Multi-source forest fire danger dynamic early warning method based on meteorological large model and light weight AI collaboration

Abstract

The application discloses a multi-source forest fire risk dynamic early warning method based on meteorological large model and light AI cooperation, which comprises the following steps: firstly, collecting multi-source data and preprocessing the data; secondly, training a multi-level forest fire probability prediction model based on the preprocessed multi-source data, and realizing cross-level cooperative calculation by introducing a probability pixel; and finally, integrating and optimizing the output of the artificial intelligence meteorological large model and the prediction result of the multi-level forest fire probability prediction model to generate the final forest fire occurrence probability. The application can realize high-precision prediction, and significantly improve the prediction accuracy of the forest fire occurrence probability through the artificial intelligence meteorological large model and multi-source data fusion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forest fire danger forecasting, in particular to a multi-source forest fire danger dynamic early warning method based on meteorological large models and light AI cooperation. BACKGROUND

[0002] Forest fires are a highly destructive natural disaster. In recent years, with the intensification of climate change and human activities, the frequency and scale of forest fires are on the rise. Therefore, developing efficient and accurate forest fire danger forecasting methods has become an important problem that needs to be solved. Currently, forest fire danger weather forecasting mainly relies on meteorological numerical prediction models. These models can provide relatively accurate meteorological data (such as temperature, humidity, wind speed, precipitation, etc.) by simulating atmospheric dynamics and thermodynamics processes, and predict fire danger weather grades based on these data. However, the existing methods have the following significant problems:

[0003] (1) Large consumption of computing resources and slow speed. Meteorological numerical prediction models usually require high-performance computing resources (such as supercomputers) to run, and the calculation process is complex and time-consuming. For example, global numerical weather prediction models such as the European Centre for Medium-Range Weather Forecasts (ECMWF) and the Weather Research and Forecasting Model (WRF) take several hours or even longer to complete a prediction, making it difficult to meet the demand for real-time forecasting. In addition, high-resolution meteorological field prediction further increases the computational burden, limiting the popularization and promotion of the model in practical applications.

[0004] (2) Single data, insufficient prediction accuracy. Current fire danger weather forecasting models mainly rely on meteorological data and do not fully consider other key factors affecting the occurrence of forest fires, such as vegetation coverage, geographical location, fire point information, etc. For example, vegetation type and coverage directly affect the distribution and burning characteristics of combustible materials, while topographic features (such as elevation, slope, aspect) affect the spread speed and direction of fire. Due to the lack of fusion analysis of these multi-source data, the prediction results of existing models often cannot comprehensively and accurately reflect the actual situation of forest fire occurrence, resulting in insufficient prediction accuracy.

[0005] (3) Limited practicality. Since existing models only consider meteorological factors, their prediction results have large deviations in practical applications. For example, in densely vegetated areas or complex terrain conditions, fire danger forecasting based only on meteorological data may severely underestimate or overestimate fire risk. In addition, existing models usually cannot provide high-resolution spatial and temporal predictions, and cannot support precise fire warning and prevention and control decisions.

[0006] In summary, forest fire prediction technology is facing severe technical bottlenecks and innovation opportunities. Although the traditional numerical weather prediction model can provide basic meteorological parameters through atmospheric dynamics simulation, its inherent supercomputer dependence results in a long time consumption of several hours for a single global prediction. When the resolution is improved to the business demand, the computing load increases exponentially. More importantly, the existing system only integrates basic meteorological indicators, completely ignoring the three key disaster-causing elements of vegetation dynamic characteristics, terrain influencing factors and historical fire point patterns, resulting in large prediction errors in dense vegetation areas or complex terrain conditions. This serious lack of data dimension and fundamental limitation of computing efficiency makes it difficult for traditional methods to meet the timeliness requirements of modern disaster prevention and mitigation, and also difficult to guarantee the accuracy of the prediction results. SUMMARY

[0007] Current fire weather prediction mainly relies on meteorological numerical prediction models. These models can provide relatively accurate meteorological data, but have problems of low computing efficiency, single data and insufficient practicality. Meteorological numerical prediction models usually require high-performance computing resources, and the calculation speed is slow, which is difficult to meet the demand of real-time prediction. The existing models only consider meteorological factors, without incorporating multi-source data such as vegetation coverage, geographic location, and fire point information, resulting in prediction results that cannot fully and accurately reflect the actual situation of forest fire occurrence. Due to the lack of fusion analysis of multi-source data, the prediction results of existing models often have large deviations in actual application, making it difficult to support precise fire warning and prevention and control decisions.

[0008] The application constructs an intelligent collaborative prediction system of "meteorological large model + fire small model", which is a forest fire danger prediction method based on artificial intelligence meteorological large model and multi-source data (meteorology, vegetation, geography, fire point information) fusion, for improving the accuracy, efficiency and practicality of forest fire occurrence probability prediction. The system first realizes the deep integration of multi-source heterogeneous data, and through grid processing, the meteorological field, vegetation index (such as normalized vegetation index, enhanced normalized vegetation index), terrain characteristics (elevation, slope, aspect) and historical fire points are time and space aligned and dynamically normalized. In terms of model architecture, a three-level linkage lightweight hybrid model group is innovatively designed: the feature extraction layer uses a 486-dimensional classification head Transformer network to capture spatial correlation; the probability correction layer parallelly deploys LSTM, LightGBM and MLP three types of differentiated models; the decision fusion layer generates the final probability output through an adaptive weighting mechanism. The specially developed probability pixel technology realizes cross-level knowledge transfer, enabling the model to maintain lightweight while improving fire detection rate. The system is deeply coupled with the Pangu meteorological large model, and only needs to call the future 24-hour meteorological prediction data to generate a global fire danger distribution map with a precision of kilometers, the overall response time is compressed to minutes, solving the contradiction between calculation efficiency and prediction accuracy of traditional methods. This innovation not only realizes the paradigm shift from single meteorological analysis to multi-source collaborative prediction, but also makes high-precision fire danger warning first available for large-scale business application through edge-friendly lightweight design.

[0009] The application proposes a multi-source forest fire danger dynamic warning method based on meteorological large model and lightweight AI model collaborative calculation, specifically including the following steps:

[0010] Step 1: Collect meteorological data, vegetation information (such as normalized vegetation index, enhanced normalized vegetation index), geographic information (such as elevation, slope, aspect), fire point information (such as historical fire records) and other multi-source data, and clean, normalize and extract features from the data:

[0011] Step 2: Based on the preprocessed multi-source data, train a multi-level forest fire probability prediction model, and realize cross-level collaborative calculation by introducing probability pixels, the model includes sequentially connected feature extraction layer, probability correction layer and decision fusion layer;

[0012] Step 3: Integrate and optimize the output of the artificial intelligence meteorological large model and the prediction results of the multi-level forest fire probability prediction model to generate the final forest fire occurrence probability, the artificial intelligence meteorological large model is the open source Pangu meteorological large model, and the meteorological large model can provide high spatiotemporal resolution meteorological prediction data to expand the time coverage of meteorological parameters required for fire analysis.

[0013] As preferred, the multi-source data comprises daily mean GTOPO30, normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI), fire point brightness temperature data, European Centre for Medium-Range Weather Forecasts (ECMWF Reanalysis 5, ERA5) reanalysis precipitation, land surface temperature, wind speed in longitude and latitude directions, mean sea level pressure, temperature, humidity, potential, wind speed in longitude and latitude directions of pressure layer at vertical height, as described in step 1.

[0014] All multi-source data are resampled into a unified resolution according to the WGS-84 projection coordinate system, and all data except the fire point brightness temperature data are normalized, and the maximum and minimum values of different types of data during the research time are recorded for unified normalization processing of historical data and predicted data, and all input data follow a unified resolution for spatial alignment.

[0015] The present application adopts a pixel training strategy to expand all multi-source data (two-dimensional image data) into one-dimensional data for training, and adopts a spatial image discretization strategy to obtain multi-dimensional pixel data. Based on the terrain elevation threshold in the daily mean GTOPO30, the effective land pixel is determined, and the above-mentioned all key features related to forest fire are converted into standardized feature vectors of environmental parameters and one-dimensional fire condition labels at a single location.

[0016] As preferred, the present application constructs a multi-level forest fire probability prediction model, and realizes cross-level collaborative calculation by introducing a probability pixel, as described in step 2. The model comprises a feature extraction layer, a probability correction layer and a decision fusion layer connected in sequence. First, the standardized feature vector and the probability pixel are spliced along the feature dimension, and the sine position coding is added to the spliced tensor to input the feature extraction layer. The layer adopts a Transformer network architecture, extracts spatial correlation features through its multi-head attention mechanism, and outputs an initial fire probability distribution. Then, the probability correction layer is deployed in parallel with three submodules of long short-term memory network (LSTM), light gradient boosting machine (LightGBM) and multilayer perceptron (MLP), each module independently receives the output feature vector of the feature extraction layer and generates three correction probabilities respectively. Finally, the decision fusion layer implements a weighted voting mechanism, integrates the three correction probabilities of the probability correction layer through a preset weight coefficient to generate the final fire probability.

[0017] The probability pixel is taken as a prefabricated input component, and the second dimension of the original probability pixel input for the first time is an integer multiple of the dimension of the normalized feature vector. Before inputting into the network, the probability pixel is segmented along the second dimension into a tensor with the same dimension as the normalized feature vector, and is spliced with the normalized feature vector in the first dimension. The spliced composite tensor is added to the sine position coding and input into the Transformer network, and after being processed by the multi-head attention layer, it is split into two groups of data along the second dimension and reconstructed into the form of the original probability pixel and the normalized feature vector. The two groups of reconstructed data are input into the parameter-independent feedforward neural network layer inside the Transformer.

[0018] The probability pixel as a dynamic storage unit can not only access global fire hazard prior knowledge to generate probability constraints through attention weights in the feature extraction layer, but also can be coupled with the feature vector in the probability correction layer, while implementing probability correction for the conservative discrimination tendency of the Transformer fire occurrence. The five probability outputs generated in the three levels of the design balance efficiency and accuracy, the high-level output has better prediction accuracy, and the low-level output can significantly reduce the demand for computing resources, meeting the deployment conditions of edge devices.

[0019] As preferred, according to step 3, the present application adopts a static digital elevation model with constant time resolution and periodically updated vegetation index features, only needs to call an artificial intelligence meteorological large model to generate future meteorological data, repeatedly performs data preprocessing and discretization work to obtain the data to be processed, and then uses a multi-level forest fire probability prediction model to predict the probability of fire occurrence, i.e. global fire occurrence probability distribution. The artificial intelligence meteorological large model is the open-source Pangu meteorological large model.

[0020] Advantages of the present application

[0021] (1) High-precision prediction can be achieved, and the prediction accuracy of forest fire occurrence probability can be significantly improved through artificial intelligence meteorological large model and multi-source data fusion.

[0022] (2) After multi-source data fusion, the prediction result is closer to the actual situation by comprehensively considering meteorological, vegetation, geographical, and fire point information.

[0023] (3) Low resource consumption can be achieved, and the multi-level forest fire probability prediction model structure is lightweight, suitable for deployment in resource-constrained environments, and reduces the computing cost.

[0024] (4) Real-time and scalability, efficient model training and prediction process, can meet the demand of real-time prediction, the method can be applied to different regions and scenes, has strong universality and adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A technical flowchart of the present application is shown;

[0026] Figure 2 A multi-source forest fire risk dynamic early warning method based on meteorological large model and light AI model collaborative calculation is shown;

[0027] Figure 3 The performance of LightGBM, MLP, LSTM and Transformer classifiers in fire detection under historical meteorological data (day 75 of 2002) is shown;

[0028] Figure 4 The generalization ability of LightGBM, MLP, LSTM and Transformer classifier models (day 89 of 2002) is shown;

[0029] Figure 5 The accuracy of various models of the present application on the fire classification task is shown;

[0030] Figure 6 The evaluation indicators of the model of the present application at each test time are shown. DETAILED DESCRIPTION

[0031] In order to better illustrate the invention and advantages of the present project, the invention is further described below in combination with the drawings and examples;

[0032] Example 1:

[0033] As shown in Figure 1 , the multi-source forest fire risk dynamic early warning method based on meteorological large model and light AI collaboration includes the following steps:

[0034] Step 1: Collect meteorological data, vegetation information (such as normalized vegetation index, enhanced normalized vegetation index), geographic information (such as elevation, slope, aspect), fire point information (such as historical fire records) and other multi-source data, and clean, normalize and feature extract the data:

[0035] Step 2: Based on the pre-processed multi-source data, train a multi-level forest fire probability prediction model, and realize cross-level collaborative calculation by introducing probability pixels, the model includes sequentially connected feature extraction layer, probability correction layer and decision fusion layer;

[0036] Step 3: Integrate and optimize the output of the artificial intelligence meteorological large model and the prediction result of the multi-level forest fire probability prediction model to generate the final forest fire occurrence probability, the artificial intelligence meteorological large model is the open-source Pangu meteorological large model;

[0037] Step 4: Predict the meteorological field of the target area with the artificial intelligence meteorological large model, output the forest fire occurrence probability of the target area according to the predicted meteorological field, and provide visual results and warning information;

[0038] As described in step 1, the multi-source data includes daily Copernicus digital elevation model, normalized vegetation index (NDVI) and enhanced vegetation index (EVI), fire point brightness temperature data, European Center for Medium-Range Weather Forecasts (ERA5) reanalysis precipitation, surface 2m temperature, 10m wind speed in longitude and latitude direction, average sea level pressure, temperature, humidity, potential, wind speed in longitude and latitude direction at 13 pressure layers in vertical height.

[0039] The Moderate Resolution Imaging Spectroradiometer (MODIS) is a large space remote sensing instrument developed by NASA, and the NDVI and EVI are derived from the MOD13A1-V6 data set, and the fire point brightness temperature data is derived from the MODIS MOD14 / MYD14 satellite data; The production data source of the Copernicus digital elevation model is obtained by interferometric processing of global radar satellite data obtained during the TanDEM-X mission from 2010 to 2015; ERA5 is the fifth generation atmospheric reanalysis data set of the European Center for Medium-Range Weather Forecasts on global climate since 1979, covering multiple variables such as atmosphere, land and ocean, and due to its high spatial and temporal resolution and wide time scale, it has become one of the most common data sets in meteorological analysis. The invention uses daily ERA5 surface 2m temperature, 10m wind speed in longitude and latitude direction, average sea level pressure, temperature, humidity, potential, wind speed in longitude and latitude direction at 13 pressure layers in vertical height, including 1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa.

[0040] In order to align the collected data in geographical space and map coordinates, all multi-source data are resampled to 0.25°x0.25° resolution according to the WGS-84 projection coordinate system; in order to eliminate the dimension effect, all data except the fire point brightness temperature data are normalized to 0-1; in order to eliminate the difference of data distribution itself, the maximum and minimum values of different types of data during the study time are recorded for unified normalization processing of historical data and prediction data, Table 1 shows the data name, time resolution, source, maximum value and minimum value involved in the invention, the time series covers the climate cycle from 2002 to 2022, and all input data follow the 0.25°x0.25° resolution for spatial alignment.

[0041] Table 1 Data name, time resolution, source, maximum value and minimum value involved in the present application

[0042]

[0043] The present application adopts a pixel training strategy, i.e. all two-dimensional image data (multi-source data) is unfolded into one-dimensional data for training. The present application adopts a strategy of spatial image discretization to obtain multi-dimensional pixel data. Based on the terrain elevation threshold in the daily Copernicus digital elevation model, the effective land pixel is determined, and all the above-mentioned key features related to forest fires are converted into standardized feature vectors of environmental parameters and one-dimensional fire condition labels at a single location. Considering that the contribution of high-level meteorological information to ground fires is low, the present application only uses the parts with vertical heights of 150 hPa, 100 hPa and 50 hPa in the meteorological data containing height information. This pixel analysis based on a specific location is advantageous, as it reduces the computational load by about four-fifths compared to traditional global analysis by excluding non-land pixels which account for about 71% of the Earth's surface. Secondly, this makes fire detection and risk assessment only rely on meteorological observation data of the target area, greatly reducing the dependence on global real-time detection networks.

[0044] According to step 2, the present application constructs a multi-level forest fire probability prediction model, and realizes cross-level collaborative calculation by introducing probability pixels. The model includes sequentially connected feature extraction layer, probability correction layer and decision fusion layer. In specific implementation, first, the standardized feature vector and the probability pixel are spliced along the feature dimension, and the sine position coding is added to the spliced tensor to input the feature extraction layer. The feature extraction layer uses the Transformer network architecture, extracts spatial correlation features through its multi-head attention mechanism, and outputs the initial fire probability distribution. Subsequently, the probability correction layer is deployed in parallel with three submodules of long short-term memory network (LSTM), light gradient boosting machine (LightGBM) and multilayer perceptron (MLP). The LightGBM model is suitable for fire detection due to its high applicability in binary classification tasks. MLP uses multiple linear layers and activation functions to realize nonlinear mapping, and is the simplest and most effective neural network. LSTM is a recurrent neural network widely used to solve long sequence dependency problems, which can effectively express and transmit information in long sequences. Each module independently receives the output feature vector of the feature extraction layer and generates three correction probabilities respectively. Finally, the decision fusion layer implements a weighted voting mechanism to integrate the three correction probabilities of the probability correction layer to generate the final fire probability through a preset weight coefficient.

[0045] The details of the deep learning network involved in the multi-level forest fire probability prediction model include: the present application uses a 4-layer hidden layer of 32, a self-attention head number of 2, a classification head hidden layer of 486, and a Transformer network with the same number of input channels as the input channel; the present application uses a 2-layer LSTM model, each layer has a hidden layer of 32, a classification head hidden layer of 486, an output dimension of 2, and respectively represents the probability of the negative class and the positive class of the sample. The present application designs a 3-layer MLP network with a hidden layer size of 64, an input dimension of 23, and an output dimension of 2, which respectively represent the probability of the negative class and the positive class of the sample. The Transformer, LSTM and MLP networks use cross-entropy loss as the loss function on the probability pixel, and Adam as the optimizer to train 10 batches.

[0046] The probability pixel is a prefabricated input component. The second dimension of the original probability pixel input for the first time is an integer multiple of the dimension of the standardized feature vector. Before inputting into the network, the probability pixel is segmented along the second dimension into a tensor with the same dimension as the standardized feature vector, and is spliced with the standardized feature vector in the first dimension. The spliced composite tensor is added to the sine position encoding and input into the Transformer network. After being processed by the multi-head attention layer, it is split into two groups of data along the second dimension, and is reconstructed into the form of the original probability pixel and the standardized feature vector. The two groups of data after reconstruction are input into the parameter-independent feedforward neural network layer inside the Transformer.

[0047] The probability pixel as a dynamic storage unit can not only access global fire risk prior knowledge to generate probability constraints through attention weights in the feature extraction layer, but also can be coupled with the feature vector in the probability correction layer, while implementing probability correction for the conservative discrimination tendency of the Transformer. The design outputs a modified probability that balances efficiency and accuracy at three levels. The high-level output has better prediction accuracy, and the low-level output can significantly reduce the demand for computing resources, meeting the deployment conditions of edge devices.

[0048] As described in step 3, the present application focuses on a lightweight way to realize fire prediction in various regions around the world. The present application combines a meteorological-driven prediction large model and a fire detection model, as shown in Figure 2 A multi-source forest fire risk dynamic early warning method is formed by the collaborative calculation of a meteorological large model and a lightweight model. The three-dimensional meteorological state modeling capability of the high-precision meteorological prediction system is used in the upstream, and the local fire risk assessment or global trend prediction is realized by the lightweight fire detection model in the downstream. It has the generalization ability from historical fire to real-time detection and future fire warning. The input parameters of the two types of models are strictly spatially aligned with the meteorological module, and the long-term stability of the digital elevation terrain and the periodic slow change characteristic of the vegetation index together constitute the theoretical boundary condition of the time-varying parameter update.

[0049] The application adopts a static digital elevation model with constant time resolution and a periodically updated vegetation index feature (the update period is 16 days), only needs to call a meteorological prediction network to generate future 24-hour meteorological data, repeatedly performs data preprocessing and discretization work to obtain the data to be processed, and then uses a multi-level forest fire probability prediction model to predict the probability of fire occurrence, that is, a global fire occurrence probability distribution map. The final warning accuracy is determined by the spatiotemporal fidelity of the meteorological field and the generalization ability of the multi-level forest fire probability prediction model, which can ensure the completeness of the physical process to the greatest extent.

[0050] The calling of the meteorological prediction network to generate future 24-hour meteorological data refers to using an artificial intelligence meteorological large model to perform high-precision prediction on the meteorological field of the target area to generate meteorological data including temperature, humidity, wind speed, precipitation, solar radiation, etc. The artificial intelligence meteorological large model is based on deep learning technology and can capture complex meteorological change rules to provide high-resolution meteorological field prediction results. The meteorological information output by the artificial intelligence meteorological large model and the real-time acquired vegetation information and geographic information are discretized and input into the trained multi-level forest fire probability prediction model to predict the probability of forest fire occurrence. By integrating the artificial intelligence meteorological large model and the forest fire probability prediction model, high-precision and high-efficiency fire probability prediction is realized. The output target area forest fire occurrence probability has a probability value range of 0 to 1. According to a preset threshold, warning information is issued to remind relevant departments to take preventive measures.

[0051] Embodiment 2:

[0052] The credibility of the integrated model proposed in the application is analyzed according to the experimental results.

[0053] 2.1 Verification of the multi-level forest fire probability prediction model on historical data fire and prediction of future fire occurrence probability

[0054] The intuitive embodiment of the model fire detection capability is to obtain the global fire occurrence probability map and analyze the quantitative indicators. The following shows the performance of the four models (LSTM, Transformer, MLP and LightGBM) involved in the present application, although the final output is only determined by the weighting of the upper three models, this part will still analyze the four models. Table 2 shows the accuracy of various models on the fire classification task. In line with the model evaluation method, the data of the 75th day of 2002 in the training set and the 89th day of 2002 in the test set are still selected for verification, and their historical real weather data and weather data predicted by the Panguxi weather big model are used respectively. In the case of extremely uneven sample distribution, the true proportion of the fire area and the non-fire area is selected to be separated and displayed to obtain a clearer accuracy representation capability. Similarly, the accuracy results will be analyzed from the aspects of training accuracy and test accuracy, model fire generalization ability and estimation ability, and AI weather big model prediction driven ability.

[0055] Table 2 Accuracy of various models of the present application on the fire classification task

[0056]

[0057] Firstly, the performance of the model on the training set is higher than that on the test set, which is reasonable and easy to explain. In terms of overall performance, the MLP classification network is slightly worse than the LightGBM network, mainly in the accuracy of the fire area; the accuracy gap between the test set and the training set is small, showing that the model has certain generalization ability. In terms of fire estimation ability, the estimation ability of the MLP network is worse than that of the LightGBM network, even though it has higher accuracy in the non-fire area, but it is not as good as the LightGBM network in the key fire prediction ability, especially on the test data. The huge negative class samples seriously affect the prediction accuracy and generalization ability of the network. Finally, the fire detection system driven by the meteorological data predicted by the AI weather big model is comparable to the result driven by the real meteorological data.

[0058] The intuitive display of the model's fire occurrence probability map in the global range can provide a method for analyzing global fire trends and intuitive inspection, Figure 3 、 Figure 4 The real meteorological data driven global fire probability distribution map of the four classification networks on the training set is respectively shown, and the actual fire point distribution of 5 days before and after the corresponding time is superimposed, represented by green points.

[0059] 2.2 Accuracy quantitative verification of the integrated system on fire occurrence probability

[0060] The performance of the various models used in this study was evaluated, mainly to verify the degree of performance improvement of the four classifier models used in this study on the final results. Tables 2 and 3 show the performance of the four models involved in the feature extraction layer and the probability correction layer of the multi-level forest fire probability prediction model on the test set of the four joint probabilities and the accuracy and importance on the fire classification task. The data of the 75th day of 2002 in the training set and the 89th day of 2002 in the test set were selected for verification, and their historical real meteorological data and meteorological data predicted by the Panguxi meteorological large model were used respectively.

[0061] Table 3 Evaluation index of the model of the present application at each test time

[0062]

[0063] Table 3 gives some valuable information. First, the evaluation indexes of the four models on the training set (the 75th day of 2002) are higher than the accuracy on the test set (the 89th day of 2002), which can be encountered, which represents the fitting ability and generalization ability of the model, showing that the model has a certain generalization ability and is not as good as the ability of the training set.

[0064] Second, the horizontal comparative analysis shows that the LSTM classifier exhibits the best overall performance. As shown in Figure 5 , Figure 6 shown, the model maintains stable performance in multiple index evaluations of fire and non-fire areas; although the Transformer has the largest parameter size (followed by LSTM, MLP, and LightGBM), its performance far exceeds the data support of the scale of this study. Focusing on the LSTM model, it leads in key indicators such as recall rate, F1 value, and AUC, except for the accuracy rate affected by class imbalance; in particular, it improves the fire area detection accuracy by nearly 20 percentage points compared to the Transformer, although it is 8 percentage points lower than the best-performing Transformer in non-fire area detection, but its performance can also reach more than 86%. Although MLP and LightGBM have basic fire identification capabilities, their stability and accuracy are significantly weaker than LSTM and Transformer.

[0065] Finally, the information of concern is that the meteorological field predicted using the AI meteorological large model can achieve accuracy comparable to the real meteorological field. By comparing the relevant indicators of the two models on real and predicted meteorological data, it can be found that they only have a slight gap. This also verifies the accuracy of using the AI meteorological large model to predict the weather and the feasibility of using the predicted meteorological data for fire prediction from another angle.

Claims

1. A multi-source dynamic early warning method for forest fire risk that combines large meteorological models with lightweight AI, characterized in that... Includes the following steps: Step 1: Collect data from multiple sources and preprocess the data: Step 2: Based on the preprocessed multi-source data, train a multi-level forest fire probability prediction model, and achieve cross-level collaborative computing by introducing probability pixels; The multi-level forest fire probability prediction model is implemented as follows: First, the standardized feature vector and probability pixels are concatenated along the feature dimension. A sinusoidal positional encoding is then added to the concatenated tensor and input into the feature extraction layer. This layer employs a Transformer network architecture, using its multi-head attention mechanism to extract spatial correlation features and output an initial fire probability distribution. Subsequently, the probability correction layer deploys three sub-modules in parallel: a Long Short-Term Memory (LSTM) network, a Lightweight Gradient Boosting Machine (LightGBM), and a Multilayer Perceptron (MLP). Each module independently receives the output feature vector from the feature extraction layer and generates three corrected probabilities. Finally, the decision fusion layer implements a weighted voting mechanism, integrating the three corrected probabilities from the probability correction layer using preset weight coefficients to generate the final fire probability. The probability pixels are used as pre-built input components. The second dimension of the initial input probability pixels is an integer multiple of the dimension of the standardized feature vector. Before being input into the network, the probability pixels are divided into tensors of the same dimension as the standardized feature vector along the second dimension, and then concatenated with the standardized feature vector along the first dimension. The concatenated composite tensor is input into the Transformer network. After being processed by the multi-head attention layer, it is split into two sets of data along the second dimension and reconstructed into the form of the original probability pixels and the standardized feature vector. The two sets of reconstructed data are respectively input into the parameter-independent feedforward neural network layer inside the Transformer. Step 3: Integrate and optimize the output of the AI ​​meteorological big data model with the prediction results of the multi-level forest fire probability prediction model to generate the final forest fire occurrence probability.

2. The multi-source forest fire risk dynamic early warning method based on the collaboration of a large meteorological model and lightweight AI as described in claim 1, is characterized in that, In step 1, the multi-source data includes the daily average Copernicus digital elevation model, the normalized difference vegetation index (NDVI) and the enhanced vegetation index (EVI), fire point brightness and temperature data, and the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis data on precipitation, surface temperature, wind speed in longitude and latitude, mean sea level pressure, and temperature, humidity, geopotential, and wind speed in longitude and latitude of the pressure layer at vertical height.

3. The multi-source dynamic forest fire risk early warning method based on the collaboration of a large meteorological model and lightweight AI as described in claim 2, is characterized in that... In step 1, the preprocessing includes: resampling all multi-source data to a uniform resolution according to the WGS-84 projection coordinate system, normalizing all data except for fire point brightness and temperature data, recording the maximum and minimum values ​​of different types of data during the study period for uniform normalization of historical and predicted data, and spatially aligning all input data according to a uniform resolution.

4. The multi-source dynamic forest fire risk early warning method based on the collaboration of a large meteorological model and lightweight AI as described in claim 3, is characterized in that... Step 1 further includes: expanding all multi-source data into one-dimensional data for training, using a spatial image discretization strategy to obtain multi-dimensional image data, determining effective land pixels based on the terrain elevation threshold in the daily average Copernicus digital elevation model, and converting all the key features related to forest fires into standardized feature vectors of environmental parameters and one-dimensional fire labels at a single location.

5. The multi-source forest fire risk dynamic early warning method based on the collaboration of a large meteorological model and lightweight AI according to claim 4, characterized in that, The multi-level forest fire probability prediction model includes a sequentially connected feature extraction layer, probability correction layer, and decision fusion layer.

6. The multi-source forest fire risk dynamic early warning method based on the collaboration of a large meteorological model and lightweight AI according to claim 5, characterized in that, Step 3 is specifically implemented as follows: using a static digital elevation model with constant time resolution and periodically updated vegetation index features, calling an artificial intelligence meteorological big data model to generate future meteorological data, repeatedly performing data preprocessing and discretization to obtain the data to be processed, and then using a multi-level forest fire probability prediction model to predict the probability of fire occurrence, i.e., a global fire probability distribution map.

Images