A brightness temperature prediction method based on multi-band joint of Himawari-8 time series data

Abstract

The application discloses a kind of based on Himawari-8 time series data multi-band joint brightness temperature prediction method, it is related to the field of forest fire monitoring, including cutting Himawari-8 satellite remote sensing data;Read the brightness temperature data of normal image element in Himawari-8 satellite remote sensing data;The brightness temperature data of the same image element before time t certain time sequence length is as input sequence;Brightness temperature data at time t is as label, as true value and prediction value are compared;After feature normalization processing is carried out to the extracted time series brightness temperature data, as the input of model, the defined label is as the output of model, these data are as training data, and the prediction model is trained;When actually predicting, first, clear sky image element is screened out, the time series brightness temperature data of these clear sky image element is as the input of prediction model, then the trained time series prediction model is used to predict brightness temperature.The application compared with simple average method, brightness temperature prediction result error is smaller, can reach the purpose of accurately predicting brightness temperature data.

Description

Technical Field

[0001] This application relates to the field of satellite data wildfire monitoring, specifically to a brightness temperature prediction method based on Himawari-8 time series data. Background Technology

[0002] To mitigate the losses to people's lives and property caused by wildfires, the power grid sector has conducted a series of satellite remote sensing wildfire monitoring operations. However, these operations have also revealed several problems. Currently, most satellite wildfire monitoring by the power grid sector relies on simple fixed threshold methods, resulting in low detection accuracy, missed detections, and false positives, leading to unsatisfactory practical application.

[0003] In satellite-based wildfire monitoring, leveraging the high temporal resolution of Himawari-8 satellite data, a multi-temporal approach can be used to detect fire points based on the differences and changes in brightness temperature values ​​across the mid-infrared and far-infrared bands. However, in time-series data forecasting, traditional methods such as simple averaging, which calculate the average of historical values ​​as the predicted value, are ineffective for brightness temperature prediction. This is because actual brightness temperature data exhibits a continuous upward and downward trend with daily temperature variations, resulting in significant errors between the predicted and actual values, failing to yield a predicted brightness temperature value close to the true value. Furthermore, brightness temperature values ​​predicted using single-band data have limited resilience to fluctuations, leading to inaccurate predictions. Summary of the Invention

[0004] This application provides a Himawari-8 brightness temperature prediction method based on time series data to solve the problem of large prediction errors in existing time series prediction methods.

[0005] The technical solution adopted in this application is as follows:

[0006] A brightness temperature prediction method based on Himawari-8 time series data includes the following steps:

[0007] Step 1: Obtain Himawari-8 satellite remote sensing data in NetCDF format for different months, different time periods within each day, and different latitudes and longitudes;

[0008] Step 2: Crop the Himawari-8 satellite remote sensing data according to the actual monitoring range;

[0009] Step 3: Read the brightness temperature data of normal pixels in the Himawari-8 satellite remote sensing data;

[0010] The normal pixel refers to a pixel that is not covered by clouds and is not subject to fire. When there are clouds or fires, the brightness temperature value of the pixel will be significantly lower or higher than the brightness temperature value under normal conditions. These abnormal data will affect the accuracy of model training, so these pixels should be excluded when creating the time-series brightness temperature dataset.

[0011] The brightness temperature data are brightness temperature data in three bands: 3.9μm, 10.4μm, and 11.2μm. This is because the brightness temperature values ​​in these bands will increase significantly when a fire occurs. At the same time, in order to ensure that there is enough training data, a sufficient amount of time-series brightness temperature data is extracted from pixels of different months, different time periods within each day, and different latitudes and longitudes for the training and testing of the model.

[0012] Step 4: Take the brightness temperature data of the same pixel before time t for a certain time series length as the input sequence, and the brightness temperature data at time t as the output predicted value;

[0013] Step 5: Use the brightness temperature data at time t obtained in Step 3 as a label and compare it with the true value and the predicted value.

[0014] Step 6: After performing feature normalization on the extracted temporal brightness and temperature data, use it as the input to the model and the defined labels as the output of the model. Use these data as training data to train the prediction model; the prediction model is a temporal convolutional network model.

[0015] Step 7: In actual prediction, it is necessary to first select clear sky pixels. Each time in the input time series should be a clear sky pixel. The time series brightness temperature data of these clear sky pixels are used as the input of the prediction model. Then, the trained time series prediction model is used to predict the brightness temperature.

[0016] Furthermore, cloud masks are obtained through a dynamic threshold algorithm. The cloud masks from eight consecutive time periods are superimposed. If any time period is a cloud pixel, then that pixel is considered a cloud pixel. The identified cloud pixels are excluded, and what remains are clear sky pixels.

[0017] Furthermore, the input size of the prediction model is 7×3, where 7 represents the length of the time series as 7 consecutive time periods, each time period being 10 minutes long, and 3 represents the brightness temperature data of 3 bands; the output is 1×3, where 3 represents the brightness temperature values ​​of the three bands.

[0018] The beneficial effects of adopting the technical solution of this application are as follows:

[0019] This application discloses a brightness temperature prediction method based on multi-band joint analysis of Himawari-8 time-series data. To reduce the prediction error of time-series brightness temperature data, pixel time-series brightness temperature data from different months, different time periods within each day, and different latitudes and longitudes under normal conditions (cloud-free, fire-free pixels) are selected as the training dataset. Furthermore, the multi-band brightness temperature data is input as a whole into the time-series prediction model. Convolution operations using real brightness temperature data from the 3.9μm, 10.4μm, and 11.2μm bands better utilize the inherent relationships between the bands. Then, by using historical time-series brightness temperature data, the brightness temperature value of pixels under normal conditions can be predicted more accurately. This method can be applied to the field of satellite wildfire monitoring, and the brightness temperature value predicted using this invention has extremely high accuracy. Attached Figure Description

[0020] Figure 1 A flowchart of a Himawari-8 brightness temperature prediction method based on time-series data provided in this application;

[0021] Figure 2 The image shows the actual brightness temperature variation curves of Himawari-8 in the 3.9μm, 10.4μm, and 11.2μm bands over a continuous 24-hour period, representing an embodiment of this application.

[0022] Figure 3 This is a 24-hour continuous brightness temperature variation curve of Himawari-8 in the 3.9μm band, according to an embodiment of this application.

[0023] Figure 4 This is a 24-hour continuous brightness temperature variation curve of Himawari-8 in the 10.4μm band, according to an embodiment of this application.

[0024] Figure 5 This is a 24-hour brightness temperature variation curve of the 11.2μm band of Himawari-8, which is an embodiment of this application.

[0025] Figure 6 The graph shows the prediction results using only the brightness temperature data in the 3.9μm band.

[0026] Figure 7 The graph shows the prediction results using only brightness temperature data in the 10.4μm band.

[0027] Figure 8 The graph shows the prediction results using only the brightness temperature data in the 11.2μm band. Detailed Implementation

[0028] The embodiments will be described in detail below, and the experimental results are shown in the accompanying drawings. The implementation methods described in the following embodiments do not represent all implementation methods consistent with this application. They are merely examples of systems and methods consistent with some aspects of this application as detailed in the claims.

[0029] See Figure 1 The flowchart below shows a Himawari-8 brightness temperature prediction method based on time-series data, provided in an embodiment of this application.

[0030] This application provides a Himawari-8 brightness temperature prediction method based on time-series data, which mainly includes the following steps:

[0031] Step 1: Obtain Himawari-8 satellite remote sensing data in NetCDF format.

[0032] Himawari-8 data can be obtained for free via FTP link (ftp.ptree.jaxa.jp). First, download the Himawari-8 satellite data in NetCDF format for different time periods.

[0033] Step 2: Based on the actual monitoring range, crop the Himawari-8 satellite remote sensing data.

[0034] The netCDF4 library in Python is used to read the raw data, and the raw data is cropped according to the latitude and longitude range of the monitored area.

[0035] Step 3: Read the brightness temperature data of normal pixels in the Himawari-8 satellite remote sensing data.

[0036] Extracting temporal brightness temperature data from normal pixels requires eliminating interference from cloud pixels and fire pixels. Cloud pixels can be identified by overlaying cloud masks from eight consecutive time periods; if a pixel is cloudy at any given time, it is considered a cloud pixel. Fire pixels can be confirmed using satellite fire products such as MODIS (Moderate-resolution Imaging Spectroradiometer) and VIRRS (Visible Infrared Imaging Radiometer). Then, pixels without clouds or fire are selected, and brightness temperature data in the 3.9μm, 10.4μm, and 11.2μm bands of Himawari-8 are read.

[0037] Step 4: Use the brightness temperature data of the same pixel at a certain time t before a certain time series as the input sequence.

[0038] The time series length is 7 consecutive time periods (Himawari-8 satellite time resolution is 10 min). The brightness temperature data of the same pixel before a certain time t is used as the input feature X of the time series prediction model. The size of the input X of the prediction model is 7×3.

[0039] Step 5: Use the brightness temperature data at a certain time t as a label and compare it with the true value and the predicted value.

[0040] The true brightness temperatures of the 3.9μm, 10.4μm, and 11.2μm bands at time t, read from the Himawari-8 raw data, are used as the label Y for prediction, meaning the output size of the prediction model is 1×3.

[0041] Step Six: The time-series brightness and temperature data extracted in Step Four are processed by feature normalization and used as input to the prediction model. The defined labels are used as the output of the model. These data are used as training data to train the prediction model.

[0042] For the time-series brightness-temperature dataset samples obtained in steps four and five, the sample data is split into training and testing sets in an 8:2 ratio, i.e., 8,800 data points are used for training and the remaining 2,200 data points are used for testing. Input features are normalized, and 5-fold cross-validation is performed to evaluate the model's predictive performance. The TCN model is a convolutional neural network, and the main parameters involved include the model's input size, learning rate, number of hidden layer nodes, and kernel size, among which the learning rate and kernel size are the most important parameters. The TCN model is built using the PyTorch framework, selecting the MSEloss loss function and the Adam optimizer. A set of commonly used values ​​are pre-set for the parameters to be optimized, and then parameter tuning is performed. The final model parameters are shown in Table 1. A TCN model is built based on the optimized parameters. The features X and labels Y of the training set are input into the built TCN model, which outputs the prediction results and variance (Mean Square Error, MSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) of the training set. The trained model parameters are then saved. Table 2 shows the model training results; the model's variance (MSE) is 0.31K. 2 The mean absolute error (MAE) is 0.28K, and the mean percentage absolute error (MAPE) is 0.10%.

[0043] Table 1. Some parameters involved in the TCN model.

[0044]

[0045]

[0046] Table 2. Results of 5-fold cross-validation experiments for the TCN model.

[0047]

[0048] Step 7: In actual prediction, it is necessary to first select clear-sky pixels, use the temporal brightness temperature data of these clear-sky pixels as input to the prediction model, and then use the trained temporal prediction model to predict the brightness temperature value.

[0049] Select a cloudless, fireless pixel (102.3°E, 25.7°N) and read the brightness temperature values ​​of this pixel in the 3.9μm, 10.4μm, and 11.2μm bands for a continuous 24-hour period on March 13, 2020. See the appendix for details. Figure 2 The input to the TCN model is obtained by processing the data, and then the data is fed into the trained TCN model to complete the brightness temperature prediction. See the appendix for the prediction results. Figure 3 , 4 The brightness temperature prediction curves for the 3.9μm, 10.4μm, and 11.2μm bands show good fitting results with the actual brightness temperature curves. The error between the predicted and actual values ​​is smaller at night, while during the day, due to the larger trend in brightness temperature variation and inaccurate data at certain time points, the predicted values ​​deviate more significantly from the actual values ​​at some time points, with a maximum error of 1.6K. The average absolute errors between the predicted and actual brightness temperature values ​​for the 3.9μm, 10.4μm, and 11.2μm bands are 0.27K, 0.21K, and 0.21K, respectively, with an average percentage error of less than 0.10%, and the prediction results for the three bands are basically consistent. Simultaneously, brightness temperature prediction was also performed using single-band data; the prediction results are shown in the appendix. Figure 6 , 7 The average absolute errors of the brightness temperature prediction results for the three bands of 3.9μm, 10.4μm, and 11.2μm are 0.71K, 0.33K, and 0.35K, respectively, which are not as good as the prediction results of the three bands simultaneously.

Claims

1. A brightness temperature prediction method based on multi-band joint analysis of Himawari-8 time series data, comprising the following steps: Step 1: Obtain Himawari-8 satellite remote sensing data in NetCDF format for different months, different time periods within each day, and different latitudes and longitudes; Step 2: Crop the Himawari-8 satellite remote sensing data according to the actual monitoring range; Step 3: Read the brightness temperature data of normal pixels in the Himawari-8 satellite remote sensing data; The normal pixel refers to a pixel that is not covered by clouds and has no fire occurrence; the brightness temperature data refers to the brightness temperature data of the three bands: 3.9μm, 10.4μm, and 11.2μm. Step 4: Use the brightness temperature data of the same pixel for a period of time before time t as the input sequence; Step 5: Use the brightness temperature data at time t as a label and compare it with the predicted value as the true value; Step 6: After performing feature normalization on the time-series brightness and temperature data, use it as the input to the model and the defined labels as the output of the model. Use these data as training data to train the prediction model; the prediction model is a temporal convolutional network model. Step 7: In actual prediction, it is necessary to first select clear sky pixels. Each time in the input time series should be a clear sky pixel. The time series brightness temperature data of these clear sky pixels are used as the input of the prediction model. Then, the trained time series prediction model is used to predict the brightness temperature.

2. The brightness temperature prediction method based on multi-band joint analysis of Himawari-8 time series data as described in claim 1, characterized in that, The cloud mask is obtained by using a dynamic threshold algorithm. The cloud masks of 8 consecutive time periods are superimposed. If any time period is a cloud pixel, then the pixel is regarded as a cloud pixel. The identified cloud pixels are excluded, and the remaining pixels are clear sky pixels.

3. The brightness temperature prediction method based on multi-band joint analysis of Himawari-8 time series data as described in claim 1, characterized in that, The input size of the prediction model is 7×3, where 7 represents the length of the time series, which consists of 7 consecutive time periods, each with a length of 10 minutes, and 3 represents the brightness temperature data of 3 bands; the output is 1×3, where 3 represents the brightness temperature values ​​of the three bands.

Images