Global wildfire prediction method and device based on meteorological data federated learning

Abstract

The invention provides a global wildfire prediction method and device based on meteorological data federated learning, and the method comprises the steps: obtaining to-be-predicted meteorological privacy data, inputting the meteorological privacy data into a trained meteorological data federated learning network model, and obtaining a global wildfire prediction probability result; the meteorological data federated learning network model is composed of a plurality of local clients and a global server, and the plurality of local clients and the global server adopt a time sequence cross network with the same architecture. The federated learning framework is introduced into the wildfire prediction task for the first time, the network model adaptive to the global data is obtained on the premise that the local client and the global server do not share data, and the bottleneck that traditional meteorological data cannot be completely shared is broken through. And through federal learning, the data privacy of each participating node can be protected, the weather information in the global range can be integrated, and the generalization ability of the prediction model can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer deep learning technology and relates to a wildfire prediction method and device, and in particular to a global wildfire prediction method and device based on meteorological data federated learning. Background Art

[0002] Wildfires are one of the most destructive natural disasters worldwide, profoundly impacting ecosystems, economic development, and human life. In recent years, climate change has triggered an increase in extreme weather events, significantly increasing the frequency and intensity of wildfires. Wildfire occurrence is influenced by a variety of meteorological factors, including temperature, precipitation, wind speed, and humidity. Furthermore, non-meteorological factors such as topography and vegetation cover also influence the spread and evolution of wildfires.

[0003] Existing wildfire prediction methods typically rely on regional meteorological observation data and numerical models. However, due to the non-independent and identically distributed nature of the data, the high dynamic nature of meteorological data, and limitations on data sharing, prediction accuracy and efficiency still need to be improved. Specifically, data unevenness refers to the differences in the density and data quality of meteorological observation networks in different regions of the world, and the different types and distribution of meteorological data held by each region and department. High dynamics means that the formation and development of wildfires are influenced by multiple complex nonlinear factors, and traditional models have difficulty capturing their dynamic changes. Data sharing limitations mean that meteorological data often has regional sensitivity or privacy restrictions, making it difficult for different regions to directly share data to support more comprehensive model training.

[0004] Federated learning is a distributed deep learning framework that allows multiple participants to collaboratively train a global model without sharing data. This framework addresses data privacy and sharing restrictions and has been widely used in data-sensitive fields such as healthcare and finance. In the meteorological field of global wildfire forecasting, the introduction of federated learning can effectively protect data privacy, integrate multi-source data, and dynamically update models.

[0005] To address the problem of global wildfire prediction, Ioannis Prapas et al. (Prapas, Ioannis, et al. "Televit: Teleconnection-driven transformers improve subseasonal to seasonalwildfire forecasting." International Conference on Computer Vision. 2023.) proposed a teleconnection-driven visual transformer method (TeleViT), which can view the Earth as an interconnected whole and integrate fine-grained local inputs with global-scale inputs, such as climate indices and coarse-grained global variables. Through comprehensive experiments, this method demonstrated the superiority of TeleViT in accurately predicting global regional patterns of burns, with a prediction window of up to four months. This advantage is particularly significant in longer prediction windows, indicating that deep learning models that use teleconnections to capture the dynamics of the Earth system have stronger capabilities. Dimitrios Michail et al. (Michail, Dimitrios, et al. "Seasonal Fire Prediction using Spatio-Temporal Deep Neural Networks." arXiv preprint arXiv:2404.06437(2024).) trained deep learning models with different architectures to capture the spatiotemporal context that contributes to wildfires. The approach focused on evaluating these models on burned areas at different global forecast horizons (up to six months), and how varying spatial or temporal contexts affect model performance. The results showed that deep learning models with longer input time series provided more robust forecasts across different forecast horizons, while incorporating spatial information to capture the spatiotemporal dynamics of wildfires significantly improved model performance. However, these methods assume the use of shared global meteorological data for training and prediction. In real-world scenarios, this assumption presents the following challenges: 1. Data privacy and security: Meteorological data from many countries and regions cannot be directly shared due to national security, commercial confidentiality, or other policy concerns; 2. Data quality and coverage: The density of meteorological monitoring networks worldwide is uneven, resulting in significant differences in data quality and timeliness across regions. At the same time, these current methods assume that time series information is fed into the model sequentially, without explicitly considering the complex dependencies between previous and subsequent time steps. Therefore, how to use unbalanced, multi-source, non-IID meteorological data for global wildfire forecasting is an important and unresolved problem.

[0006] The federated learning framework was first proposed by Google in 2016 (McMahan, Brendan, et al. "Communication-efficient learning of deep networks from decentralized data." Artificial intelligence and statistics. PMLR, 2017). This approach proposes a deep federated learning method based on iterative model parameter averaging and undergoes extensive empirical evaluation across five different model architectures and four datasets. Experimental results demonstrate that this method is robust to unbalanced and non-IID data distributions, a hallmark of federated learning scenarios. Research demonstrates that this method requires 10 to 100 times fewer communication rounds than synchronous stochastic gradient descent. However, in this method, since the data distribution between clients usually varies significantly, this will lead to unstable global model updates and even convergence to a poor local solution. Therefore, Tian Li et al. (Li, Tian, ​​et al. "Federated optimization in heterogeneous networks." Proceedings of Machine learning and systems 2 (2020): 429-450.) proposed FedProx, which can be regarded as a generalization and reparameterization of the current mainstream federated learning method FedAvg. Although this reparameterization only makes minor modifications to the method itself, these modifications are of great significance both in theory and practice. The present invention is based on the improvement of such a baseline method framework. Summary of the Invention

[0007] To address these issues, the present invention proposes a global wildfire prediction method and device based on federated learning of meteorological data. Specifically, based on the existing FedAvg framework, the present invention addresses the privacy and security concerns of global wildfire meteorological data, as well as the imbalance of meteorological data, by proposing a federated learning method for wildfire data. Furthermore, to address the issue of missing long-range information during the training of time series features in deep learning models, the present invention proposes a time series cross-network architecture to construct long-term feature connections.

[0008] The technical solution adopted in the present invention is as follows:

[0009] A global wildfire prediction method based on federated learning of meteorological data includes the following steps:

[0010] Obtain the meteorological privacy data to be predicted, input the meteorological privacy data into a trained meteorological data federated learning network model, and obtain the global wildfire prediction probability result; the meteorological data federated learning network model is composed of several local clients and a global server, and the several local clients and the global server adopt a time-series cross network with the same architecture.

[0011] Furthermore, the training process of the meteorological data federated learning network model includes:

[0012] 1) Obtain a global wildfire image dataset, preprocess the dataset, and divide it into a training set, a validation set, and a test set;

[0013] 2) Given an initial network weight, each local client is trained multiple times using local-specific data from a global wildfire image dataset to obtain a local network weight;

[0014] 3) Sending the local network weight to the global server, the global server performs weighted averaging on the local network weight to obtain a global weighted average network weight, uses the global weighted average network weight as the initial network weight of the global server, and trains the global server using global shared data to obtain a global network weight;

[0015] 4) Without sharing local data, the global network weight is sent to the local client, and the global network weight is used as the new initial network weight of the local client;

[0016] 5) Repeat steps 2)-4) to obtain the trained meteorological data federated learning network model.

[0017] Furthermore, the global wildfire image dataset is a time series dataset that includes multiple seasonal fire driving factors, including locally specific data and globally shared data. The locally specific data is the time series data of several fire driving variables, and the globally shared data is the time series data of ocean and climate indices.

[0018] Furthermore, the time series data of each fire driving variable is trained through a different local client.

[0019] Furthermore, the local client is trained separately using the local specific data to obtain the local network weight, specifically:

[0020] The local client is fed with local data and forward propagation is performed to obtain an output prediction value. A loss value is calculated based on the output prediction value and the real labeled data in the dataset. The loss value is used to perform a backpropagation operation to update the model parameters. The local client is trained iteratively multiple times to obtain the local client network weight. The loss function of the local client is:

[0021]

[0022] Among them, L local_fire is the loss function of the mth local client, represents the weight of the mth local client network, Z represents the number of local client network parameters, t is the current time, To provide real-world wildfire data around the world. is the global wildfire prediction result of the mth local client at time t.

[0023] Furthermore, the local network weight and the global network weight are both transmitted in an encrypted manner.

[0024] Furthermore, the global weighted average network weight calculation formula is:

[0025]

[0026] in, is the global weighted average network weight after weighted averaging of each local network weight, W m is the local network weight of the mth local client, M is the number of local clients, n m is the number of samples of the mth local client.

[0027] Furthermore, the global server is trained using the global shared data, specifically in the following steps:

[0028] The global shared data is input into the global server, and forward propagation calculation is performed to obtain a global output prediction value. A global loss value is calculated based on the global output prediction value and the true value in the global shared data. The global loss value is used to perform a backpropagation operation to update the model parameters and obtain the global network weight. The loss function of the global server is:

[0029]

[0030] Among them, L global_fire represents the loss function of the global server network, {w 1 ,w 2 ,…,w Z} represents the weight of the global server network, Z represents the number of global server network parameters that are the same as the local client network, t is the current time t, y t To provide real-world wildfire data around the world. is the global wildfire prediction result of the global server at time t.

[0031] Furthermore, the propagation process of the time-series cross network is:

[0032]

[0033]

[0034] Among them, X T Represents a four-dimensional feature vector with a feature size of T×C×W×H, where T represents the temporal size of the feature, and C, W, and H represent the number of channels, width, and length of the feature, respectively; Represents the feature point position of the cth channel, hth row, and wth column of the three-dimensional feature at time step t. r represents the number of feature time series crossovers, s is the time series size increased by the current convolution effect, represents the new features after 3D convolution processing, and f(·) represents the activation function.

[0035] A global wildfire prediction device based on federated learning of meteorological data, including:

[0036] one or more processors;

[0037] a memory for storing one or more programs;

[0038] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned global wildfire prediction method based on federated learning of meteorological data.

[0039] The beneficial effects of the present invention are:

[0040] The present invention introduces a federated learning framework for the first time in the wildfire prediction task, breaking through the bottleneck that traditional meteorological data cannot be fully shared. Through federated learning, it is possible to integrate global meteorological information while protecting the data privacy of each participating node (such as meteorological agencies in different countries or regions), thereby improving the generalization ability of the prediction model. The globally unified Ocean and Climate Index (OCIs) is used as public data for retraining of the global server. These public data can significantly enhance the synergy between models in different regions and capture the key global driving factors of wildfires. At the same time, the present invention proposes a time series cross-network framework for modeling feature associations over long time spans. This framework can effectively capture the deep connections between different time steps, avoiding the traditional method of relying solely on time sequence and ignoring the correlation between previous and subsequent meteorological data. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of a global wildfire prediction method based on federated learning of meteorological data according to an embodiment of the present invention.

[0042] Figure 2 This is a structural diagram of the meteorological data federated learning network according to an embodiment of the present invention.

[0043] Figure 3 This is a timing cross-network structure diagram of an embodiment of the present invention.

[0044] Figure 4 This is a graph showing the results of global wildfire prediction using the Seasfire Cube dataset according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The method of the present invention is described in detail below with reference to the accompanying drawings and examples. This example is implemented based on the technical solution of the present invention, and provides an implementation method and specific operating process. However, the protection scope of the present invention is not limited to the following examples.

[0046] A global wildfire prediction method based on federated learning of meteorological data includes the following steps:

[0047] Obtain the meteorological privacy data to be predicted, input the meteorological privacy data into a trained meteorological data federated learning network model, and obtain the global wildfire prediction probability result; the meteorological data federated learning network model is composed of several local clients and a global server, and the several local clients and the global server adopt a time-series cross network with the same architecture.

[0048] Furthermore, the training process of the meteorological data federated learning network model includes:

[0049] 1) Obtain a global wildfire image dataset, preprocess the dataset, and divide it into a training set, a validation set, and a test set;

[0050] 2) Given an initial network weight, each local client is trained multiple times using local-specific data from a global wildfire image dataset to obtain a local network weight;

[0051] 3) Without sharing local data, the local network weight is sent to the global server. The global server performs weighted averaging on the local network weight to obtain a global weighted average network weight. The global weighted average network weight is used as the initial network weight of the global server. The global server is trained using the globally shared data to obtain a global network weight.

[0052] 4) Sending the global network weight to the local client and using the global network weight as the new initial network weight of the local client;

[0053] 5) Repeat steps 2)-4) to obtain the trained meteorological data federated learning network model.

[0054] Furthermore, the global wildfire image dataset is a time series dataset that includes multiple seasonal fire driving factors, including locally specific data and globally shared data. The locally specific data is the time series data of several fire driving variables, and the globally shared data is the time series data of ocean and climate indices.

[0055] Furthermore, the time series data of each fire driving variable is trained through a different local client.

[0056] Furthermore, the local client is trained separately using the local specific data to obtain the local network weight, specifically:

[0057] The local client is fed with local data and forward propagation is performed to obtain an output prediction value. A loss value is calculated based on the output prediction value and the real labeled data in the dataset. The loss value is used to perform a backpropagation operation to update the model parameters. The local client is trained iteratively multiple times to obtain the local client network weight. The loss function of the local client is:

[0058]

[0059] Among them, L local_fire is the loss function of the mth local client, represents the weight of the mth local client network, Z represents the number of local client network parameters, t is the current time, To provide real-world wildfire data around the world. is the global wildfire prediction result of the mth local client at time t.

[0060] Furthermore, the local network weight and the global network weight are both transmitted in an encrypted manner.

[0061] Furthermore, the global weighted average network weight calculation formula is:

[0062]

[0063] in, is the global weighted average network weight after weighted averaging of each local network weight, W m is the local network weight of the mth local client, M is the number of local clients, n m is the number of samples of the mth local client.

[0064] Furthermore, the global server is trained using the global shared data, specifically in the following steps:

[0065] The global shared data is input into the global server, and forward propagation calculation is performed to obtain a global output prediction value. A global loss value is calculated based on the global output prediction value and the true value in the global shared data. The global loss value is used to perform a backpropagation operation to update the model parameters and obtain the global network weight. The loss function of the global server is:

[0066]

[0067] Among them, L global_fire represents the loss function of the global server network, {w 1 ,w 2 ,…,w Z} represents the weight of the global server network, Z represents the number of global server network parameters that are the same as the local client network, t is the current time t, y t To provide real-world wildfire data around the world. is the global wildfire prediction result of the global server at time t.

[0068] The local client can use the global network parameters trained by the global server to train the local client, and obtain a network model that adapts to the global data without sharing data.

[0069] Furthermore, the propagation process of the time-series cross network is:

[0070]

[0071]

[0072] Among them, X T Represents a four-dimensional feature vector with a feature size of T×C×W×H, where T represents the temporal size of the feature, and C, W, and H represent the number of channels, width, and length of the feature, respectively; Represents the feature point position of the cth channel, hth row, and wth column of the three-dimensional feature at time step t. r represents the number of feature time series crossovers, represents the new features after 3D convolution processing, and f(·) represents the activation function.

[0073] Example

[0074] See also Figure 1 , the implementation of the embodiment of the present invention includes the following steps:

[0075] A. Prepare global wildfire image training, validation, and test datasets, and perform data preprocessing and data partitioning;

[0076] The dataset used in this embodiment is the public dataset Seasfire Cube, which can be used for global seasonal fire prediction. The dataset contains 21 years of data (2001-2021), with a temporal resolution of once every 8 days and a spatial resolution of 0.25 degrees longitude and latitude grid. The dataset covers a variety of seasonal fire drivers, extending from atmospheric and climate variables to vegetation variables, socioeconomic factors, and fire-related target variables such as burned area, fire radiant power, and fire-related carbon dioxide emissions. In this embodiment, the dataset is divided into three parts: training, validation, and test sets. The data from 2002 to 2017 are used as the training set, the data from 2018 are used as the validation set, and the data from 2019 are used as the test set.

[0077] To process local client input data, this example extracts 80×80 data blocks, dividing global meteorological data represented at 0.25-degree resolution (1440×720 cells) into 18×9 = 162 local input patches. Ten fire-driving variables extracted from the Seasfire Cube (mean sea level pressure, total precipitation, vapor pressure, sea surface temperature, mean air temperature at 2 meters, surface solar radiation, soil water content, daytime surface temperature, normalized difference vegetation index, and population density) are distributed to 10 different local clients for training.

[0078] For the processing of the global shared data in the global server, namely the ocean and climate index, this embodiment extracts 10 ocean and climate indices (West Pacific Index, Pacific North American Index, North Atlantic Oscillation Index, Southern Oscillation Index, Global Average Temperature, Pacific Decadal Oscillation Index, East Asia / West Russia Index, East Pacific / North Pacific Oscillation Index, 3.4 Regional anomalies, bivariate ENSO time series). Since these indices represent global indicators, there is only one value for each period of time, and its value is , which finally forms a global time series data of size (10, 10).

[0079] B. Design a meteorological data federated learning network model for global wildfire prediction. The model consists of multiple local clients and a global server.

[0080] like Figure 2 As shown in the figure, the entire meteorological data federated learning network model training process is divided into local client training and global server training. Each of the 10 local clients has 10 different types of meteorological data with different distributions. The local network and the global network have the same structure. After the local network weights are trained for each client, they are uploaded to the global network on the global server. The global network is then trained again using global time series data based on the weighted average of these local network weights. The global network weights from the global server are then distributed to each local client.

[0081] C. Place the local data in the local client for training, output wildfire predictions, and compare them with the labeled data in the dataset. Use the backpropagation algorithm for end-to-end training to obtain a local global wildfire prediction network model.

[0082] Assume that there are M local clients (in this embodiment, M=10), and each local client has its own unique local weather wildfire prediction data X m ∈R T×W×H , where m represents the mth client, T, W, and H represent the total time span, length, and width of the selected data, respectively. When the data of each local client is input into the local network designed by the present invention, it is used to optimize the non-convex neural network objective function:

[0083]

[0084] For one of the samples Through the timing cross network Ne in the local network T After the forward propagation of , we get the global wildfire prediction results

[0085] D. After multiple rounds of training in step C, the trained model weights are sent to the global server without leaking any data.

[0086] There are M participating local clients in the system (such as weather stations or monitoring stations in different regions), each client m Holds its local dataset D m , the amount of data is n m =|D m After completing local training in step C, each local client uploads the following information to the global server:

[0087] Model weights The mth client m Using local dataset D m Train the model and obtain better model parameters W after multiple rounds of back propagation m This weight contains all the parameters of the network model, such as the weights and bias values ​​of the neural network, but it does not contain specific data, ensuring that the local client's meteorological privacy data is not leaked. The transmission of weights can be encrypted to ensure data security.

[0088] Local data volume n m :Local client Client m The number of data samples n m The data is also sent to the global server for weighted average calculation. In FedAvg, the data size is an important weight aggregation factor, which can reflect the contribution of each local client to the global model.

[0089] E. The global server is responsible for integrating the local network weights obtained by multiple local clients to obtain the global weighted average network weight, and uses the global ocean and climate index to train the global network weight to obtain the global wildfire prediction network model;

[0090] After the global server receives the training parameters of all local clients, it performs a weighted average of the model weights based on the data share of each client. The formula is as follows:

[0091]

[0092] in, Represents the global weighted average network weight after weighted average of each client. After obtaining the corresponding global weighted average network weight, the global server also needs to perform training similar to the local client, but the input of the global server becomes a global shared parameter, namely the ocean and climate index o tThe retraining of the global server can enable the global server to combine unique global ocean and climate index data (such as sea surface temperature, El Nino-Southern Oscillation index, global climate change trends, etc.) to build richer data input features for Perform additional optimization. The global training goal is to further improve the accuracy and robustness of the model in global wildfire prediction. The training loss function can be expressed as:

[0093]

[0094] For one of the samples o t , through the temporal cross network Ne in the global network T After the forward propagation of , we get the global wildfire prediction results The recovered optimal global network weight is expressed as W global ={w 1 ,w 2 ,...,W Z}.

[0095] F. The trained global network weights obtained in step E are re-distributed to the local client, the output wildfire predictions are compared with the labeled data in the dataset, and the local global wildfire prediction network model is obtained by end-to-end training using the backpropagation algorithm.

[0096] The redistribution of weights aims to efficiently and securely distribute the global network weights after integration and training optimization of the global server to all participating local clients, providing a unified starting point for the next round of local training. global Packaging and encryption ensure that it cannot be intercepted or tampered with during distribution. Encryption methods can use symmetric encryption or public key encryption (such as AES or RSA), combined with transport layer security protocols (such as TLS) to ensure transmission security. The encrypted global network weight is transmitted to each client through the network.

[0097] After receiving the encrypted global network weights, each local client uses the decryption key distributed by the server or its own key to decrypt the global network weights to ensure that the local client can correctly load the latest global model. After decrypting and verifying the global network weights, each local client uses them as the new initial model weights to replace the local network weights after the previous round of training. The replacement operation is expressed as:

[0098]

[0099] Each local client will use the new local network weight W m Combined with its local meteorological time series data X mA new round of training begins. The updated local network weights can better adapt to the shared information between local clients and provide higher-quality initial parameters for the local model.

[0100] G. Such as Figure 3 As shown in Figure 2, in steps C and E, a temporal cross network with the same architecture is designed to learn the relevant features between long time series.

[0101] The time series cross network framework uses a shared structure and parameters to extract features and model interactions from input multidimensional time series data. The core of the network is to utilize a cross mechanism to handle the global correlations between time series data, thereby capturing long-range dependencies and deep interactions between specific time points in the temporal dimension. At the same time, the network preserves local patterns in time series and, by fusing short-term and long-term features, provides comprehensive support for complex time series modeling. This design not only improves model training efficiency but also reduces the number of parameters through a unified architecture, improving generalization performance and making it more suitable for long-span forecasting and analysis tasks.

[0102] Suppose we have a four-dimensional feature vector X with a feature size of T×C×W×H T , where T represents the time length of the data, C, W, and H represent the number of channels, width, and length of the feature respectively. It represents the position of the feature point of the cth channel, hth row, and wth column of the three-dimensional feature at time step t. The propagation process of the temporal cross network is as follows:

[0103]

[0104] in, r represents the number of feature time series crossovers, s is the time series size increased by the current convolution effect, represents the new features after 3D convolution processing, and f(·) represents the activation function.

[0105] H. Repeat steps C and E until training is complete, and input the private meteorological data to be tested into the meteorological data federated learning network model to obtain the global wildfire prediction probability results.

[0106] The resolution of the input image is divided into 80×80 data blocks and fed into the trained global wildfire prediction network. Finally, the output prediction results are subjected to a Sigmoid operation and concatenated in longitude and latitude order to obtain the final global wildfire prediction probability result, as shown in the figure below. Figure 4As shown, the first row represents the prediction results of this embodiment on January 1, 2019, and the second row represents the corresponding target label results. The model accurately captures the areas with high wildfire probability in central Africa and parts of South America. The color depth is basically consistent with the actual occurrence probability, indicating that the model has high sensitivity to these areas. The present invention achieves high-precision global wildfire prediction while protecting the privacy of local client data. It can accurately predict the probability of wildfire occurrence in different regions of the world at different time points and capture the geographical distribution characteristics of wildfire occurrence.

[0107] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0108] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0109] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0110] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1A step that specifies a function in one or more boxes.

[0111] The above specific embodiments are used to illustrate the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A global wildfire prediction method based on meteorological data federated learning, characterized in that: The following steps are involved: Obtain the meteorological privacy data to be predicted, input the meteorological privacy data into a trained meteorological data federated learning network model, and obtain the global wildfire prediction probability result; the meteorological data federated learning network model is composed of a number of local clients and a global server, and the number of local clients and the global server adopt a timing cross network with the same architecture.

2. A global wildfire prediction method based on meteorological data federated learning according to claim 1, characterized in that: The training process of the meteorological data federated learning network model includes: 1) Obtain a global wildfire image dataset, preprocess the dataset, and divide the dataset into a training set, a validation set, and a test set; 2) Given an initial network weight, each local client is trained multiple times iteratively using local-specific data in the global wildfire image dataset to obtain a local network weight; 3) Sending the local network weight to the global server, the global server performs weighted averaging on the local network weight to obtain a global weighted average network weight, using the global weighted average network weight as the initial network weight of the global server, and training the global server using global shared data to obtain a global network weight; 4) Sending the global network weight to the local client and using the global network weight as the new initial network weight of the local client; 5) Repeat steps 2)-4) to obtain the trained meteorological data federated learning network model.

3. A global wildfire prediction method based on meteorological data federated learning according to claim 2, characterized in that: The global wildfire image dataset is a time series dataset that includes multiple seasonal fire driving factors, including locally specific data and globally shared data. The locally specific data is the time series data of several fire driving variables, and the globally shared data is the time series data of ocean and climate indices.

4. A global wildfire prediction method based on meteorological data federated learning according to claim 3, characterized in that: The time series data of each fire driving variable is trained through a different local client.

5. A global wildfire prediction method based on meteorological data federated learning according to claim 2, characterized in that: The method of using the local specific data to train each local client to obtain the local network weight is specifically as follows: The local specific data is input into the local client, and the forward propagation calculation is performed to obtain the output prediction value. The loss value is calculated according to the output prediction value and the real labeled data in the data set. The loss value is used to perform the back propagation operation, update the model parameters, and perform multiple iterative training on the local client to obtain the local client network weight; the loss function of the local client is: Among them, L local_fire is the loss function of the mth local client, represents the weight of the mth local client network, Z represents the number of local client network parameters, t is the current time t, To label the real data of global wildfires, is the global wildfire prediction result of the mth local client at time t.

6. A global wildfire prediction method based on meteorological data federated learning according to claim 2, characterized in that: The local network weight and the global network weight are both transmitted in an encrypted manner.

7. A global wildfire prediction method based on meteorological data federated learning according to claim 2, characterized in that: The global weighted average network weight calculation formula is: in, is the global weighted average network weight after weighted average of each local network weight, W m is the local network weight of the mth local client, M is the number of local clients, n m is the number of samples of the mth local client.

8. A global wildfire prediction method based on meteorological data federated learning according to claim 2, characterized in that: The global server is trained by using the global shared data, and the specific steps are: The global shared data is input into the global server, and forward propagation calculation is performed to obtain the global output prediction value. The global loss value is calculated according to the global output prediction value and the true value in the global shared data. The global loss value is used to perform back propagation operation, update the model parameters, and obtain the global network weight. The loss function of the global server is: Among them, L global_fire represents the loss function of the global server network, {w 1 ,w 2 ,…,w Z } represents the weight of the global server network, Z represents the number of global server network parameters that are the same as the local client network, t is the current time t, y t To label the real data of global wildfires, It is the global wildfire prediction result of the global server at time t.

9. A global wildfire prediction method based on meteorological data federated learning according to claim 1, characterized in that: The propagation process of the timing cross network is: Among them, X T Represents a four-dimensional feature vector with a feature size of T×C×W×H, where T represents the temporal size of the feature, and C, W, and H represent the number of channels, width, and length of the feature, respectively; represents the feature point position of the cth channel, hth row, and wth column of the three-dimensional feature at time step t. r represents the number of feature time series crossovers, s is the time series size increased by the current convolution effect, represents the new features after 3D convolution, and f(·) represents the activation function.

10. A global wildfire prediction device based on meteorological data federated learning, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the global wildfire prediction method based on federated learning of meteorological data as described in any one of claims 1-9.

Images