Tropical forest fire monitoring method based on multi-element composite deep learning
By constructing a multivariate composite deep learning model, integrating multi-source data on tropical forest fires, and utilizing convolutional neural networks and long short-term memory networks to capture spatial heterogeneity and seasonal temporal dependencies, the problem of insufficient fire prediction accuracy in existing technologies has been solved, achieving high-precision fire monitoring and prediction.
Patent Information
- Application Number
- CN202511108808.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-08
- Publication Date
- 2025-11-21
AI Technical Summary
Existing technologies are unable to effectively capture the nonlinear interactions and seasonal dynamics among the driving factors of tropical forest fires, and their ability to integrate spatial and temporal features is limited, resulting in limited prediction accuracy and making it difficult to meet the needs of precise prevention and control.
A multi-source deep learning model is constructed by integrating convolutional neural networks, long short-term memory networks, and attention mechanisms to combine the spatiotemporal features of multi-source data and build a seasonally optimized deep learning model to achieve high-precision fire monitoring and prediction.
It significantly improves the accuracy of fire prediction, provides the contribution distribution of driving factors and potential spread paths, offers more comprehensive decision support for the allocation of prevention and control resources, solves the model training problem for small sample seasons, and ensures the prediction stability during seasonal transitions.
Smart Images
Figure CN120995388A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of forest fire monitoring technology, and in particular to a tropical forest fire monitoring method based on multivariate composite deep learning. Background Technology
[0002] Due to their unique climatic conditions, tropical forest ecosystems exhibit significant seasonality in forest fire occurrence, with the driving factors (such as climate, vegetation, topography, and human activities) exhibiting complex and differing mechanisms of action across seasons. Current research on forest fires largely employs traditional statistical models (such as Logistic Regression) or single machine learning models (such as Random Forest). On the one hand, these models struggle to effectively capture the nonlinear interactions and seasonal dynamics among driving factors, lack adaptability to adjusting factor weights between dry and rainy seasons, and have limited ability to integrate spatial features (such as topographic heterogeneity) and temporal features (such as seasonal sequence dependence), resulting in prediction accuracy limited by spatiotemporal scales. On the other hand, model outputs are often single probability values, lacking a refined representation of potential fire spread paths and the contribution of driving factors, thus failing to meet the needs of precise fire prevention and control.
[0003] Therefore, this invention proposes a tropical forest fire monitoring method that can integrate spatiotemporal characteristics, dynamically adapt to seasonal changes, and improve prediction accuracy. Summary of the Invention
[0004] This invention aims to overcome the shortcomings of existing technologies and provide a method for dynamic monitoring and prediction of tropical forest fires based on spatiotemporal fusion deep learning. By integrating the spatiotemporal characteristics of multi-source data, a seasonally optimized deep learning model is constructed to achieve high-precision fire monitoring and prediction.
[0005] The technical solution of this invention: A tropical forest fire monitoring method based on multi-dimensional composite deep learning, comprising:
[0006] Step (1) Multi-source data acquisition and preprocessing: Collect historical forest fire data, meteorological data, vegetation data, topographic data and human activity data of the study area, standardize and normalize the data, and construct a spatiotemporal dataset;
[0007] Step (2) Feature extraction: Based on the preprocessed data, extract the seasonal fluctuation characteristics of meteorological factors, the spatiotemporal variation characteristics of vegetation index, the spatial heterogeneity characteristics of topographic factors, and the spatiotemporal distribution characteristics of human activity factors.
[0008] Step (3) Construct a multi-component deep learning model: Construct a deep learning model that integrates convolutional neural networks, long short-term memory networks and attention mechanisms. The convolutional neural network is used to extract spatial features, the long short-term memory network is used to capture seasonal dynamic temporal features, and the attention mechanism is used to strengthen the weights of key driving factors.
[0009] Step (4), Seasonal Model Training and Optimization: The model is trained using a spatiotemporal dataset, the hyperparameters of the model are adjusted using the Bayesian optimization algorithm, and the model performance is evaluated and optimized using cross-validation.
[0010] Step (5), Fire monitoring and risk prediction: Input the data to be predicted into the optimized model, output the probability of forest fire occurrence, and draw fire risk distribution maps for different seasons in conjunction with the geographic information system.
[0011] Optionally, in step (1), the historical forest fire data includes forest fire archive data and MODIS fire point data for the past 20 years; meteorological data includes monthly average temperature, monthly average rainfall, monthly average evapotranspiration, monthly average humidity, and monthly average wind speed; vegetation data includes normalized vegetation index; topographic data includes altitude, slope, and aspect; and human activity data includes building distance and road distance.
[0012] Optionally, in step (2), the seasonal fluctuation characteristics are obtained by calculating the mean, standard deviation and coefficient of variation of meteorological factors in different seasons;
[0013] Spatiotemporal variation characteristics were extracted using the sliding window method, which showed the rate of change and spatial distribution gradient of the normalized vegetation index over time.
[0014] Spatial heterogeneity characteristics are obtained through spatial interpolation of topographic factors and zonal statistics.
[0015] Optionally, step (2) also includes a seasonal sensitivity algorithm for driving factors, calculated using the following formula: Where S f,s X is the sensitivity index of factor f in season s. f,s Let X be the mean of factor f over season s. f Let f be the annual mean of factor f, and σ be the mean of factor f. f W is the annual standard deviation of factor f. s Seasonal weighting.
[0016] Optionally, in step (3), the fused convolutional neural network uses multi-scale convolutional kernels to extract micro-topographic features within a 3×3 km grid and macro-vegetation distribution features at the regional scale.
[0017] Optionally, in step (3), the long short-term memory network introduces a seasonal mask layer, which strengthens seasonal specific information by assigning dynamic weights to the input data of different seasons; the attention mechanism automatically assigns high weights to key factors based on the ranking of the importance of driving factors.
[0018] Optionally, in step (3), the multi-component deep learning model includes:
[0019] The spatial feature extraction module adopts a fusion convolutional neural network architecture, which includes three convolutional layers with kernel sizes of 3×3, 5×5, and 7×7, respectively. It is used to extract micro-topographic features such as slope combinations and macro-distribution features of regional vegetation within the grid.
[0020] The seasonal time series feature module uses a long short-term memory network. It takes seasonal sequence data of meteorological factors as input and strengthens seasonal specificity through a seasonal masking layer. The seasonal masking layer is set with a weight of 1.2 for the dry season and 0.8 for the rainy season.
[0021] Attention mechanism layer: Based on the ranking of the importance of driving factors, the intersection of the geographic detector and the random forest can be referenced to assign dynamic weights to key factors such as humidity in the dry season and road distance, with a weight range of 1.5-2.0.
[0022] Fusion output layer: Spatial features, temporal features and attention-weighted factor features are concatenated and output as the probability of forest fire occurrence through a fully connected layer.
[0023] Optionally, the seasonal time series feature module adopts a two-layer attention mechanism, and the calculation formula is as follows: Where qf is the weight of the key factor jointly identified by the geographic detector and the random forest, and rf,s is the Pearson correlation coefficient of factor f in season s.
[0024] Optionally, in step (4), the Bayesian optimization algorithm optimizes the learning rate, number of convolutional kernels, and dimension of the hidden layer of the long short-term memory network of the model; the cross-validation method adopts time series cross-validation to ensure the generalization ability of the model in different seasons.
[0025] Optionally, in step (4), a seasonal adaptive learning rate formula is introduced in the seasonal model training: Where η s Let η be the learning rate for season s, η0 be the initial learning rate, and N be the learning rate for season s. s Let N be the sample size for season s, and N be the mean sample size. s Let be the average absolute error of the seasonal forecast s, and 'a' be the adjustment coefficient, preferably 0.5.
[0026] Optionally, in step (5), the fire risk distribution map is superimposed with a probability grid generated by deep learning through spatial interpolation to output a refined risk level map with a resolution of 3×4 kilometers, and the seasonal dynamic change trajectory of high-risk areas is marked.
[0027] Compared with the prior art, this application includes at least one of the following beneficial technical effects:
[0028] This invention combines convolutional neural networks and bidirectional long short-term memory networks to simultaneously capture the spatial heterogeneity and seasonal temporal dependence of forest fires, significantly improving the prediction AUC value compared to single models. At the same time, the seasonal dynamic weight adjustment layer and multi-head attention mechanism enable the model to automatically adapt to the differences in the effects of factors in different seasons, effectively capturing the nonlinear coupling relationship of driving factors in different seasons and significantly improving the accuracy of fire prediction.
[0029] Furthermore, this invention not only outputs the probability of fire occurrence, but also provides the contribution distribution of driving factors and potential spread paths, providing more comprehensive decision support for the allocation of prevention and control resources. Through data augmentation and transfer learning, it solves the model training problem of small sample seasons, and the seasonal cross-validation accuracy reaches 0.82, ensuring the prediction stability during the seasonal transition period. Attached Figure Description
[0030] Figure 1 This is a flowchart of a tropical forest fire monitoring method based on multivariate composite deep learning according to the present invention. Detailed Implementation
[0031] The technical solutions of this disclosure will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. The components of the embodiments of this disclosure described and shown in the accompanying drawings can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely to represent selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.
[0032] Please see as follows Figure 1 As shown, a tropical forest fire monitoring method based on multivariate composite deep learning is proposed, including:
[0033] Step 1: Multi-source data acquisition and preprocessing
[0034] Data collection: Collect historical forest fire data (including archived data on fire occurrence time, location, and area, as well as MODIS fire point data) for the study area from 2000 to 2024, meteorological data (monthly average temperature, rainfall, evapotranspiration, humidity, and wind speed), vegetation data (normalized difference vegetation index), topographic data (elevation, slope, and aspect), and human activity data (distances between buildings and roads).
[0035] Spatiotemporal alignment: Using the WGS84 coordinate system, all data are unified into 3×4 km grid cells. The time dimension is divided into four seasons (spring March-May, summer June-August, autumn September-November, winter December-February) and dry season (November-April) and rainy season (May-October), forming a three-dimensional data structure of grid cells, seasons, and driving factors.
[0036] Data augmentation: For seasons with fewer fire samples, such as the rainy season, the SMOTE algorithm is used to generate synthetic samples. Gaussian noise is injected into meteorological data to enhance the robustness of the model. The data is standardized and normalized, and a spatiotemporal dataset is constructed.
[0037] Step 2: Dynamic Feature Extraction
[0038] Seasonal fluctuation feature extraction: Calculate the cumulative anomaly of monthly average temperature during the dry season (reflecting the deviation from the historical same period) and the skewness coefficient of rainfall during the rainy season (to identify abnormal rainfall events), and identify temperature abrupt change points (such as the inflection point of spring warming) through a sliding t-test.
[0039] Phenological abrupt change characteristics: The normalized vegetation index sequence was subjected to first-order difference to extract the vegetation cover abrupt change slope at the beginning of the rainy season (May) and the beginning of the dry season (November). The larger the absolute value of the slope, the more drastic the vegetation change.
[0040] Spatial interaction characteristics: Construct an interaction term of "slope × road distance" to reflect the accessibility of human activities in areas with steep terrain, and an interaction term of normalized vegetation index × humidity to reflect the coupling relationship between vegetation and water.
[0041] Dynamic weighting features: Based on historical data, the contribution decay coefficient of each factor in different seasons is calculated. For example, the contribution of road distance increases by 20% during the Spring Festival, which is used as a time decay feature.
[0042] In the above, the seasonal fluctuation characteristics are obtained by calculating the mean, standard deviation and coefficient of variation of meteorological factors in different seasons; the spatiotemporal variation characteristics are obtained by extracting the rate of change and spatial distribution gradient of the normalized vegetation index over time series using the sliding window method; and the spatial heterogeneity characteristics are obtained by spatial interpolation and zonal statistics of topographic factors.
[0043] As one implementation method, feature extraction also includes a driving factor seasonal sensitivity algorithm, the calculation formula of which is: Where S f,s X is the sensitivity index of factor f in season s. f,s Let X be the mean of factor f over season s. f Let f be the annual mean of factor f, and σ be the mean of factor f. f W is the annual standard deviation of factor f. s Seasonal weighting.
[0044] Step 3: Construct a multi-dimensional composite deep learning model
[0045] The spatial feature extraction module employs a fusion convolutional neural network architecture, containing three layers of dilated convolutions with a kernel size of 3×3 and an dilation rate of 1-3. It captures spatial correlation features within 3×3, 5×5, and 7×7 neighborhoods, outputting a 32-dimensional spatial feature vector after pooling layers. The fusion convolutional neural network uses multi-scale convolutional kernels to extract microscopic terrain features within a 3×3 km grid and macroscopic vegetation distribution features at the regional scale.
[0046] The seasonal temporal feature module employs a Long Short-Term Memory (LSTM) network. It takes seasonal sequence data of meteorological factors as input and enhances seasonal specificity through a seasonal masking layer. For example, the seasonal masking layer has a weight of 1.2 for the dry season and 0.8 for the rainy season. The LTM network incorporates a seasonal masking layer, assigning dynamic weights to the input data for different seasons to strengthen seasonal specificity information. The attention mechanism automatically allocates high weights to key factors based on the importance ranking of driving factors. This invention, by combining a convolutional neural network and a bidirectional LTM network, simultaneously captures the spatial heterogeneity and seasonal temporal dependence of forest fires, significantly improving the prediction AUC value compared to a single model. Furthermore, the seasonal dynamic weight adjustment layer and multi-head attention mechanism enable the model to automatically adapt to the differences in factor effects across different seasons, effectively capturing the nonlinear coupling relationships of driving factors in different seasons and significantly improving fire prediction accuracy.
[0047] The seasonal time-series feature module employs a two-layer attention mechanism, and the calculation formula is as follows: Where qf is the weight of the key factor jointly identified by the geographic detector and the random forest, and rf,s is the Pearson correlation coefficient of factor f in season s.
[0048] Attention mechanism layer: Based on the ranking of the importance of driving factors, the intersection of the geographic detector and the random forest can be referenced to assign dynamic weights to key factors such as humidity in the dry season and road distance, with a weight range of 1.5-2.0.
[0049] Fusion output layer: Spatial features, temporal features and attention-weighted factor features are concatenated and output as the probability of forest fire occurrence through a fully connected layer.
[0050] Step 4: Seasonal Model Training and Optimization
[0051] Dataset partitioning: The training set (2000-2016) and the test set (2017-2024) were divided in a 7:3 ratio, and seasonal stratified sampling was used to ensure that the sample proportions were consistent in each season.
[0052] Transfer learning training: First, a base model is trained using data from the entire year (50 iterations). Then, the base model weights are added seasonally for fine-tuning (20 iterations). During the dry season, the weights of fire samples are increased (1.5 times) to address the sample imbalance problem. Through data augmentation and transfer learning, the challenge of model training with small seasonal samples is solved. The seasonal cross-validation accuracy reaches 0.82, ensuring the predictive stability during seasonal transitions.
[0053] Hyperparameter optimization: An improved Bayesian optimization algorithm is used to optimize the learning rate, number of convolutional kernels, and the dimension of the long short-term memory network hidden layer. The cross-validation method adopts time series cross-validation to ensure the generalization ability of the model in different seasons. The objective function is "AUC×0.6+SCV×0.4". For the dry season model, the learning rate (0.001-0.005) and the number of CNN convolutional kernels (32-64) are optimized. For the rainy season model, the dimension of the LSTM hidden layer (128-256) and the number of attention heads (2-4) are optimized.
[0054] Performance evaluation: The model was evaluated using AUC (target ≥ 0.9), PA (target ≥ 0.85), and SCV (target ≥ 0.8). The SCV calculation method was cross-validation of accuracy between adjacent seasons (e.g., using the spring model to predict fires from late April to early May).
[0055] The above seasonal adaptive learning rate formula is introduced in the training of the seasonal model: Where η s Let η be the learning rate for season s, η0 be the initial learning rate, and N be the learning rate for season s. s Let N be the sample size for season s, and N be the mean sample size. s Let be the average absolute error of the seasonal forecast s, and 'a' be the adjustment coefficient, preferably 0.5.
[0056] Step 5: Fire Monitoring and Risk Prediction
[0057] Real-time monitoring: Input real-time updated meteorological data (temperature and humidity of the previous 3 days) and vegetation data (latest normalized vegetation index), and the model outputs the probability of fire occurrence in the next 7 days.
[0058] Seasonal forecasting: For each season, the model outputs the probability of fire occurrence in 3×4 km grid cells, and combines it with ArcGIS to generate a 5-level risk map.
[0059] Driver contribution analysis: By visualizing attention weights, a spatial distribution map of the contribution of key factors in each season is generated. For example, in spring, the high contribution factors are temperature and road distance.
[0060] Potential spread path prediction: Based on the grid cells with the highest prediction probability, the potential spread path of the fire within 48 hours is simulated by combining slope (slopes spread faster upwards) and wind speed (downwinds spread faster). The spread direction and speed are marked with arrows.
[0061] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.
Claims
1. A method for monitoring tropical forest fires based on multivariate composite deep learning, characterized in that, include: Step (1) Multi-source data acquisition and preprocessing: Collect historical forest fire data, meteorological data, vegetation data, topographic data and human activity data of the study area, standardize and normalize the data, and construct a spatiotemporal dataset; Step (2) Feature extraction: Based on the preprocessed data, extract the seasonal fluctuation characteristics of meteorological factors, the spatiotemporal variation characteristics of vegetation index, the spatial heterogeneity characteristics of topographic factors, and the spatiotemporal distribution characteristics of human activity factors. Step (3) Construct a multi-component deep learning model: Construct a deep learning model that integrates convolutional neural networks, long short-term memory networks and attention mechanisms. The convolutional neural network is used to extract spatial features, the long short-term memory network is used to capture seasonal dynamic temporal features, and the attention mechanism is used to strengthen the weights of key driving factors. Step (4), Seasonal Model Training and Optimization: The model is trained using a spatiotemporal dataset, the hyperparameters of the model are adjusted using the Bayesian optimization algorithm, and the model performance is evaluated and optimized using cross-validation. Step (5), Fire monitoring and risk prediction: Input the data to be predicted into the optimized model, output the probability of forest fire occurrence, and draw fire risk distribution maps for different seasons in conjunction with the geographic information system.
2. The tropical forest fire monitoring method based on multivariate composite deep learning according to claim 1, characterized in that, In step (1), the historical forest fire data includes forest fire archive data and MODIS fire point data for the past 20 years; meteorological data includes monthly average temperature, monthly average rainfall, monthly average evapotranspiration, monthly average humidity, and monthly average wind speed; vegetation data includes normalized vegetation index; topographic data includes altitude, slope, and aspect; and human activity data includes building distance and road distance.
3. The tropical forest fire monitoring method based on multivariate composite deep learning according to claim 2, characterized in that, In step (2), the seasonal fluctuation characteristics are obtained by calculating the mean, standard deviation and coefficient of variation of meteorological factors in different seasons; Spatiotemporal variation characteristics were extracted using the sliding window method, which showed the rate of change and spatial distribution gradient of the normalized vegetation index over time. Spatial heterogeneity characteristics are obtained through spatial interpolation of topographic factors and zonal statistics.
4. The tropical forest fire monitoring method based on multi-element composite deep learning according to claim 3, characterized in that, Step (2) also includes a seasonal sensitivity algorithm for driving factors, calculated using the following formula: Where S f,s X is the sensitivity index of factor f in season s. f,s Let X be the mean of factor f over season s. f Let f be the annual mean of factor f, and σ be the mean of factor f. f W is the annual standard deviation of factor f. s Seasonal weighting.
5. A tropical forest fire monitoring method based on multivariate composite deep learning according to claim 1, characterized in that, In step (3), the fusion convolutional neural network uses multi-scale convolutional kernels to extract micro-topographic features within a 3×3 km grid and macro-vegetation distribution features at the regional scale.
6. A tropical forest fire monitoring method based on multivariate composite deep learning according to claim 5, characterized in that, In step (3), the long short-term memory network introduces a seasonal mask layer, which strengthens seasonal specific information by assigning dynamic weights to input data of different seasons; the attention mechanism automatically assigns high weights to key factors based on the ranking of the importance of driving factors.
7. A tropical forest fire monitoring method based on multivariate composite deep learning according to claim 6, characterized in that, In step (3), the multi-component deep learning model includes: The spatial feature extraction module adopts a fusion convolutional neural network architecture, which includes three convolutional layers, to extract micro-topographic features and macro-distribution features of regional vegetation within the grid. The seasonal time series feature module uses a long short-term memory network, inputs seasonal sequence data of meteorological factors, and enhances seasonal specificity through a seasonal mask layer; Attention mechanism layer: Based on the ranking of the importance of driving factors, key factors are assigned dynamic weights; Fusion output layer: Spatial features, temporal features and attention-weighted factor features are concatenated and output as the probability of forest fire occurrence through a fully connected layer.
8. A tropical forest fire monitoring method based on multivariate composite deep learning according to claim 7, characterized in that, The seasonal temporal feature module employs a two-layer attention mechanism, and the calculation formula is as follows: Where qf is the weight of the key factor jointly identified by the geographic detector and the random forest, and rf,s is the Pearson correlation coefficient of factor f in season s.
9. A tropical forest fire monitoring method based on multivariate composite deep learning according to claim 8, characterized in that, In step (4), the Bayesian optimization algorithm optimizes the learning rate, number of convolutional kernels, and dimension of the hidden layer of the long short-term memory network; the cross-validation method adopts time series cross-validation.
10. A tropical forest fire monitoring method based on multivariate composite deep learning according to claim 9, characterized in that, In step (4), the seasonal adaptive learning rate formula is introduced in the seasonal model training: Where η s Let η be the learning rate for season s, η0 be the initial learning rate, and N be the learning rate for season s. s Let N be the sample size for season s, and N be the mean sample size. s Let be the average absolute error of the seasonal forecast for season s, and 'a' be the adjustment coefficient.