases are iteratively updated by adjusting the model structure and optimizing the model hyperparameters, which include the learning rate, optimizer type, and Dropout ratio.

[0031] The model with the highest coefficient of determination and the smallest root mean square error is selected as the evaluation criterion, and its network parameters are saved as the pre-trained model.

[0032] Furthermore, the fine-tuning of the pre-trained model using partial isoprene concentration observation data from the target urban area includes:

[0033] Initialize the weights and biases in the artificial neural network using the network parameters of the pre-trained model;

[0034] By fixing the parameters of the output prediction layer and retraining only the parameters of the feature extraction layer, the model can be adapted to the vegetation distribution characteristics, emission structure and meteorological conditions of the target urban area.

[0035] The beneficial effects of this invention are:

[0036] (1) This invention integrates satellite remote sensing data, emission inventory data and meteorological reanalysis data to construct a multi-source data-driven isoprene concentration prediction model, which improves the accuracy of isoprene concentration prediction in urban areas.

[0037] (2) This invention does not rely on vegetation emission factor parameters in traditional biological emission models, but learns the variation law of isoprene concentration directly from the observation data through machine learning methods, thereby reducing the impact of emission factor uncertainty on the estimation results;

[0038] (3) By introducing physical constraints based on emission mechanisms, this invention enables the model prediction results to meet the physical law that isoprene concentration increases with the increase of vegetation and emissions, thereby improving the physical consistency and reliability of the model.

[0039] (4) The present invention adopts a transfer learning strategy, which enables the model to have strong generalization ability by pre-training on observation data in multiple regions, and can achieve high-precision prediction in regions with less observation data.

[0040] (5) Compared with traditional methods that rely on chemical transport models or biological emission models, the required input data is easy to obtain, the computational efficiency is high, and it can achieve rapid prediction of isoprene concentration at the urban scale. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0042] Figure 1 This is a schematic diagram of the process framework of a method for predicting urban isoprene concentration based on deep transfer learning proposed in an embodiment of the present invention.

[0043] Figure 2 This is a schematic diagram of the deep learning model structure of a method for predicting urban isoprene concentration based on deep transfer learning proposed in an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of a terminal device for a method for predicting urban isoprene concentration based on deep transfer learning, as proposed in an embodiment of the present invention.

[0045] Figure 4 This is a schematic diagram of a readable storage medium for a method for predicting urban isoprene concentration based on deep transfer learning, as proposed in an embodiment of the present invention.

[0046] In the diagram, 200 is the terminal device, 210 is the memory, 211 is the RAM, 212 is the cache, 213 is the ROM, 214 is the program / utility, 215 is the program module, 220 is the processor, 230 is the bus, 240 is the external device, 250 is the I / O interface, 260 is the network adapter, and 300 is the program product. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0048] Example 1

[0049] refer to Figure 1 This embodiment discloses a method for predicting urban isoprene concentration based on deep transfer learning. This method integrates satellite remote sensing vegetation information, emission inventory data, and meteorological reanalysis data, and achieves urban-scale isoprene concentration prediction through a deep transfer learning model that incorporates physical constraints. Figure 1 middle Which is the hidden layer? This represents the number of neurons in each layer. These are the intermediate features or hidden representations learned by each neuron in the hidden layer.

[0050] Unlike traditional methods that rely on biological emission models, this embodiment does not require the introduction of vegetation emission factor parameters. Instead, it uses machine learning methods to automatically learn the variation pattern of isoprene concentration from multi-source environmental data and observation data. This embodiment will be explained in detail using a target city as an example.

[0051] In the input data acquisition and preprocessing stage of the model, vegetation information, emission inventory data, and meteorological reanalysis data retrieved from satellite remote sensing were comprehensively acquired first. Specifically, the vegetation information included leaf area index and normalized difference vegetation index, sourced from MODIS satellite products; the emission data included traffic-related black carbon emission inventory data from the EDGAR Global Atmospheric Research Emissions Database; and the meteorological data came from ERA5 meteorological reanalysis data provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), covering meteorological elements such as temperature, solar radiation, wind speed, precipitation, and boundary layer height.

[0052] The detailed time span, temporal resolution, spatial coverage, and spatial resolution information of the input variables used to predict isoprene concentration in this embodiment are shown in Table 1: Input variables Time Coverage Time resolution Spatial coverage Spatial resolution Leaf area index 1990-2023 8 days worldwide 0.05° Normalized Difference Vegetation Index 1990-2023 half a month worldwide 0.083° Black carbon emissions from transportation sources 1990-2023 1 year worldwide 0.1° 2 m temperature 1990-2023 Hour worldwide 0.1° Sun radiation down to the Earth's surface 1990-2023 Hour worldwide 0.25° Soil moisture 1990-2023 Hour worldwide 0.1° relative humidity 1990-2023 Hour worldwide 0.1° Surface pressure 1990-2023 Hour worldwide 0.1° 10 m zonal wind 1990-2023 Hour worldwide 0.1° 10 m meridional wind 1990-2023 Hour worldwide 0.1° Evaporation caused by vegetation transpiration 1990-2023 Hour worldwide 0.1° Boundary layer height 1990-2023 Hour worldwide 0.25° Total rainfall 1990-2023 Hour worldwide 0.1°

[0053] Table 1. Input variable information used for predicting isoprene concentration.

[0054] Because the multi-source data obtained above have inherent differences in spatial and temporal resolution, they must undergo rigorous spatial reprojection and resampling operations. Simultaneously, data within the latitude and longitude grid of the corresponding target study area stations must be accurately extracted. After this processing, data from different sources are uniformly aligned to a daily temporal resolution and a spatial resolution of 0.1°, thus obtaining input variables that can be directly used for training the underlying network. In the vegetation feature construction stage, considering that both leaf area index (LAI) and normalized vegetation index (NVI) can characterize vegetation growth and are strongly correlated with each other, to avoid information redundancy and the curse of dimensionality, this embodiment uses principal component analysis (PCA) to perform deep coupling analysis between the two. By calculating the covariance matrix between them and performing eigenvalue decomposition, principal components that can highly characterize vegetation change characteristics are extracted. The principal component with the largest contribution to explaining variance is selected as the final model input variable. This effectively reduces variable redundancy and significantly improves the computational efficiency and robustness of the model.

[0055] To address the inherent limitation of low temporal resolution in the annual emissions data from transportation sources provided by the EDGAR emissions inventory, this embodiment introduces monthly black carbon allocation coefficients from transportation sources obtained from statistics of different countries or regions. These coefficients are used to weight and redistribute annual emissions over time, accurately converting them into monthly emissions. This provides the model with input driving data of higher temporal resolution.

[0056] At the network architecture level, this embodiment constructs a deep transfer learning model with physical constraints based on the PyTorch framework, referencing... Figure 2 The main body of the model employs an artificial neural network with residual connections to enhance its ability to express highly nonlinear features of input variables and effectively prevent gradient vanishing during training of deep networks, thereby improving stability. Logically, the model is divided into two parts:

[0057] One is the feature extraction layer, which is specifically used to learn the deep, latent, implicit feature relationships between input variables such as vegetation, emissions, and meteorology and isoprene concentration;

[0058] The second is the output prediction layer, which is responsible for mapping the feature vectors to generate the final scalar result of isoprene concentration.

[0059] Before formal training, the system performed reasonable random initialization configurations for all weights and bias parameters in the neural network. To ensure that the data-driven model does not violate objective scientific laws, the implementation example rigidly introduced physical constraints based on emission mechanisms during the model training optimization process. Since it is generally recognized in the physics community that isoprene emissions are positively correlated with vegetation cover and emission intensity, a monotonicity constraint mechanism was embedded in the model's total loss function, forcing the first-order partial derivatives of the isoprene concentration output by the model with respect to both vegetation-related and emission-related input variables to remain positive. The specific mathematical expression for the partial derivatives is as follows:

[0060] ;

[0061] ;

[0062] in, This represents the predicted concentration of isoprene given by the model. This represents the input variables representing the vegetation-related features after fusion processing. This represents the emission-related input variables after time redistribution. This constraint effectively prevents the model from misfitting to noise, ensuring consistency between the predicted results and the actual physical mechanisms of vegetation emissions. Based on the above constraint logic, the mathematical expression of the total loss function optimized during backpropagation in model training is as follows:

[0063]

[0064] In the above total loss function, the data loss term represents the degree of data fit. The calculation formula is:

[0065]

[0066] The monotonic partial derivatives of the physical constraint loss term representing the mechanism cognition The calculation formula is:

[0067]

[0068] Within this constraint, the sign function applied to the derivative result. The definition of is:

[0069]

[0070] To prevent the network from overfitting to the training set samples, a structure regularization loss term is used to smooth the network weights. Defined as:

[0071]

[0072] In all the above formulas, and These are the trade-off hyperparameters used to balance the magnitudes of various losses. The unified representation refers to the total number of training samples within a batch. Indicates the first A sample index, Indicates the first The true observed value of each sample Indicates the measured values ​​of the model. Representing traffic-related variables, express right The partial derivatives, Indicates the sensitivity of traffic variables. This represents the total number of learnable network layers in the residual neural network. Indicates the first Layer weight parameters, Indicates the first Layer bias parameters.

[0073] After establishing the model and objective function, the system enters the pre-training and transfer fine-tuning process. Due to the limited observation data in the target area, the system first uses hourly isoprene concentration observation data collected over a long period of time in multiple source area cities with different climate and ecological characteristics as the source domain for sufficient pre-training.

[0074] At this stage, the system removes the data of the target city and, in conjunction with the high-dimensional input variables obtained from the aforementioned preprocessing, uses mini-batch gradient descent to continuously adjust the hidden layer structure of the model and iteratively optimize the network hyperparameters, including the learning rate and Dropout ratio, to update the weights and biases. After the iteration, the network weight snapshot with the highest coefficient of determination and the smallest root mean square error is selected as the base of the pre-trained model and saved.

[0075] Subsequently, in the model transfer phase, the local residual network was initialized using the parameter matrix of the pre-trained model. By introducing trace amounts of isoprene observation data from the target region, the network feature layers were fine-tuned at a low learning rate. This enabled the deep network to effectively retain its generalization feature extraction capabilities while adapting to the spatial specificity of vegetation distribution, emission source structure, and local meteorological conditions in the target city. This achieved high-dimensional spatial and temporal accuracy prediction of isoprene concentration in the target city area while significantly saving computing power and data acquisition costs.

[0076] Example 2

[0077] Based on the prediction method of Example 1, this example quantitatively backtracks and estimates the long-term isoprene concentration in the target city during the summers of 1990 to 2023, and performs rigorous model evaluation. The verification results show that the coupled physical constraint transfer model of this invention exhibits excellent prediction performance, with a determination coefficient of 0.76 on the test set. Its advantages are extremely significant compared to various traditional statistical regression models and conventional machine learning models. A detailed comparison of the specific performance evaluations of each model for the target city's isoprene concentration is shown in Table 2. Model Coefficient of determination Mean Absolute Error Mean square error Root mean square error This invention model 0.76 0.169 0.065 0.255 Residual Neural Network 0.74 0.170 0.069 0.263 XGBoost 0.62 0.249 0.103 0.320 Random Forest 0.58 0.250 0.112 0.335 Multilayer perceptron 0.57 0.240 0.117 0.341 GBDT 0.54 0.264 0.122 0.350 SVM 0.50 0.262 0.135 0.367 Multiple linear regression 0.34 0.329 0.176 0.419

[0078] Table 2. Performance comparison of different prediction models in predicting isoprene concentration in target cities.

[0079] Further analysis of the predicted sequence reveals that, in recent decades, driven by the continuous improvement of the city's green area and quality, as well as external meteorological factors such as regional and even global climate warming, the intensity of vegetation-dominated biogenic emissions in the target city has shown a gradually increasing trend over time. This has led to a steady increase in the overall concentration of isoprene base in the troposphere of the region, with an average annual growth rate approaching 18.1 pptv. Monitoring of key characterization variables in the model input data shows that the measured temperature and extracted normalized vegetation index features in the target area also exhibit a continuous upward trend consistent with this growth rate.

[0080] Compared to bio-sources, the region’s transport-related carbon emissions have shown a significant precipitous decline since 1998 due to optimized transport infrastructure, suggesting that the historical fluctuations in the concentration of this particular chemical substance may have become less significant due to anthropogenic transport emissions.

[0081] To verify this statistical hypothesis, this method analyzed the Pearson correlation between traffic emissions and atmospheric benzene concentration. The results showed that there was only a low correlation between the two (R = 0.28). This quantitative data further confirms that the absolute contribution of mobile anthropogenic emission sources such as traffic to the total atmospheric isoprene concentration in this region is quite limited.

[0082] To break down the long-standing black-box barrier of deep learning, this embodiment also successfully introduced the SHapley Additive exPlanations (SHAP) interpretability attribution analysis framework for the trained deep transfer learning model. The analysis results of SHAP values ​​clearly show that the long-term positive trend of isoprene concentration has a very high statistical correlation with the marginal contribution value of SHAP of vegetation index in the input variable (R = 0.95), while its correlation with the temperature variable alone is at a secondary medium confidence level (R = 0.63).

[0083] Using ablation experiments with fixed target input variables, the system accurately isolated and quantified the independent contributions of major driving factors to isoprene concentration. Differential comparison tests were performed across different historical time slices. The experimental results irrefutably demonstrate that the presence of high-density green spaces within cities is the absolute key factor dominating isoprene concentration fluctuations in the target city: in the defined past 17 years and the initial 17-year baseline period, the ecological expansion of urban green spaces directly contributed to an increase of approximately 290 pptv in isoprene concentration. In contrast, the increase in temperature due to global warming contributed only about 51 pptv; and the numerical fluctuations caused by traffic emissions, including black carbon, are negligible due to their extremely low weight.

[0084] A more intuitive system dynamic simulation indirectly proves the above conclusion: if the feature vector matrix representing urban green space is forcibly locked as a constant in the environmental model simulation, the statistical coefficient of variation of the predicted annual average isoprene concentration will show a dramatic decrease of approximately 10%, 12.0%, and 6.0% compared to the scenarios where climate warming factors are locked in unchanged, and where traffic emissions are locked in unchanged, respectively. This series of detailed data comparisons and simulations further highlights and anchors the decisive role of urban green vegetation in the generation, change, and long-term trend evolution of urban isoprene concentration.

[0085] Overall, the end-to-end methodology disclosed and deployed in this embodiment not only bridges the regional data gap and achieves extremely high prediction accuracy through parameter migration and gradient partial derivative physical constraints, but also demonstrates scientific applicability and high-level environmental engineering reliability with great potential for widespread application in capturing the evolution of long-term climate characteristics and analyzing the intrinsic causes of urban-level biogenic volatile organic compound concentrations.

[0086] Example 3

[0087] refer to Figure 3 Based on Example 1, this example proposes a terminal device for predicting urban isoprene concentration based on deep transfer learning. The terminal device 200 includes at least one memory 210, at least one processor 220, and a bus 230 connecting different platform systems.

[0088] The memory 210 may include a readable medium in the form of volatile memory, such as RAM 211 and / or cache memory 212, and may further include ROM 213.

[0089] The memory 210 also stores a computer program that can be executed by the processor 220, causing the processor 220 to perform any of the above-described applications of the deep transfer learning-based urban isoprene concentration prediction method in this application. The specific implementation and technical effects are consistent with those described in the above-described application embodiments, and some details will not be repeated here. The memory 210 may also include a program / utility 214 having a set (at least one) of program modules 215. Such program modules include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment.

[0090] Accordingly, processor 220 can execute the aforementioned computer program, as well as executable program / utility 214.

[0091] Bus 230 can represent one or more of several types of bus structures, including a memory bus or memory controller, peripheral bus, graphics acceleration port, processor, or a local bus using any of the various bus structures.

[0092] Terminal device 200 can also communicate with one or more external devices 240, such as keyboards, pointing devices, Bluetooth devices, etc., and with one or more devices capable of interacting with it, and / or with any device that enables it to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication can be performed via I / O interface 250. Furthermore, terminal device 200 can communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapter 260. Network adapter 260 can communicate with other modules of terminal device 200 via bus 230. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with terminal device 200, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0093] Example 4

[0094] This embodiment proposes a readable storage medium for a method for predicting urban isoprene concentration based on deep transfer learning. The computer-readable storage medium stores instructions that, when executed by a processor, implement any of the aforementioned methods for predicting urban isoprene concentration based on deep transfer learning. The specific implementation method and the technical effects achieved are consistent with those described in the embodiments of the above applications, and some details will not be repeated.

[0095] Figure 4 The present embodiment illustrates a program product 300 for implementing the above-described applications. This product may employ a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product 300 of the present invention is not limited thereto. In this embodiment, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device. The program product 300 may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0096] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. Program code for performing operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on a user computing device, partially on a user device, as a standalone software package, partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to user computing devices via any type of network, including local area networks (LANs) or wide area networks (WANs), or they can be connected to external computing devices (e.g., via the Internet using an Internet service provider).

[0097] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting urban isoprene concentration based on deep transfer learning, characterized in that, Includes the following steps: S1. Acquire multi-source environmental data, including vegetation data retrieved from satellite remote sensing, emission inventory data, and meteorological reanalysis data, and perform spatial and temporal resolution alignment processing on the multi-source environmental data to obtain model input variables. S2. Perform principal component analysis on the vegetation data to extract the principal components that characterize the regional vegetation change features as vegetation feature variables; for the emission inventory data, use the time allocation coefficient to convert it into emission input data with a target time resolution. S3. Construct a deep learning model with physical constraints. The deep learning model includes a feature extraction layer and an output prediction layer. Add a monotonicity constraint term to the loss function of the deep learning model so that the first-order partial derivative of the isoprene concentration output by the model with respect to the vegetation feature variables and the emission input data remains positive. S4. The deep learning model is pre-trained using isoprene concentration observation data from multiple source regions and the corresponding model input variables to obtain a pre-trained model. S5. Fine-tune the pre-trained model using partial isoprene concentration observation data of the target city area, input the model input variables of the target city area into the fine-tuned model, and output the isoprene concentration prediction result of the target city area.

2. The method for predicting urban isoprene concentration based on deep transfer learning according to claim 1, characterized in that, The vegetation data includes leaf area index and normalized difference vegetation index; the emission inventory data includes black carbon emission data from transportation sources; the meteorological reanalysis data includes air temperature, solar radiation, and wind speed. Alignment processing of multi-source environmental data with spatial and temporal resolution includes: The above multi-source environmental data are resampled, spatially reprojected, and region extracted to ensure that different data maintain consistency in temporal and spatial resolution.

3. The method for predicting urban isoprene concentration based on deep transfer learning according to claim 2, characterized in that, Principal component analysis was performed on the vegetation data to extract principal components characterizing regional vegetation change as vegetation feature variables, including: Calculate the covariance matrix of the leaf area index and the normalized vegetation index and perform eigenvalue decomposition. Principal components that can characterize vegetation change are extracted, and the principal component with the largest contribution rate to explaining variance is selected as the vegetation characteristic variable.

4. The method for predicting urban isoprene concentration based on deep transfer learning according to claim 2, characterized in that, The process of converting the emission inventory data into emission input data with a target time resolution using a time allocation factor includes: Obtain the monthly traffic source allocation coefficients obtained from statistics of different countries or regions; The annual black carbon emissions from traffic sources are time-weighted and redistributed using the monthly traffic source allocation coefficient, and then converted into monthly black carbon emissions from traffic sources to obtain the emission input data with higher time resolution.

5. The method for predicting urban isoprene concentration based on deep transfer learning according to claim 1, characterized in that, The deep learning model employs an artificial neural network with a residual connection structure. The feature extraction layer is used to learn the potential feature relationship between the model input variables and the isoprene concentration, and the output prediction layer is used to generate the isoprene concentration prediction result.

6. The method for predicting urban isoprene concentration based on deep transfer learning according to claim 1, characterized in that, A monotonicity constraint term is added to the loss function of the deep learning model to ensure that the first-order partial derivatives of the isoprene concentration output by the model with respect to the vegetation characteristic variables and the emission input data remain positive. This is achieved through the following formula: ； ； in, This indicates the predicted concentration of isoprene. Indicates vegetation-related input variables, Indicates emission-related input variables; The loss function includes a data fitting loss term, a monotonicity constraint term, and a structure regularization loss term.

7. The method for predicting urban isoprene concentration based on deep transfer learning according to claim 1, characterized in that, The process of pre-training the deep learning model using isoprene concentration observation data from multiple source regions and the corresponding model input variables to obtain a pre-trained model includes: The model weights and biases are iteratively updated by adjusting the model structure and optimizing the model hyperparameters, which include the learning rate, optimizer type, and Dropout ratio. The model with the highest coefficient of determination and the smallest root mean square error is selected as the evaluation criterion, and its network parameters are saved as the pre-trained model.

8. The method for predicting urban isoprene concentration based on deep transfer learning according to claim 5, characterized in that, The fine-tuning of the pre-trained model using partial isoprene concentration observation data from the target urban area includes: Initialize the weights and biases in the artificial neural network using the network parameters of the pre-trained model; By fixing the parameters of the output prediction layer and retraining only the parameters of the feature extraction layer, the model can be adapted to the vegetation distribution characteristics, emission structure and meteorological conditions of the target urban area.

Images