A spatio-temporal adaptive neural network forest fire inversion method

Abstract

The application discloses a kind of based on space-time adaptive neural network forest fire inversion method, it is related to artificial intelligence remote sensing field, first by comparing the difference between actual reflection signal in GNSS-R data and predicted signal, calculate the residual signal related to fire, effectively filter out non-fire related interference;STANN model is developed, the spatial feature processing ability of CNN and the time series data processing ability of RNN are fused, especially suitable for analyzing data with complex space-time characteristics;By dynamically adjusting the convolution kernel through the adaptive feature extraction layer, STANN can accurately extract the key features related to fire.The application not only improves the accuracy of forest fire detection, but also provides an efficient tool for fire management and emergency response.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence remote sensing, and in particular to a forest fire inversion method based on spatiotemporal adaptive neural networks. Background Technology

[0002] Forest fires, as a natural disaster, have a tremendous impact on ecosystems, biodiversity, and human society. With global climate change and increased human activity, the frequency and intensity of forest fires are rising worldwide. Therefore, effective fire monitoring and prediction systems are crucial for mitigating the impacts of these disasters. Traditional fire monitoring methods, such as satellite remote sensing and ground observation, while successful in some aspects, still suffer from problems such as insufficient timeliness and limited coverage.

[0003] In recent years, Global Navigation Satellite System Reflected Signal (GNSS-R) technology has attracted attention as an emerging remote sensing method due to its unique advantages. GNSS-R technology can provide high spatiotemporal resolution data, which is particularly important for monitoring and predicting dynamically changing natural phenomena, such as forest fires. However, extracting fire-related information directly from GNSS-R data remains a challenge because the data is affected by various factors, and the fire information contained in the signal may not be directly obvious. Summary of the Invention

[0004] To overcome the aforementioned problems in the existing technology, this invention proposes a forest fire inversion method based on spatiotemporal adaptive neural networks.

[0005] The technical solution adopted by this invention to solve its technical problem is: a forest fire inversion method based on spatiotemporal adaptive neural network, comprising the following steps:

[0006] Step 1, Data Preparation: Regularly grid the surface temperature, soil moisture, and surface reflectance data of the same area and time period;

[0007] Step 2, Calculation of forest fire reflectivity: Perform time series decomposition on the data obtained in Step 1, group and reconstruct the decomposed data according to the characteristics of the singular value vector, apply a nonlinear prediction model for fitting and prediction, and calculate the reflectivity residual.

[0008] Step 3: Input the soil moisture, surface temperature obtained in Step 1, and the reflectance residual obtained in Step 2 into the spatiotemporal adaptive neural network for training.

[0009] Step 4: Use standardized fire reflectivity, soil moisture, and surface temperature data as input to the model obtained in Step 3 to reconstruct the forest fire situation.

[0010] The above-mentioned forest fire inversion method based on spatiotemporal adaptive neural network includes the following step 1: calculating the surface reflectance. This step involves selecting GNSS-R point data with an incident angle less than 65° and an antenna gain greater than 0, filtering the data based on the quality label parameters in the metadata, and calculating the effective surface reflectance (SR). The specific formula for calculating the effective surface reflectance (SR) is as follows:

[0011]

[0012] Among them, P coh To delay the peak power of the Doppler graph, P t G t P is the equivalent isotropic radiated power of the transmitter. t G represents the transmit power of the right-hand circularly polarized navigation signal of a GNSS satellite. t For transmitter antenna gain, G r R is the receiver antenna gain, λ is the carrier wavelength of the transmitted signal, and R is the receiver antenna gain. ts and R rs These represent the distances from the transmitter to the ground reflection point and the distances from the receiver to the ground reflection point, respectively.

[0013] The above-mentioned forest fire inversion method based on spatiotemporal adaptive neural networks, specifically includes step 2 as follows:

[0014] Step 2.1: For the effective reflectance time series data of each grid, select the window length according to the time series characteristics and construct the trajectory matrix;

[0015] Step 2.2: Perform singular value decomposition on the trajectory matrix obtained in Step 2.1;

[0016] Step 2.3: Group the singular value vectors according to their characteristics. After grouping, reconstruct the components of each group independently.

[0017] Step 2.4: Apply a nonlinear prediction model for fitting and prediction;

[0018] Step 2.5, calculate the reflectance residual. The calculation formula is:

[0019]

[0020] Among them, Y r Y represents the reflectance residual. t This represents the true value of reflectivity. This represents the total predicted value at time t.

[0021] The above-mentioned forest fire inversion method based on spatiotemporal adaptive neural networks, specifically includes step 2.4: determining the orders p and q of the nonlinear prediction model using ACF and PACF plots; and fitting and predicting each component obtained in step 2.3, using the following formula:

[0022] y t =c+φ1y t-1 +φ2y t-2 +...+φ p y t-p +∈ c +θ1∈ t-1 +θ2∈ t-2 +...+θ q ∈ t-q

[0023] Where, φ i It is the autoregressive coefficient, θ j It is the moving average coefficient, ∈ t c is the error term, and c is the constant term;

[0024] Based on the eigenvalue contribution rate of each component, a weighted method is used to synthesize the prediction results:

[0025]

[0026] in, This represents the total predicted value at time t. w is the predicted value of the i-th component. i These are the corresponding weights.

[0027] The above-mentioned forest fire inversion method based on spatiotemporal adaptive neural network includes a spatiotemporal input layer, an adaptive feature extraction layer, a spatiotemporal fusion layer, a deep fully connected layer, and an output layer in step 3.

[0028] The above-mentioned forest fire inversion method based on spatiotemporal adaptive neural networks specifically includes step 3 as follows:

[0029] Step 3.1: The spatiotemporal input layer standardizes or normalizes the data and divides it into training set, validation set and test set. The input data is organized into a three-dimensional array, including time series and spatial features. The training set is used to train the model.

[0030] Step 3.2, the adaptive feature extraction layer uses an adaptive convolutional kernel W. adaptive Dynamic adjustments are made using backpropagation and gradient descent methods.

[0031] Step 3.3: The spatiotemporal fusion layer sets up an adaptive weighted fusion mechanism, dynamically adjusting the weights based on the importance of the features; spatial and temporal features are weighted and fused using the following formula:

[0032] Y fusion =αY T +βY S

[0033] Where α and β are weights adaptively adjusted based on feature importance; Y T As a time feature, Y S Spatial features;

[0034] Step 3.4: The deep fully connected layer performs in-depth feature analysis and learning. The specific formula is as follows:

[0035] Y FC =ReLU(W FC *Y fusion +b FC )

[0036] Among them, W FC and b FC Y represents the weights and biases of a deep fully connected layer. FC Indicates the output of the fully connected layer;

[0037] Step 3.5: The output layer uses the Softmax function for classification, with the following formula:

[0038] Y output =Softmdx(W output *Y FC +b output )

[0039] Among them, W output and b output These are the weights and biases of the output layer;

[0040] Step 3.6: Evaluate the model using the validation set. If the model accuracy meets the requirements, proceed to the next step. If the model accuracy does not meet the requirements, repeat steps 3.2-3.5.

[0041] The beneficial effect of this invention is that it first uses GNSS-R data to calculate fire-related residual signals, that is, it identifies potential fire events by comparing the difference between the actual reflected signal and the expected (or normal) signal; this residual method can effectively filter out non-fire-related signal interference, thereby improving the accuracy of fire detection.

[0042] This invention's STANN model combines the spatial feature processing capabilities of Convolutional Neural Networks (CNNs) with the time-series data processing capabilities of Recurrent Neural Networks (RNNs), making it particularly suitable for handling data with complex spatiotemporal characteristics. STANN's adaptive feature extraction layer dynamically adjusts the convolutional kernel to selectively extract key fire-related features from GNSS-R data. This adaptive capability makes STANN more efficient and accurate when processing complex and variable spatiotemporal data. The STANN model can learn and identify complex patterns related to fire events from GNSS-R residual data, thereby achieving effective inversion of forest fire data. This invention effectively improves the detection accuracy and prediction efficiency of forest fires, providing a more effective tool for fire management and emergency response. Attached Figure Description

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0045] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1 As shown in the figure, this embodiment discloses a forest fire inversion method based on GNSS-R residuals and spatiotemporal adaptive neural networks, which specifically includes the following steps:

[0047] (1) Data preparation

[0048] GNSS-R point data with an incident angle less than 65° and an antenna gain greater than 0 were selected, and the data were further filtered based on the quality label parameters in the metadata. Then, the effective surface reflectivity was calculated according to formula (1).

[0049]

[0050] In the formula, P coh To delay the peak power of the Doppler graph, P t G t P is the equivalent isotropic radiated power of the transmitter. t G represents the transmit power of the right-hand circularly polarized navigation signal of a GNSS satellite. t For transmitter antenna gain, G r R is the receiver antenna gain, λ is the carrier wavelength of the transmitted signal, and R is the receiver antenna gain. ts and R rs These represent the distances from the transmitter to the ground reflection point and the distances from the receiver to the ground reflection point, respectively.

[0051] Select surface temperature and soil moisture data for the same time period and the same area.

[0052] Regular gridding: The surface reflectance data, soil moisture and surface temperature data are regularly processed into daily data in 5km×5km grids.

[0053] (2) Calculation of forest fire reflectivity (GNSS-R residual)

[0054] ①Time series decomposition

[0055] For the effective reflectance time series data of each grid, a suitable window length L is selected based on the characteristics of the time series. Then, the trajectory matrix is ​​constructed according to equation (2):

[0056]

[0057] Perform singular value decomposition on the trajectory matrix:

[0058] X=U∑V T (3)

[0059] In the formula, U and V are orthogonal matrices, and Σ is a diagonal matrix containing singular values.

[0060] After time series decomposition, the components are grouped according to the characteristics of the singular value vectors. Typically, the singular vectors (components) corresponding to the first or first few singular values ​​are considered as trend terms, those appearing in pairs are considered as periodic terms, and the rest are considered as noise. Once the grouping is complete, the components of each group will be reconstructed independently.

[0061] For each group, reconstruct the components:

[0062]

[0063] In the formula, U i ∑ i V i It is the matrix corresponding to the i-th group.

[0064] ② Applying nonlinear prediction models

[0065] First, the orders p and q of the nonlinear prediction model are determined using ACF and PACF plots. Then, for each component, the Model(p,q) formula (4) is used for fitting and prediction.

[0066] y t =c+φ1y t-1 +φ2y t-2 +...+φ p y t-p +∈ t +θ1∈ t-1 +θ2∈ t-2+...+θ q ∈ t-q (5)

[0067] Where, φ i It is the autoregressive coefficient, θ j It is the moving average coefficient, ∈ t is the error term, and c is the constant term.

[0068] Based on the eigenvalue contribution rate of each component, a weighted method is used to synthesize the prediction results:

[0069]

[0070] in, It is the total predicted value at time t. w is the predicted value of the i-th component. i These are the corresponding weights.

[0071] Mean squared error and root mean square error are used to assess the accuracy of predictions.

[0072] ③ Calculation of reflectivity residual

[0073]

[0074] Among them, Y r Y is the reflectivity residual. t This represents the true reflectance value.

[0075] (3) Fire inversion based on the Spatio-Temporal Adaptive Neural Network (STANN) model

[0076] ① Spatiotemporal input layer settings

[0077] Data standardization or normalization. The data is split into training, validation, and test sets. The training set is used to train the neural network model; during this process, the model learns to recognize patterns and features in the data. The validation set is used to verify the model's performance after training; the validation set is not involved in the training process. After model training is complete, the model is validated using the validation set, and then the test set is used for final evaluation. The input data Y is organized into a three-dimensional array containing time series and spatial features.

[0078] ② Adaptive Feature Extraction Layer

[0079] To dynamically adjust the convolution kernel to adapt to the characteristics of the input data, an adaptive convolution kernel W is used. adaptive Its size and shape are dynamically adjusted based on the features of the input data through backpropagation and gradient descent. Convolution operations and the ReLU activation function are applied, as shown in the formula:

[0080] Z = ReLU(W) adaptive *Y+b) (8)

[0081] Where Y represents the input data, including 3D GNSS-R residual data, 3D surface temperature data, and 3D soil moisture data, W adaptive is the weight, b is the bias, and * is the convolution.

[0082] ③ Spatiotemporal fusion layer settings

[0083] In Spatiotemporal Adaptive Neural Networks (STANN), the extraction of spatial and temporal features is achieved by combining Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). CNNs are responsible for extracting spatial features; they process the input data using a set of learnable filters through convolutional layers, capturing spatial patterns in the data. These filters generate feature maps as they move across the input data, revealing key spatial features. CNNs also include pooling layers to reduce the spatial dimensionality of the feature maps while preserving important information, thus improving the robustness of feature extraction. On the other hand, RNNs are responsible for extracting temporal features; they process time-series data through their recurrent structure. The core of RNNs lies in their hidden states, which are the network's internal memory, capable of capturing sequence information observed so far. This allows RNNs to remember and utilize past information to influence the processing of current and future data points, thereby revealing key temporal features.

[0084] To combine spatial and temporal features, an adaptive weighted fusion mechanism is designed, dynamically adjusting the weights based on feature importance. The spatial and temporal features are then weighted and fused using the following formula:

[0085] Y fusion =αY T +βY S (9)

[0086] Here, α and β are weights that are adaptively adjusted based on feature importance. T As a time feature, Y S It is a spatial feature.

[0087] ④ Deep fully connected layer

[0088] To perform in-depth feature analysis and learning, a deep fully connected layer was designed, with the following formula:

[0089] Y FC =ReLU(W FC *Y fusion +b FC (10)

[0090] Among them, W FC and bFC These are the weights and biases for a deep fully connected layer.

[0091] ⑤ Output layer

[0092] The output layer has 2 neurons (two neurons for binary classification), and the Softmax function is used for classification, as shown in the formula:

[0093] Y output =Softmax(W output *Y FC +b output (11)

[0094] Among them, W output and b output These are the weights and biases of the output layer.

[0095] ⑥ Selecting the loss function and optimizer

[0096] For classification tasks, use the cross-entropy loss function and either the Adam optimizer or stochastic gradient descent. The learning rate is typically started at 0.001 or 0.01 and adjusted as needed.

[0097] ⑦ Training the model

[0098] During training, the number of iterations was set to 100, and the batch size was set to 64. The training data was traversed during training, with forward and backward propagation performed to update the weights.

[0099] ⑧ Model Evaluation

[0100] Evaluate the model on the validation set: Use metrics such as accuracy and recall to assess model performance. In this case, a model with 95% accuracy and 85% recall is considered to have excellent performance.

[0101] 9. Hyperparameter tuning

[0102] Adjust the learning rate, batch size, number of epochs, etc., based on the performance on the validation set. Finally, construct the optimal model.

[0103] ⑩ Fire inversion

[0104] Standardized spatiotemporal data such as fire reflectivity, soil moisture, and surface temperature are used as model inputs to invert the forest fire situation in the region.

[0105] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its spirit and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.

Claims

1. A forest fire inversion method based on spatiotemporal adaptive neural network, characterized in that, Includes the following steps: Step 1, Data Preparation: Regularly grid the surface temperature, soil moisture, and surface reflectance data of the same area and time period; Step 2, Calculation of forest fire reflectivity: Perform time series decomposition on the data obtained in Step 1, group and reconstruct the decomposed data according to the characteristics of the singular value vector, apply a nonlinear prediction model for fitting and prediction, and calculate the reflectivity residual. Step 3: Input the soil moisture, surface temperature obtained in Step 1, and the reflectance residual obtained in Step 2 into the spatiotemporal adaptive neural network for training. Step 4: Use standardized fire reflectivity, soil moisture, and surface temperature data as input to the model obtained in Step 3 to reconstruct forest fire conditions. Step 1, the method for calculating surface reflectivity, includes: selecting GNSS-R point data with an incident angle less than 65° and an antenna gain greater than 0; filtering the data based on the quality label parameters in the metadata; and calculating the effective surface reflectivity (SR). The specific formula for calculating the effective surface reflectivity (SR) is as follows: in, To delay the peak power of the Doppler graph, This represents the equivalent isotropic radiated power of the transmitter. This refers to the transmission power of the right-hand circularly polarized navigation signal of the GNSS satellite. For transmitter antenna gain, For receiver antenna gain, The carrier wavelength of the transmitted signal. and These represent the distances from the transmitter to the ground reflection point and the distances from the receiver to the ground reflection point, respectively. Step 2 specifically includes: Step 2.1: For the effective reflectance time series data of each grid, select the window length according to the time series characteristics and construct the trajectory matrix; Step 2.2: Perform singular value decomposition on the trajectory matrix obtained in Step 2.1; Step 2.3: Group the singular value vectors according to their characteristics. After grouping, reconstruct the components of each group independently. Step 2.4: Apply a nonlinear prediction model for fitting and prediction; Step 2.5, calculate the reflectance residual. The calculation formula is: in, Represents the reflectivity residual. This represents the true value of reflectivity. This represents the total predicted value at time t; The spatiotemporal adaptive neural network in step 3 includes a spatiotemporal input layer, an adaptive feature extraction layer, a spatiotemporal fusion layer, a deep fully connected layer, and an output layer.

2. The forest fire inversion method based on spatiotemporal adaptive neural network according to claim 1, characterized in that, Step 2.4 specifically includes: determining the orders p and q of the nonlinear prediction model using ACF and PACF plots; and fitting and predicting each component obtained in step 2.3, using the following formulas: in, It is the autoregressive coefficient. It is the moving average coefficient. c is the error term, and c is the constant term; Based on the eigenvalue contribution rate of each component, a weighted method is used to synthesize the prediction results: in, This represents the total predicted value at time t. It is the predicted value of the i-th component. These are the corresponding weights.

3. The forest fire inversion method based on spatiotemporal adaptive neural network according to claim 1, characterized in that, Step 3 specifically includes: Step 3.1: The spatiotemporal input layer standardizes or normalizes the data and divides it into training set, validation set and test set. The input data is organized into a three-dimensional array, including time series and spatial features. The training set is used to train the model. Step 3.2, the adaptive feature extraction layer uses an adaptive convolutional kernel. Dynamic adjustments are made using backpropagation and gradient descent methods. Step 3.3: The spatiotemporal fusion layer sets up an adaptive weighted fusion mechanism, dynamically adjusting the weights based on the importance of the features; spatial and temporal features are weighted and fused using the following formula: in, α and β The weights are adaptively adjusted based on the importance of the features; As a time feature, Spatial features; Step 3.4: The deep fully connected layer performs in-depth feature analysis and learning. The specific formula is as follows: in, and The weights and biases of the deep fully connected layer; Indicates the output of the fully connected layer; Step 3.5: The output layer uses the Softmax function for classification, with the following formula: in, and These are the weights and biases of the output layer; Step 3.6: Evaluate the model using the validation set. If the model accuracy meets the requirements, proceed to the next step. If the model accuracy does not meet the requirements, repeat steps 3.2-3.5.

Images