A forest fire temperature downscaling method and system based on machine learning and energy conservation constraints

Abstract

The application discloses a forest fire temperature downscaling method and system based on machine learning and energy conservation constraints, comprising the following steps: step one, data preparation and preprocessing; step two, temperature downscaling based on a machine learning algorithm; and step three, correction of the temperature downscaling result based on the energy conservation constraint. The application firstly obtains a 50 m spatial resolution sensor end brightness temperature by using an XGBoost machine learning model, and then calculates an energy conservation factor to correct the 50 m spatial resolution sensor end brightness temperature, so that the problem of distortion of the temperature downscaling result when the energy conservation principle is not considered is weakened, and the forest fire temperature downscaling precision is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing imagery, specifically relating to a method and system for downscaling forest fire temperature based on machine learning and energy conservation constraints. Background Technology

[0002] With the increasing frequency and intensity of global warming and extreme drought events, forest fires have become a major natural disaster threatening ecosystem security and the sustainable development of forestry resources. Utilizing satellite remote sensing technology to obtain temperature information of forest fire sites is crucial for fire monitoring and precise firefighting. Currently, mid-infrared and thermal infrared remote sensing data are widely used in forest fire monitoring. However, the spatial resolution of mid-infrared and thermal infrared sensors is generally low, making it difficult to detect the complex temperature changes within a fire zone.

[0003] Downscaling algorithms are a key technology for improving the spatial resolution of remote sensing data. They have been widely used in the field of temperature downscaling. The most commonly used temperature downscaling methods include temperature downscaling based on image fusion methods and temperature downscaling based on machine learning algorithms.

[0004] The principle of surface temperature downscaling based on image fusion methods is to fuse the rich spatial detail information of high spatial resolution optical data with the thermal information of low spatial resolution temperature data, thereby generating a temperature product that combines high spatial resolution and high thermal information accuracy.

[0005] However, temperature downscaling algorithms based on image fusion are suitable for scenarios with small spatial temperature changes and insignificant boundary effects. Their application in forest fire temperature downscaling, where temperature changes are drastic and boundary effects are significant, is not ideal.

[0006] The principle of surface temperature downscaling based on machine learning algorithms is to automatically learn the complex nonlinear mapping relationship from low spatial resolution temperature data and high spatial resolution auxiliary data to high spatial resolution temperature data, and then estimate the high spatial resolution temperature by combining the high spatial resolution auxiliary data.

[0007] However, the automatic learning process does not take into account the law of conservation of radiative energy that should be followed in spatial resolution conversion. That is, the sum of the radiative energy of the high spatial resolution pixels after downscaling should be equal to the radiative energy of the corresponding low spatial resolution pixels. Therefore, temperature downscaling based on machine learning algorithms still has the problem of result distortion.

[0008] Therefore, temperature downscaling algorithms based on image fusion are suitable for scenarios with small spatial temperature variations and insignificant boundary effects, but their application in forest fire temperature downscaling, where temperature changes are drastic and boundary effects are significant, is not ideal. While machine learning-based temperature downscaling algorithms can improve accuracy by automatically learning complex nonlinear mapping relationships, the automatic learning process does not consider the conservation of radiative energy during spatial resolution conversion, resulting in distortion of the downscaled temperature data.

[0009] In summary, existing temperature downscaling algorithms cannot be directly applied to downscaling temperatures in forest fires. Therefore, designing a temperature downscaling method suitable for forest fires, based on their temperature characteristics, to obtain detailed temperature information within the fire area remains an urgent problem to be solved for forest fire monitoring and precise firefighting. Summary of the Invention

[0010] To address the aforementioned issues, this invention discloses a method and system for downscaling forest fire temperature based on machine learning and energy conservation constraints.

[0011] To achieve the above objectives, the technical solution of the present invention is as follows:

[0012] A method for downscaling forest fire temperature based on machine learning and energy conservation constraints includes the following steps:

[0013] Step 1: Preparation of remote sensing images and auxiliary data:

[0014] Data on forest fires is collected; the forest fire data includes optical band data, near-infrared band data, and mid-infrared band data from satellite imagery, as well as slope aspect and slope data of the forest fire; wherein the spatial resolution of the near-infrared band data is H m, and the spatial resolution of the mid-infrared band data is L m; L > H

[0015] The forest fire data was resampled to obtain two sets of raster image data with spatial resolutions of H m and L m, respectively. The sensor end brightness temperature with a spatial resolution of L m was calculated based on the mid-infrared band data. The enclosed fire area was obtained based on the sensor end brightness temperature of L m. The raster image data includes seven effective factors: altitude, slope, aspect, normalized difference vegetation index (NDVI), land cover type, GF-4 satellite red band reflectance, and near-infrared band reflectance.

[0016] Step 2: Temperature downscaling based on machine learning algorithms:

[0017] A training set is constructed based on raster image data with spatial resolution of L m and corresponding sensor brightness temperature with spatial resolution of L m. An XGBoost temperature downscaling model is then constructed. The training set is input into the XGBoost temperature downscaling model to obtain the trained XGBoost temperature downscaling model. Raster image data with spatial resolution of H m is input into the trained XGBoost temperature downscaling model to obtain the sensor brightness temperature of each pixel with spatial resolution of H m.

[0018] Step 3: Correction of temperature downscaling results based on energy conservation constraints:

[0019] The brightness temperature of each pixel sensor end with spatial resolution of H m is corrected based on the energy conservation constraint.

[0020] ;

[0021] As the energy conservation constraint factor, For the corrected first Pixel temperature, The number of pixels in the raster image data with a spatial resolution of H m; For the first time before correction Individual pixel temperature.

[0022] Further improvements were made, with H=50 and L=400.

[0023] Further improvements are made, and the specific steps of step one are as follows:

[0024] Step 1-1: Obtain GF-4 satellite remote sensing images of the forest fire area. After radiometric calibration of the mid-infrared band, obtain radiance values. Obtain reflectance data for the optical and near-infrared bands through atmospheric correction. Calculate the Normalized Difference Vegetation Index (NDVI) based on the reflectance data for the red and near-infrared bands. Using the optical, near-infrared, and NDVI data as inputs, perform supervised classification using a support vector machine algorithm to obtain the land cover type. Download the 90 m spatial resolution digital elevation model (DEM) and calculate the aspect and slope data of the forest fire area in ArcGIS software. Resample the optical and near-infrared reflectance data, NDVI, land cover type, and aspect and slope data of the forest fire area to obtain two sets of raster image data with spatial resolutions of 400 m and 50 m, respectively.

[0025] Step 1-2: Calculation of brightness temperature at the sensor end:

[0026] Using the infrared radiance values ​​of the GF-4 satellite The brightness temperature of each pixel at the sensor end with a spatial resolution of 400 m was calculated using Planck's radiation law. :

[0027]

[0028] These are the infrared radiance values ​​of the GF-4 satellite, with K1 and K2 being 150158.90 and 3785.45 respectively.

[0029] Steps 1-3: Based on the brightness temperature data of the sensor end with a spatial resolution of 400 m, the adaptive threshold method is used to identify fire points. Using the fire points as seed points, the region growing algorithm is used to capture high-temperature pixels around the fire points to form a closed fire field range.

[0030] A further improvement is made to the auxiliary factor screening method in step one, as follows:

[0031] Based on a sensor-end brightness temperature and auxiliary dataset with a spatial resolution of 400 m, the correlation between each factor and the sensor-end brightness temperature was evaluated by calculating the Pearson correlation coefficient, and strongly correlated factors were obtained. Multicollinearity among the factors was detected by calculating the variance inflation factor, and multicollinear variables were eliminated, resulting in 7 effective factors, including altitude, slope, aspect, normalized difference vegetation index (NDVI), land cover type, and red band and near-infrared band reflectance of GF-4 satellite. The auxiliary dataset includes optical band data, near-infrared band data, and mid-infrared band data of GF-4 satellite imagery, as well as slope aspect and slope data of forest fires, normalized difference vegetation index (NDVI), and land cover type.

[0032] In a further improvement, in step two: auxiliary data with a spatial resolution of 400 m is used as the independent variable, and the corresponding sensor end brightness temperature data is used as the dependent variable. A sample set is obtained through spatial registration and extraction, and the sample set is divided into a training set and a test set.

[0033] A further improvement is made to step three, which involves an energy conservation constraint factor. The calculation method is as follows:

[0034]

[0035] in, , This indicates the brightness temperature (K) at the sensor end, with subscripts L and H representing low and high resolution, respectively. The fourth power of the brightness temperature at the low-resolution pixel sensor end. For the first time before correction The fourth power of the brightness temperature at the high-resolution pixel sensor end. .

[0036] A forest fire temperature downscaling system based on machine learning and energy conservation constraints is provided, wherein the forest fire satellite remote sensing image reconstruction system is used to run the above-mentioned forest fire temperature downscaling method based on machine learning and energy conservation constraints.

[0037] Advantages of this invention:

[0038] This invention first uses the XGBoost machine learning model to obtain the brightness temperature at the sensor end with a spatial resolution of 50 m. Then, it calculates the energy conservation factor to correct the brightness temperature at the sensor end with a spatial resolution of 50 m, which reduces the distortion of temperature downscaling results when the energy conservation principle is not considered, and effectively improves the accuracy of temperature downscaling in forest fires. Attached Figure Description

[0039] Figure 1 Flowchart for downscaling GF-4 IRS images.

[0040] Figure 2 This is a comparison image before and after downscaling. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments. Example 1

[0042] like Figure 1 The method for downscaling forest fire temperature based on machine learning and energy conservation constraints, as shown, includes the following steps:

[0043] Step 1: Data preparation and preprocessing.

[0044] Step 1-1: Preparation of Remote Sensing Imagery and Auxiliary Data. Taking the "10.17 Xintian County Forest Fire" as the research object, six GF-4 satellite remote sensing images from the period of the fire were downloaded. Preprocessing was performed on the mid-wave infrared data, including radiometric calibration. Atmospheric correction was used to obtain reflectance data in the optical and near-infrared bands, and vegetation indices were calculated using the reflectance data in the near-infrared and red bands. Based on the optical and near-infrared band reflectance and normalized vegetation index data, a support vector machine algorithm was used to perform supervised classification of the images to obtain land cover type data. A 90 m spatial resolution digital elevation model (DEM) was downloaded, and slope aspect and slope data were calculated in ArcGIS software. In this invention, the spatial resolutions required for training and running the temperature downscaling model are 400 m and 50 m, respectively; therefore, the auxiliary data underwent resampling. The original spatial resolution of the visible and near-infrared bands of the GF-4 satellite is 50 m, which can be directly used in the model running stage without upsampling. However, in the model training stage, to maintain resolution consistency, it needs to be downsampled to 400 m. In addition, auxiliary data such as DEM, land cover type, and NDVI were resampled to two spatial resolutions of 400 m and 50 m, respectively, to construct two raster datasets required for matching model training and operation. The final auxiliary data includes: terrain data (slope, aspect, elevation), land cover type (SVM classification results), normalized vegetation index, and GF-4 satellite visible and near-infrared reflectance.

[0045] Step 1-2: Calculation of brightness temperature at the sensor end. This is done using the infrared radiance values ​​from the GF-4 satellite obtained in Step 1-1. The brightness temperature at the sensor tip (400 m) was calculated using Planck's law of radiation, and the unit is K. Although the influence of atmospheric and other environmental factors is not considered, the brightness temperature at the sensor tip clearly reflects the surface temperature state, and therefore it is widely used in forest fire monitoring. Therefore, the forest fire temperature downscaling in this invention refers to the scale of the brightness temperature downscaling at the sensor tip. The brightness temperature at the sensor tip... The calculation method is shown in formula (1):

[0046] (1)

[0047] In the formula: These are the infrared radiance values ​​of the GF-4 satellite, where K1 and K2 are relative to the wavelength. Relevant constants, center wavelength of the mid-infrared band of the GF-4 satellite The value is 3.08 μm, and the corresponding values ​​for K1 and K2 are 150158.90 and 3785.45, respectively.

[0048] Steps 1-3: Fire point identification and fire range determination. Based on the sensor brightness temperature data obtained in Step 1-2, an adaptive threshold method is used for fire point identification. Using these data as seed points, a region growing algorithm is used to capture high-temperature pixels around the fire points, forming a closed fire range.

[0049] Step 2-1: Screening of input factors for the temperature downscaling model. Based on the sensor-end brightness temperature and auxiliary dataset with a spatial resolution of 400 m, the correlation between each factor and the sensor-end brightness temperature was evaluated by calculating the Pearson correlation coefficient, and strongly correlated factors were obtained. Multicollinearity among the factors was detected by calculating the variance inflation factor, and multicollinear variables were eliminated, finally resulting in 7 effective factors, including altitude, slope, aspect, NDVI, land cover type, and GF-4 satellite red light and near-infrared reflectance.

[0050] Step 2-2: Sample Set Construction. Using the seven effective factors (400 m) obtained in Step 2-1 as independent variables and the sensor end brightness temperature (400 m) obtained in Step 1-2 as the dependent variable, a sample set of 11,700 samples covering both the interior and exterior of the fire scene was obtained through spatial registration and extraction. The sample set was divided into a training set and a test set in a 7:3 ratio, which were used for training the XGBoost downscaling model in Step 2-3 and for accuracy testing, respectively.

[0051] Steps 2-3: Construction of the XGBoost Temperature Downscaling Model. This invention utilizes the open-source XGBoost machine learning library to train and build the XGBoost model within a Windows 10 operating system, PyCharm 2022.2 integrated development environment (IDE), and Python 3.10 software environment. Based on the sample set in Step 2-1, with the sensor-end brightness temperature as the dependent variable (target variable) and seven effective factors as independent variables, the independent and dependent variables are input into the XGBoost machine learning model. During training, cross-validation and grid search are used to optimize the model parameters to obtain the best predictive performance. The optimal parameters selected are as follows: ① `colsample_bytree`: represents the proportion of features randomly sampled when building a new weak learner (tree) in each iteration; in this invention, it is set to 0.8. ② `gamma`: represents the minimum instance weight required when branching to leaf nodes; a larger value indicates a more conservative algorithm; in this invention, it is set to 0. ③ `learning_rate`: represents the learning rate; a lower value indicates a less prone model to overfitting; in this invention, it is set to 0.01. ④ `max_depth`: represents the maximum depth of the weak learner (tree); a larger value indicates that the model will learn more specific features; in this invention, it is set to 8. ⑤ `min_child_weight`: represents the minimum sum of sample weights required for child nodes; a larger parameter can prevent overfitting. The model learns noise, which is set to 5 in this invention; ⑥ n_estimators: represents the total number of weak learners (regression trees) fitted in the model, i.e., the number of iterations, which is set to 1000 in this invention; ⑦ random_state: is the random seed, which is set to 100 in this invention to control the random behavior in the model and ensure that the results are repeatable; ⑧ reg_alpha and ⑨ reg_lambda are the weights of L1 and L2 regularization terms, respectively, used to constrain model complexity and prevent overfitting, and are set to 1 and 5 respectively; ⑩ subsample represents the proportion of samples used when training each tree, set to 0.8, which helps to improve the generalization ability of the model; other parameter settings of the XGBoost algorithm use default values, so they are not listed one by one. The model performance evaluation result of the constructed model on the test set is the coefficient of determination R. 2 =0.96, and the root mean square error (RMSE) is 0.62 K, indicating that the constructed XGBoost model has good fitting effect and high prediction accuracy. After saving the trained XGBoost model in .joblib format, the model is loaded in steps 2-4 using the joblib.load() function.

[0052] Step 2-4: Temperature downscaling based on the XGBoost temperature downscaling model. Input the 50 m resolution raster image with elevation, slope, aspect, NDVI, land cover type, and GF-4 satellite red and near-infrared reflectance into the XGBoost temperature downscaling model constructed in Step 2-3 to estimate the sensor-end brightness temperature at 50 m spatial resolution.

[0053] Step 3: Correction of temperature downscaling results based on energy conservation constraints.

[0054] Step 3-1: Energy Conservation Constraint Factor The calculation involves calculating the energy conservation of total thermal radiation energy of ground objects at different spatial scales during scale transformation. To ensure physical consistency, this constraint must be met to ensure the conservation of total thermal radiation energy of ground objects at different spatial scales. According to this principle, the total thermal radiation energy of a 400 m resolution pixel should be equal to the sum of the energy of its corresponding 50 m resolution pixel. This invention utilizes the Stefan-Boltzmann law to calculate the energy conservation relationship between the brightness temperature at the sensor end of the low-resolution (400 m) and corresponding high-resolution (50 m) sensors, thus constructing an energy conservation constraint factor. This is used to perform energy consistency correction on the downscaled brightness temperature results. Specifically, each 400 μm pixel corresponds to... For a high-resolution sub-pixel, the energy conservation condition can be expressed as equation (2). Based on the Stefan-Boltzmann law, the energy expression for thermal radiation is shown in equations (3) and (4), where: Image energy (W) Emissivity of ground features It is the Stefan-Boltzmann constant (natural constant, whose value is...) ), The brightness temperature (K) at the sensor end is given, with subscripts L and H representing low and high resolution, respectively. For the first High-resolution pixel energy, For low-resolution pixel energy, For the first The fourth power of the brightness temperature at the high-resolution pixel sensor end. The fourth power of the brightness temperature at the low-resolution pixel sensor end. ; Let L be the blackbody radiative surface area corresponding to the pixel, used to calculate the total thermal radiation energy of the pixel. The subscripts L and H represent low and high resolution, respectively. If the surface emissivity is uniform, and Then the energy conservation equation can be simplified to equation (5); from this, the energy conservation constraint factor can be derived. The calculation formula is shown in (6), where: The first before correction High-resolution pixel temperature, .

[0055] (2)

[0056] (3)

[0057] (4)

[0058] (5)

[0059] (6)

[0060] Step 3-2: Calculate the corrected high-resolution pixel temperature. Use the energy conservation constraint factor calculated in Step 3-1. The brightness temperature at the sensor end with a spatial resolution of 50 m is corrected using the method shown in equation (7). The final downscaling result is obtained after correction, as shown in equation (7). Figure 2 As shown.

[0061] (7)

[0062] In the formula: For the corrected first High-resolution pixel temperature, .

[0063] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. Other modifications can be easily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and shown herein.

Claims

1. A method for downscaling forest fire temperature based on machine learning and energy conservation constraints, characterized in that, Includes the following steps: Step 1: Preparation of remote sensing images and auxiliary data: Data on forest fires is collected; the forest fire data includes optical band data, near-infrared band data, and mid-infrared band data from satellite imagery, as well as slope aspect and slope data of the forest fire; wherein the spatial resolution of the near-infrared band data is H m, and the spatial resolution of the mid-infrared band data is L m; L > H The forest fire data was resampled to obtain two sets of raster image data with spatial resolutions of H m and L m, respectively. The sensor end brightness temperature with a spatial resolution of L m was calculated based on the mid-infrared band data. The enclosed fire area was obtained based on the sensor end brightness temperature of L m. The raster image data includes seven effective factors: altitude, slope, aspect, normalized difference vegetation index (NDVI), land cover type, GF-4 satellite red band reflectance, and near-infrared band reflectance. Step 2: Temperature downscaling based on machine learning algorithms: A training set is constructed based on raster image data with spatial resolution of L m and corresponding sensor brightness temperature with spatial resolution of L m. An XGBoost temperature downscaling model is then constructed. The training set is input into the XGBoost temperature downscaling model to obtain the trained XGBoost temperature downscaling model. Raster image data with spatial resolution of H m is input into the trained XGBoost temperature downscaling model to obtain the sensor brightness temperature of each pixel with spatial resolution of H m. Step 3: Correction of temperature downscaling results based on energy conservation constraints: The brightness temperature of each pixel sensor end with spatial resolution of H m is corrected based on the energy conservation constraint. ； As the energy conservation constraint factor, For the corrected first Pixel temperature, The number of pixels in the raster image data with a spatial resolution of H m; For the first time before correction Individual pixel temperature.

2. The forest fire temperature downscaling method based on machine learning and energy conservation constraints as described in claim 1, characterized in that, H=50, L=400.

3. The forest fire temperature downscaling method based on machine learning and energy conservation constraints as described in claim 1, characterized in that, The specific steps of step one are as follows: Step 1-1: Obtain GF-4 satellite remote sensing images of the forest fire area. After radiometric calibration of the mid-infrared band, obtain radiance values. Obtain reflectance data for the optical and near-infrared bands through atmospheric correction. Calculate the Normalized Difference Vegetation Index (NDVI) based on the reflectance data for the red and near-infrared bands. Using the optical, near-infrared, and NDVI data as inputs, perform supervised classification using a support vector machine algorithm to obtain the land cover type. Download the 90 m spatial resolution digital elevation model (DEM) and calculate the aspect and slope data of the forest fire area in ArcGIS software. Resample the optical and near-infrared reflectance data, NDVI, land cover type, and aspect and slope data of the forest fire area to obtain two sets of raster image data with spatial resolutions of 400 m and 50 m, respectively. Step 1-2: Calculation of brightness temperature at the sensor end: Using the infrared radiance values ​​of the GF-4 satellite The brightness temperature of each pixel at the sensor end with a spatial resolution of 400m was calculated using Planck's radiation law. : ； These are the infrared radiance values ​​of the GF-4 satellite, with K1 and K2 being 150158.90 and 3785.45 respectively. Steps 1-3: Based on the brightness temperature data of the sensor end with a spatial resolution of 400 m, the adaptive threshold method is used to identify fire points. Using the fire points as seed points, the region growing algorithm is used to capture high-temperature pixels around the fire points to form a closed fire field range.

4. The forest fire temperature downscaling method based on machine learning and energy conservation constraints as described in claim 1, characterized in that, The auxiliary factor screening method in step one is as follows: Based on the sensor-end brightness temperature and auxiliary dataset with a spatial resolution of 400 m, the correlation between each factor and the sensor-end brightness temperature was evaluated by calculating the Pearson correlation coefficient, and the strongly correlated factors were obtained. Multicollinearity among the factors was detected by calculating the variance inflation factor, and multicollinear variables were eliminated, resulting in seven effective factors, including altitude, slope, aspect, normalized difference vegetation index (NDVI), land cover type, and red band and near-infrared band reflectance of GF-4 satellite imagery. The auxiliary dataset included optical band data, near-infrared band data, and mid-infrared band data of GF-4 satellite imagery, as well as aspect and slope data of forest fires, normalized difference vegetation index (NDVI), and land cover type.

5. The forest fire temperature downscaling method based on machine learning and energy conservation constraints as described in claim 1, characterized in that, In step two: using auxiliary data with a spatial resolution of 400 m as the independent variable and the corresponding sensor end brightness temperature data as the dependent variable, a sample set is obtained through spatial registration and extraction, and the sample set is divided into a training set and a test set.

6. The forest fire temperature downscaling method based on machine learning and energy conservation constraints as described in claim 1, characterized in that, In step three, the energy conservation constraint factor The calculation method is as follows: ; in, , This indicates the brightness temperature (K) at the sensor end, with subscripts L and H representing low and high resolution, respectively. The fourth power of the brightness temperature at the low-resolution pixel sensor end. For the first time before correction i The fourth power of the brightness temperature at the high-resolution pixel sensor end. .

7. A forest fire temperature downscaling system based on machine learning and energy conservation constraints, characterized in that, The forest fire satellite remote sensing image reconstruction system is used to run the forest fire temperature downscaling method based on machine learning and energy conservation constraints as described in any one of claims 1-5.

Images