WO2022085869A1 - Forest fire risk medium-range forecast device and method - Google Patents

Abstract

A forest fire risk medium-range forecast device and method, according to a preferred embodiment of the present invention, generate a forest fire risk medium-range forecast model optimized for the Korean Peninsula and make a forest fire risk medium-range forecast (i.e. weekly forecast) by using the generated forest fire risk medium-range forecast model, such that it is possible to minimize damage through efficient on-site preparedness for forest fires (such as forward deployment of firefighting resources, etc.), and further reduce the frequency of forest fires by providing information necessary for prevention of forest fires.

Description

Forest fire risk medium-term forecasting device and method

The present invention relates to a forest fire risk medium-term forecasting apparatus and method, and more particularly, generating a forest fire risk medium-term forecast model optimized for the Korean Peninsula, and using the generated forest fire risk medium-term forecast model, medium-term forecast (i.e., It relates to an apparatus and method for making a weekly forecast). This study is related to the establishment of an integrated forest fire risk forecasting system (No. 1405004074) using meteorological big data conducted under the supervision of the National Institute of Forest Science with the funds of the Korea Forest Service for 2018-2020.

As the number of dry days in the Korean Peninsula increases due to climate change, the risk of forest fires is increasing, such as a longer period of wildfires throughout the year. Currently, it is difficult to preemptively respond to the risk of forest fires as only short-term forecasts are being performed for forest fire risk forecasting.

An object of the present invention is to create a medium-term forecasting model of forest fire risk optimized for the Korean Peninsula, and a medium-term forecasting device for forest fire risk that uses the generated medium-term forecasting model for forest fire risk to make a medium-term forecast (ie, weekly forecast) for the risk of forest fire and to provide a method.

Other objects not specified in the present invention may be additionally considered within the scope that can be easily inferred from the following detailed description and effects thereof.

A forest fire risk medium-term forecasting device according to a preferred embodiment of the present invention for achieving the above object, using daily data of a preset past period as training data, based on a machine learning algorithm, The medium-term forest fire risk medium-term forecast model, which uses the forest fire risk index, drought index, and weather forecast data obtained from the Global Data Assimilation and Prediction System (GDAPS) as input variables for the past 7 days, and uses medium-term forecast index values as output variables a model generator generating each forecast day; and a forecasting unit that acquires a medium-term forecast index value for each medium-term forecast date for the forest fire risk based on the forecast date as a reference date based on the plurality of medium-term forecast models for forest fire risk generated for each medium forecast date; Containing, obtained from the GDAPS The used weather forecast data includes air temperature, precipitation amount, relative humidity, surface temperature and wind speed, and the precipitation amount is a cumulative value for the past 7 days, and the remainder is an average value for the past 7 days.

Here, the model generation unit generates the forest fire risk medium-term forecast model for each medium-term forecast date in real time by using the daily data of the input variable for a preset past period with the forecast date as the reference date as training data, The forecasting unit may obtain a medium-term forecast index value for each medium-term forecast date for the forest fire risk with the forecast date as a reference date, based on the plurality of forest fire risk medium-term forecast models generated in real time for each medium-term forecast date.

Here, the forest fire risk index acquisition unit for obtaining the forest fire risk index by using the forest fire frequent area map, the modified Fine Fuel Moisture Code (FFMC), the drought index and monthly weights; further comprising, the monthly weights, Using the ratio of monthly wildfire occurrences obtained based on all forest fires that have occurred in the past, the greater the number of monthly wildfire occurrences, the greater the weight can be given.

Here, the forest fire risk index acquisition unit may obtain the forest fire risk index through the formula (the wildfire frequent area map + 0.5) * (the modified FFMC) * (1.5 - the drought index) * (the monthly weight). there is.

Here, the modified FFMC is, in the conventional FFMC that calculates the FFMC based on the equilibrium moisture content, drying rate and minimum moisture content of the day calculated using relative humidity, precipitation measured at noon, temperature, and wind speed, FFMC initial value , the coefficient of the equation for calculating the precipitation reference value, the FFMC reference value, and the drying rate is corrected, the coefficient of the equation for calculating the moisture content of the previous day is corrected, the coefficient of the equation for calculating the FFMC value using the minimum moisture content is corrected, and at noon The accumulated precipitation per day is used instead of the measured precipitation, and the precipitation reference value may be larger than that of the conventional FFMC.

Here, a drought index obtaining unit for obtaining the drought index by using the soil moisture index downscaled to a grid size of 1 km, a Normalized Different Water Index (NDWI), and a Temperature Condition Index (TCI); may further include.

Here, the drought index obtaining unit may obtain the drought index through the formula 0.4 * (the downscaled soil moisture index) + 0.3 * (the NDWI) + 0.3 * (the TCI).

Here, using data upscaled to a grid size of 25 km as training data, based on a machine learning algorithm, TRMM (Tropical Rainfall Measuring Mission) precipitation data, ASCAT (Advanced SCATterometter) soil moisture data, NDVI (Normalized Difference Vegetation Index) ), Land Surface Temperature (LST), and Digital Elevation Model (DEM) as input variables, and Global Land Data Assimilation System (GLDAS) soil moisture data as output variables, a soil moisture index downscaling model is created, It may further include; a soil moisture index acquisition unit for obtaining the downscaled soil moisture index by inputting the input variable converted to the grid size into the soil moisture index downscaling model.

Here, the soil moisture index acquisition unit, using the data upscaled to a grid size of 25 km as training data, the TRMM precipitation data, the ASCAT soil moisture data, the NDVI, the LST, and the DEM as input variables, Using the GLDAS soil moisture data as an output variable, a first soil moisture index downscaling model is created, and the TRMM precipitation data, the NDVI, the LST and A second soil moisture index downscaling model is generated using the DEM as an input variable and the GLDAS soil moisture data as an output variable, and the input variable converted to a grid size of 1 km is used as the first soil moisture index downscaling model to obtain a first soil moisture index downscaled to a grid size of 1 km by inputting into A second soil moisture index may be obtained, and the part missing from the first soil moisture index may be replaced with the second soil moisture index to obtain the downscaled soil moisture index.

The medium-term forest fire risk forecasting method according to a preferred embodiment of the present invention for achieving the above object is a forest fire risk medium-term forecasting method performed by a forest fire risk medium-term forecasting device, and using daily data of a preset past period as training data Therefore, based on a machine learning algorithm, using as input variables the map of wildfire areas, altitude, forest fire risk index for the past 7 days, drought index, and weather forecast data obtained from the Global Data Assimilation and Prediction System (GDAPS), Generating a forest fire risk medium-term forecasting model for each medium-term forecast day, using the forecast index value as an output variable; and obtaining a medium-term forecast index value for each medium-term forecast date for the forest fire risk based on the forecast date as a reference date based on the plurality of medium-term forecast models for forest fire risk generated for each medium forecast date; Containing, obtained from the GDAPS The weather forecast data includes air temperature, precipitation amount, relative humidity, surface temperature and wind speed, wherein the precipitation amount is a cumulative value for the past 7 days, and the remainder is an average value for the past 7 days.

Here, the generating of the forest fire risk medium-term forecast model step includes using the daily data of the input variable for a preset past period with the forecast date as the reference date as training data, and generating the forest fire risk medium-term forecast model in real time for each medium forecast date. The medium-term forecast index value acquisition step is based on the plurality of forest fire risk medium-term forecast models generated in real time for each medium-term forecast date, and the medium-term forecast for the risk of forest fire with the forecast date as the reference date. It may consist of obtaining a daily medium-term forecast index value.

Here, obtaining the forest fire risk index by using the forest fire frequent area map, the modified Fine Fuel Moisture Code (FFMC), the drought index and monthly weights; further comprising, wherein the monthly weights are all past occurrences Using the ratio of the number of monthly wildfire occurrences obtained based on the forest fire, the greater the number of monthly wildfire occurrences, the greater the weight can be given.

The computer program according to a preferred embodiment of the present invention for achieving the above technical problem is stored in a computer-readable recording medium and executes any one of the above-described medium-term forest fire risk forecasting methods on the computer.

According to the medium-term forest fire risk forecasting apparatus and method according to a preferred embodiment of the present invention, a medium-term forecasting model for forest fire risk is generated optimized for the Korean Peninsula, and medium-term forecasting of forest fire risk (that is, using the generated forest fire risk medium-term forecasting model) Weekly forecasting), it is possible to minimize damage through efficient preparation for wildfires on the site (forward deployment of firefighting resources, etc.) and further reduce the frequency of wildfires by providing information necessary for forest fire prevention.

Effects of the present invention are not limited to the above-mentioned effects, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

1 is a block diagram for explaining a medium-term forest fire risk forecasting device according to a preferred embodiment of the present invention.

2 is a view for explaining a soil moisture index downscaling model according to a preferred embodiment of the present invention.

3 is a view for explaining the results of the soil moisture index downscaling model according to a preferred embodiment of the present invention.

4 is a view for explaining the results of the soil moisture index downscaling model excluding the ASCAT soil moisture data according to a preferred embodiment of the present invention.

5 is a view for explaining the final downscaled soil moisture index according to a preferred embodiment of the present invention. The left side of FIG. 5 shows the results of the first soil moisture index downscaling model, and the right side of FIG. 5 shows the first It shows that the missing part in the result of the soil moisture index downscaling model is replaced with the result of the second soil moisture index downscaling model.

6 is a view for explaining a drought factor according to a preferred embodiment of the present invention.

7 is a diagram for explaining a result of a drought index model to which weights are applied according to a preferred embodiment of the present invention.

8 is a view for explaining a map of a forest fire frequent area according to a preferred embodiment of the present invention.

9 is a view for explaining the reason for using the FFMC as a factor of the forest fire risk index according to a preferred embodiment of the present invention. (b) compares FFMC with the number of wildfires by 10 days in 2015.

10 is a view for explaining a result of comparing the spatial distribution of a modified FFMC according to a preferred embodiment of the present invention and a conventional FFMC.

11 is a view for explaining a result of comparing the number of forest fires per month between the modified FFMC and the conventional FFMC according to a preferred embodiment of the present invention.

12 is a view for explaining a drought index used to obtain a forest fire risk index according to a preferred embodiment of the present invention.

13 is a view for explaining a first monthly weight used to obtain a forest fire risk index according to a preferred embodiment of the present invention.

14 is a diagram for explaining a second monthly weight used to obtain a forest fire risk index according to a preferred embodiment of the present invention.

15 is a view for explaining a forest fire risk medium-term forecast model according to a preferred embodiment of the present invention.

16 is a view for explaining the accuracy comparison results of the forest fire risk medium-term forecasting model according to a preferred embodiment of the present invention.

17 is a diagram for explaining the importance of input variables of a mid-term forest fire risk forecasting model generated based on a random forest according to a preferred embodiment of the present invention.

18 is a view for explaining the results of a forest fire risk medium-term forecast model generated based on a random forest according to a preferred embodiment of the present invention.

19 is a view for explaining the accuracy of the forest fire risk medium-term forecast model generated in real time according to a preferred embodiment of the present invention.

20 is a flowchart for explaining a medium-term forest fire risk forecasting method according to a preferred embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. Advantages and features of the present invention, and a method for achieving them will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments published below, but may be implemented in various different forms, and only these embodiments make the publication of the present invention complete, and common knowledge in the art to which the present invention pertains It is provided to fully inform those who have the scope of the invention, and the present invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout.

Unless otherwise defined, all terms (including technical and scientific terms) used herein may be used with the meaning commonly understood by those of ordinary skill in the art to which the present invention belongs. In addition, terms defined in a commonly used dictionary are not to be interpreted ideally or excessively unless clearly defined in particular.

In the present specification, terms such as “first” and “second” are for distinguishing one component from other components, and the scope of rights should not be limited by these terms. For example, a first component may be termed a second component, and similarly, a second component may also be termed a first component.

In the present specification, identification symbols (eg, a, b, c, etc.) in each step are used for convenience of description, and identification symbols do not describe the order of each step, and each step is clearly Unless a specific order is specified, the order may differ from the specified order. That is, each step may occur in the same order as specified, may be performed substantially simultaneously, or may be performed in the reverse order.

In this specification, expressions such as “have”, “may have”, “include” or “may include” indicate the existence of a corresponding feature (eg, a numerical value, function, operation, or component such as a part). and does not exclude the presence of additional features.

In addition, the term '~ unit' as used herein means software or a hardware component such as a field-programmable gate array (FPGA) or ASIC, and '~ unit' performs certain roles. However, '-part' is not limited to software or hardware. '~' may be configured to reside on an addressable storage medium or may be configured to refresh one or more processors. Accordingly, as an example, '~' indicates components such as software components, object-oriented software components, class components, and task components, and processes, functions, properties, and procedures. , subroutines, segments of program code, drivers, firmware, microcode, circuitry, data structures and variables. The functions provided in the components and '~ units' may be combined into a smaller number of components and '~ units' or further separated into additional components and '~ units'.

Hereinafter, with reference to the accompanying drawings, it will be described in detail a preferred embodiment of the forest fire risk medium-term forecasting apparatus and method according to the present invention.

First, with reference to Figure 1 will be described with respect to the forest fire risk medium-term forecasting device according to a preferred embodiment of the present invention.

1 is a block diagram for explaining a medium-term forest fire risk forecasting device according to a preferred embodiment of the present invention.

Referring to Figure 1, the forest fire risk medium-term forecasting apparatus 100 according to a preferred embodiment of the present invention generates a forest fire risk medium-term forecast model optimized for the Korean Peninsula, and uses the generated forest fire risk medium-term forecast model to predict forest fire risk. Make medium-term forecasts (ie weekly forecasts).

To this end, the forest fire risk medium-term forecasting device 100 includes a soil moisture index acquisition unit 110 , a drought index acquisition unit 130 , a forest fire risk index acquisition unit 150 , a model generation unit 170 and a forecasting unit 190 ). may include

The soil moisture index acquisition unit 110 is based on a machine learning algorithm, TRMM (Tropical Rainfall Measuring Mission) precipitation data, ASCAT (Advanced SCATterometter) soil moisture data, NDVI (Normalized Difference Vegetation Index), LST (Land Surface Temperature) and A soil moisture index downscaling model is generated using a digital elevation model (DEM) as an input variable and global land data assimilation system (GLDAS) soil moisture data as an output variable.

Here, the machine learning algorithm may use one of a random forest (RF), support vector regression (SVR), and an artificial neural network (ANN). In particular, the present invention can generate a soil moisture index downscaling model using a random forest (RF).

At this time, the soil moisture index acquisition unit 110 uses, as training data, the upscaled data to a grid size of 25 km, which is the same as the grid size of the GLDAS soil moisture data, as training data for the unity of spatial resolution between the data. create

Then, the soil moisture index acquisition unit 110 obtains the soil moisture index downscaled to the grid size of 1 km by inputting the input variable converted to the 1km grid size to the soil moisture index downscaling model.

In more detail, the soil moisture index acquisition unit 110 uses the data upscaled to a grid size of 25 km as training data, and uses TRMM precipitation data, ASCAT soil moisture data, NDVI, LST and DEM as input variables, Using the GLDAS soil moisture data as an output variable, a first soil moisture index downscaling model may be generated.

Then, the soil moisture index acquisition unit 110 uses the upscaled data to a grid size of 25 km as training data, TRMM precipitation data, NDVI, LST, and DEM as input variables, and GLDAS soil moisture data as output variables. , it is possible to generate a second soil moisture index downscaling model. That is, the soil moisture index acquisition unit 110 may generate the second soil moisture index downscaling model by using the remaining input variables except for the ASCAT soil moisture data among the input variables of the first soil moisture index downscaling model.

Then, the soil moisture index acquisition unit 110 inputs the input variable converted to a grid size of 1 km into the first soil moisture index downscaling model to obtain a first soil moisture index downscaled to a grid size of 1 km, The input variable converted to a grid size of 1 km is input to the second soil moisture index downscaling model to obtain a second soil moisture index downscaled to a grid size of 1 km, and the missing part in the first soil moisture index is the second By substituting the soil moisture index, the final soil moisture index downscaled to a grid size of 1 km can be obtained.

The drought index acquisition unit 130 uses the soil moisture index downscaled to a grid size of 1 km obtained through the soil moisture index acquisition unit 110, NDWI (Normalized Different Water Index), and TCI (Temperature Condition Index) during drought get an index

That is, the drought index obtaining unit 130 may obtain the drought index through the following equation.

Drought Index = 0.4 * (Downscaled Soil Moisture Index) + 0.3 * (NDWI) + 0.3 * (TCI)

In more detail, the drought index acquisition unit 130 may acquire the drought index by using the downscaled soil moisture index, NDWI, and TCI, which are factors showing a high correlation with drought among the following drought-related factors.

- Soil moisture index downscaled to a grid size of 1 km

- NDDI (Normalized Different Drought Index)

- NDWI (Normalized Different Water Index)

- NMDI (Normalized Multi-band Drought Index)

- TCI (Temperature Condition Index)

- VCI (Vegetation Condition Index)

- TRMM (Tropical Rainfall Measuring Mission) 1 (accumulated precipitation per week) / TRMM 2 (accumulated precipitation in 2 weeks)

Of course, the drought index acquisition unit 130 may obtain three drought indices through the following three equations, and obtain an average value of the obtained three drought indices as the final drought index.

- Equation 1: The drought exponential equation described above

- Equation 2: Drought index = 0.4 * (downscaled soil moisture index) + 0.2 * (NDWI) + 0.4 * (TRMM)

- Equation 3: Drought index = 0.3 * (downscaled soil moisture index) + 0.2 * (NDWI) + 0.3 * (TRMM) + 0.1 * (TCI)

On the other hand, the drought index obtaining unit 130 may generate a drought index model using the One-Class SVM, and may obtain the drought index using the generated drought index model, but the drought index obtaining unit 130 according to the present invention ) obtains the drought index through the above equation.

Of course, the drought index acquisition unit 130 obtains the first drought index through Equation 1 above using the downscaled soil moisture index, NDWI, and TCI, and uses the One-Class SVM to generate the drought index model. A second drought index may be obtained through the process, and an average value of the obtained first and second drought indexes may be obtained as a final drought index. In this case, the drought index acquisition unit 130 may use the average value of the three drought indices obtained through Equations 1 to 3 above as the first drought index.

The forest fire risk index acquisition unit 150 obtains a forest fire risk index using the drought index and monthly weights obtained through the forest fire frequent area map, the modified Fine Fuel Moisture Code (FFMC), and the drought index acquisition unit 130 .

Here, the monthly weight may be given a greater weight as the number of monthly forest fires increases, using the ratio of the number of monthly wildfires obtained based on all forest fires that have occurred in the past. For example, the monthly weight is the first monthly weight that gives a greater weight as the number of monthly wildfires increases according to the ratio of the number of monthly wildfires based on all forest fires that have occurred in the past It may be a second monthly weight that weights only a specific month (March-May) with the highest number of cases. In particular, the present invention can obtain a forest fire risk index by using the first monthly weight.

And, the modified FFMC is the conventional FFMC that calculates the FFMC based on the equilibrium moisture content, drying rate and minimum moisture content of the day calculated using the relative humidity, precipitation measured at noon, temperature, and wind speed. say FFMC. That is, in the modified FFMC, the coefficient of the formula for calculating the FFMC initial value, the precipitation reference value, the FFMC reference value and the drying rate is modified in the conventional FFMC, the coefficient of the formula calculating the moisture content of the previous day is modified, and the minimum moisture content is used It refers to an FFMC in which the coefficient of the equation for calculating the FFMC value is modified, the accumulated precipitation per day is used instead of the precipitation measured at noon, and the precipitation reference value is larger than that of the conventional FFMC.

That is, the forest fire risk index acquisition unit 150 may acquire the forest fire risk index through the following equation.

Wildfire Risk Index = (Wildfire Hotspot Map + 0.5) * (Adjusted FFMC) * (1.5 - Drought Index) * (Monthly Weighted)

At this time, the forest fire risk index acquisition unit 150 obtains the first forest fire risk index through the above formula using the “first monthly weight” as the monthly weight, and uses the “second monthly weight” as the monthly weight to obtain the above The second wildfire risk index is obtained through the ceremony, and the average value of the obtained first wildfire risk index and the second wildfire risk index may be obtained as the final forest fire risk index.

The model generation unit 170 is based on the machine learning algorithm, the forest fire risk index for the past 7 days, the drought index acquisition unit 130 obtained through the forest fire frequent area map, altitude, and the forest fire risk index acquisition unit 150 A forest fire risk medium-term forecast model is created for each medium-term forecast day, using the drought index obtained through the program and the weather forecast data obtained from the Global Data Assimilation and Prediction System (GDAPS) as input variables and medium-term forecast index values as output variables. .

Here, the machine learning algorithm may use one of a random forest (RF), support vector regression (SVR), and a deep neural network (DNN). Random forest (RF) uses 500 trees, support vector regression analysis (SVR) uses a Gaussian kernel, and deep neural network (DNN) has 3 hidden layers and the number of neurons is 5, 4, 3 It can be composed of dogs, and the training function can use Levenberg-Marquardt, and the performance function can use cross-entropy. In particular, the present invention can generate a forest fire risk medium-term forecast model using a random forest (RF).

And, the weather forecast data obtained from GDAPS includes air temperature, precipitation, relative humidity, surface temperature and wind speed, and the precipitation is an accumulated value for the past 7 days, and the remainder (air temperature, relative humidity, surface temperature, wind speed) ) is the average value for the past 7 days.

In addition, the medium-term forecast date refers to the weekly forecast such as 1 day later, 2 days later, 3 days later, 4 days later, 5 days later, 6 days later, and 7 days later.

In this case, the model generating unit 170 may generate a forest fire risk medium-term forecasting model for each medium-term forecast day by using the daily data of the preset past period as training data. For example, by using daily data on input variables for a specific past period (July 1, 2016 to May 30, 2017) as training data, a medium-term forecasting model for forest fire risk may be generated for each medium-term forecast date.

On the other hand, the model generator 170 may generate a forest fire risk medium-term forecasting model for each medium-term forecast date in real time by using daily data of input variables for a preset past period with the forecast date as the reference date as training data. . For example, in consideration of GDAPS weather forecast data that is periodically updated, daily data on input variables for the past 30 days as of the forecast date are used as training data to generate a medium-term forecasting model for forest fire risk by medium-term forecast date in real time. .

The forecasting unit 190 obtains a medium-term forecast index value for each medium-term forecast day for the forest fire risk with the forecast date as a reference date, based on a plurality of medium-term forecast models for forest fire risk generated for each medium forecast date.

On the other hand, when the forest fire risk medium-term forecasting model is generated in real time through the model generating unit 170, the forecasting unit 190 based on a plurality of forest fire risk medium-term forecasting models generated in real time for each medium-term forecast date, It is possible to obtain the medium-term forecast index value for each medium-term forecast date for the risk of forest fire as the reference date.

Then, a soil moisture index downscaling model according to a preferred embodiment of the present invention will be described in more detail with reference to FIGS. 2 to 5 .

In the present invention, GLDAS (Global Land Data Assimilation System) soil moisture data, ASCAT (Advanced SCATterometter) soil moisture data, NDVI (Normalized Difference Vegetation Index), Soil moisture field observation data after constructing a soil moisture index downscaling model using 5-day and 7-day cumulative precipitation data using LST (Land Surface Temperature), DEM (Digital Elevation Model), and daily TRMM (Tropical Rainfall Measuring Mission) precipitation data to verify the model.

First, GLDAS used as soil moisture data is a data assimilation data system based on three surface models of Mosaic, Noah, and Community Land Model. In the present invention, Noah model was used among the three surface models. The Noah model provided the moisture data of the 1 cm ~ 10 cm soil layer with a spatial resolution of 25 km 4 times a day (03, 09, 15, and 12:00), and the corresponding soil moisture data (m ³ /m ³ ) was averaged daily and used. For soil moisture data calculated by ASCAT, daily ascending and descending pass data were averaged and used. As auxiliary variables, MOD13A2, which is 16-day synthetic NDVI data with 1 km spatial resolution calculated from MODIS (Moderate Resolution Imaging Spectroradiometer) of Terra satellite, and MOD11A2 output, which is 8-day synthetic LST data with 1 km spatial resolution, was used. In addition, the DEM used the global DEM data of the Shuttle Radar Topography Mission (SRTM) with a spatial resolution of 90 m. In the case of precipitation data, TRMM 3B42 daily data with a spatial resolution of 25 km were accumulated over a period of 5 and 7 days, respectively. The soil moisture field observation data (%) provided by the Agricultural Promotion Agency was used to validate the model after unit conversion to m ³ /m ³ . All data except field observation data were used after masking in the range of the study area.

2 is a view for explaining a soil moisture index downscaling model according to a preferred embodiment of the present invention.

Referring to FIG. 2 , in the present invention, a downscaling model for soil moisture detailing was developed using various machine learning and artificial intelligence techniques. As the machine learning and artificial intelligence methods used, three techniques were used, respectively: random forest (RF), support vector regression analysis (SVR), and artificial neural network (ANN). The downscaling model was developed for each technique by setting the GLDAS soil moisture data as the dependent variable and setting a total of five data except the soil moisture field observation data as independent variables. For the unity of spatial resolution between data, all input data were applied to model building after upscaling to a spatial resolution of 25 km, the same grid size as GLDAS. After that, the 5 input data converted to the 1km grid size is applied to the developed model to finally obtain the downscaling soil moisture result of the 1km grid. In the present invention, the model was constructed by dividing the training data and the validation data by dividing the data from 2013 to 2014 at a ratio of 8:2, and additionally 10-fold cross-validation was performed with data for the same period. In addition, the soil moisture downscaling result produced through each model was verified by comparing it with the field observation soil moisture data of the Agricultural Promotion Administration.

3 is a view for explaining the results of the soil moisture index downscaling model according to a preferred embodiment of the present invention.

To find the optimal method for the soil moisture index downscaling model in the Korean Peninsula, three methods (RF, SVR, and ANN) were used to develop and compare the downscaling model. 3 is a scatter plot showing the soil moisture downscaling results of each model. In the order of RF, ANN, and SVR, the correlation coefficient value between the actual variable value and the model predicted value was large, and the RMSE (Root Mean Square Error) also showed a small value. In particular, the distribution of RF results was the narrowest in the scatterplot related to cross-validation. All three models tended to predict lower-than-actual values for high soil moisture values and high for low soil moisture values overall. Accordingly, the soil moisture index downscaling model according to the present invention was generated using RF.

Figure 4 is a view for explaining the results of the soil moisture index downscaling model excluding the ASCAT soil moisture data according to a preferred embodiment of the present invention, Figure 5 is the final downscaled soil moisture index according to a preferred embodiment of the present invention As a diagram to explain It represents substitution with the result of the downscaling model.

It can be seen that the results of the soil moisture index downscaling model are missing for the Jeju Island area and near some coastlines. This depends on the value of ASCAT soil moisture data with a spatial resolution of 25 km used as an input variable. In order to solve the omission problem, the soil moisture index downscaling model was additionally built with a total of four input variables excluding ASCAT from the input variables in the soil moisture downscaling method mentioned above. was to be replaced with the result of the additional model. 4 and 5 show examples of the resultant performance and soil moisture map of the additional model, respectively.

To verify the results of each soil moisture index downscaling model, comparison verification was performed with soil moisture data based on field observations provided by the Agricultural Promotion Administration. Time series analysis was performed to verify the soil moisture data of the observatory used for verification with the results of the soil moisture index downscaling model. As a result of time series analysis, it was confirmed that the soil moisture increase/decrease pattern was similar at all observation stations.

Then, the drought index according to a preferred embodiment of the present invention will be described in more detail with reference to FIGS. 6 and 7 .

The main factors that cause wildfires in Korea include misfires of mountain climbers and cigarette butts, and the occurrence of wildfires due to drought is extremely rare. However, if a forest fire occurs during a drought, the damage (area) may be increased. Therefore, in the present invention, the correlation between drought and forest fires was analyzed using data on actual wildfire occurrence of 1 ha or more from 2013 to 2018. Later, satellite data available in real time were used to develop the forest fire risk index. The drought-related factors used include the above-mentioned soil moisture downscaling data, NDDI (Normalized Different Drought Index), NDWI (Normalized Different Water Index) 5, 6, 7 (Divided into 5, 6, 7 depending on the use of SWIR band) , NMDI (Normalized Multi-band Drought Index), TCI (Temperature Condition Index), VCI (Vegetation Condition Index), TRMM 1 and 2 (accumulated precipitation for one and two weeks) data. The formula below is expressed based on MODIS (based on MODIS band).

NDDI = (NDVI - NDWI) / (NDVI + NDWI)

NDWI = (band2 - SWIR) / (band2 + SWIR)

NMDI = (band2 - (band6 - band7)) / (band2 + (band6-band7))

6 is a view for explaining a drought factor according to a preferred embodiment of the present invention.

As a result of analyzing the value distribution of each factor and index based on a forest fire occurrence area of 1 ha or more (left side of FIG. 6) and 10 ha or more (right side of FIG. 6), the same results as those of FIG. 6 were obtained. A value closer to 1 means a non-drought state, and a value closer to 0 means more severe drought. Above 1 ha, the range of 1-week cumulative precipitation (TRMM 1) and 2-week cumulative precipitation (TRMM 2) values showed the most drought phenomenon, and soil moisture (SM) and NDWI showed a correlation distribution in that order. NDDI, NMDI, and VCI showed values close to non-drought, indicating that the correlation was poor with the occurrence of forest fires (left side of FIG. 6). It showed the same correlation as 1 ha even at more than 10 ha, and relatively few cases of non-drought (right side of FIG. 6). For forest fire area of 50 ha or more (13 cases), the average value for each factor is soil moisture 0.37, NDDI 0.49, NDWI 5 0.32, NDWI 6 0.28, NDWI 7 0.38, NMDI 0.41, TCI 0.45, VCI 0.56, TRMM 1 0.01, TRMM 2 showed a greater correlation with drought, such as 0.02.

In the present invention, a drought index model was developed by applying a weight to the drought factor selected through correlation analysis and by applying One-Class SVM, one of the types of machine learning. Indices were developed using soil moisture, NDWI, TCI, NMDI, and TRMM, which showed a high correlation.

(1) weighting

Through correlation analysis, it was found that the precipitation factor had the greatest influence. However, in the case of precipitation, the drought index shows extreme values depending on the occurrence of precipitation, and drought or non-drought in all regions spatially. Therefore, as shown in [Table 1] below, soil moisture and NDWI, which had the highest correlation after precipitation, were combined with other factors.

Scheme	Equation
One	0.7 * Soil Moisture Index + 0.3 * NDWI
2	0.4 * Soil Moisture Index + 0.3 * NDWI + 0.3 * TCI
3	0.4 * Soil Moisture Index + 0.3 * NDWI + 0.3 * NMDI
4	0.4 * Soil Moisture Index + 0.2 * NDWI + 0.4 * TRMM
5	0.3 * Soil Moisture Index + 0.2 * NDWI + 0.3 * TRMM + 0.1 * TCI

7 is a diagram for explaining a result of a drought index model to which weights are applied according to a preferred embodiment of the present invention. 7 shows the results of each scheme for actual wildfire occurrence. Schemes 4 and 5 including precipitation factors showed a severe drought state compared to other schemes, and showed a drought state irrespective of the area damaged by the forest fire. However, even in areas where wildfires did not occur, drought conditions were indicated. In addition, since TCI has a relatively large correlation compared to NMDI, Scheme 2, which does not include a precipitation factor, and Scheme 4 and 5, which includes a precipitation factor, were used to develop the forest fire risk index.

(2) One-Class SVM application

One-Class SVM uses a hyperplane to classify one class and outliers in the same way as the existing binary classification and multi-classification SVMs. Based on the margin support vectors, internal vectors are assigned a class, and external vectors are identified as outliers.

When developing a machine learning model using the existing binary classification, the developer's subjective opinion about the sample extraction of non-existing forest fires is included, so in the present invention, we tried to develop a drought index related to the occurrence of forest fires through One-Class SVM. The kernel function used Gaussian, and the result was derived by normalizing the score calculated through the model. The results were compared by dividing the narrow range normalization (One-Class SVM 1, min: 0, max: 0.5) and wide range normalization (One-Class SVM 2, min: 0, max: 0.1). As a result of comparison, the smaller the area damaged by forest fires, the more non-drought in the One-Class SVM model, and the area of non-drought was higher than when weight was applied. However, as mentioned above, the occurrence of forest fires in Korea is largely due to human ignition, so wildfires can occur in the absence of drought. Therefore, the results of the One-Class SVM were also tested in the development of a forest fire risk index.

Then, the forest fire risk index according to a preferred embodiment of the present invention will be described in more detail with reference to FIGS. 8 to 14 .

8 is a view for explaining a map of a forest fire frequent area according to a preferred embodiment of the present invention.

In order to add information on places vulnerable to forest fires, a map of areas prone to forest fires provided by the National Academy of Forest Sciences was used. The map of areas prone to forest fires is a numerical map of location information for all forest fires (10,560 cases) that occurred from 1991 to 2015. This is the selected map. In order to select an area with a high incidence of forest fires, the average distance method and nearest neighbor analysis were performed to estimate the average distance between forest fires, and the density function first suggested by Kernel was used to finally select the area with a high incidence of forest fires. The causes of wildfires were divided into six categories: true stories of mountain climbers, incineration of paddy fields/field headlands, incineration of garbage, true stories of cemetery guests, true stories of cigarette fires, and others. From 1991 to 2015, an annual average of 422 wildfires occurred, and 2,102 ha of forest was burned every year. The region with the highest number of wildfires was Busan Metropolitan City (477 cases), followed by Seoul (412 cases), Incheon (391 cases), Ulsan (350 cases), Daegu (285 cases), Daejeon (263 cases), and Gwangju (203 cases). case), the incidence of wildfires was high in metropolitan areas (see Fig. 8). The most common month for wildfires occurred in April, with 3,250 wildfires occurring, accounting for 30.8% of the total. Looking at the statistical status of forest fires by cause, 4,392 cases of wildfires caused by mishaps by mountain climbers accounted for 41.6% of the total forest fires, followed by incineration of paddy fields/bedheads at 1,952 cases (18.5%), garbage incineration (830 cases (7.9%), and cigarette fires). was followed by 813 cases (7.7%).

9 is a view for explaining the reason for using the FFMC as a factor of the forest fire risk index according to a preferred embodiment of the present invention. (b) compares FFMC with the number of wildfires by 10 days in 2015.

The Canada Fire Weather Index (CFWI) is a Fine Fuel Moisture Code (FFMC), Duff Moisture Code (DMC), and Drought Code (DC) calculated using weather information such as temperature, relative humidity, wind speed and precipitation. , ISI (Initial Spread Index), and BUI (Build Up Index) information is a meteorological index calculated by synthesizing information. FFMC predicts the moisture content of the fine fuel in the forest on the ground, and DMC predicts the humidity of the surface fuel layer in the forest. DC predicts the likelihood of seasonal drought and geological formation by predicting the moisture of deep organic matter layers and coarse fuel in the ground, and ISI is calculated by combining FFMC and wind speed factors. Among them, FFMC predicts the moisture content of fine fuel using temperature, relative humidity, wind speed, and precipitation information, and the range is 0 to 99. A higher number means a higher probability of ignition. Among existing studies, when CFWI is applied to Korea, FFMC has a higher correlation than CFWI, which is the final output index, and as a result of predicting the probability of a forest fire with a regression analysis model using FFMC, it has a statistical significance of 5% ( Park Heung-seok et al., 2009). When CFWI and FFMC were compared with DWI (Daily Weather Index), an existing forest fire index in Korea, based on the 2015 forest fire (see FIG. 9), it was found that FFMC was more suitable for Korea than CFWI, and FFMC was used in the present invention. decided to do

10 is a diagram for explaining the result of comparing the spatial distribution of the modified FFMC according to the preferred embodiment of the present invention and the conventional FFMC, and FIG. 11 is the modified FFMC according to the preferred embodiment of the present invention and the conventional FFMC It is a diagram to explain the results of comparing with the number of forest fires per month.

9, it was determined that FFMC was more suitable for Korea than CFWI, so that FFMC was used in the present invention. FFMC was developed in the 1970s and has been continuously updated since then, and in order to create a more accurate forest fire risk index model, the process of optimizing the existing FFMC for the Korean environment was carried out. FFMC calculates the equilibrium moisture content, drying rate, and minimum moisture content of the day using relative humidity, precipitation, temperature, and wind speed data, and finally calculates the FFMC. Accordingly, in the present invention, FFMC was optimized by modifying some coefficients according to the Korean environment. That is, the formula coefficients for calculating the FFMC initial value, the precipitation standard value, the FFMC standard value, and the drying rate were modified, and the coefficients of the formula for calculating the moisture content of the previous day and the formula for calculating the FFMC using the minimum amount were modified. After determining the range of coefficients by comparing the first formula developed in the 1970s with the formula currently used, the coefficients were modified within the range of each coefficient, and the optimal coefficients were found by comparing with the number and area of wildfires in Korea. Conventional FFMC used precipitation measured at noon, but in the present invention, daily cumulative precipitation was used to make daily FFMC. The precipitation reference value was increased by changing the precipitation data from instantaneous precipitation to cumulative precipitation. 10 is a result of comparing the spatial distribution of the existing FFMC and the optimized FFMC (that is, the modified FFMC according to the present invention), and since precipitation determines the degree of drying and is an important variable in FFMC calculation, the FFMC most simulates the pattern of precipitation. I could see that a lot was followed. The FFMC was modified to suit the Korean environment, and the existing FFMC, which was heavily biased toward precipitation, was modified to more gently affect it.

11 is a result of comparing the existing FFMC and the optimized FFMC with the number of wildfires in Korea, and it can be seen that the optimized FFMC shows the time series better than the existing FFMC (refer to the purple arrow).

As shown in [Table 2] below, the CDF (Cumulative Distribution Function) value of wildfire pixels was calculated after calculating the FFMC index of all wildfire pixels and non-fired pixels for each wildfire day for the existing FFMC and the optimized FFMC. The average was derived and then compared. CDF is a probability that a random variable is less than or equal to a certain value for a certain probability distribution, for example, if the CDF is 0.95, it is a high index corresponding to the top 5%. Since the distribution of the existing FFMC and the optimized FFMC value is different, each index value was converted into a CDF and the accuracy was analyzed for comparison.

	wildfire area	>= 50ha	< 50 ha	< 10ha	< 1 ha
existing FFMC	FFMC average	87.0244	86.0494	83.8506	80.9751
	CDF average	0.769	0.791	0.575	0.429
optimized FFMC	FFMC average	60.9135	58.9741	51.8598	49.2367
	CDF average	0.972	0.978	0.960	0.924

Compared to the existing FFMC, the average CDF value of the optimized FFMC was higher than 0.9, and the larger the damage area, the higher the average CDF value. Through this, it can be seen that the optimized FFMC is more suitable for Korea than the existing FFMC.

The FFMC, modified for the Korean environment, was fused with the map of areas prone to forest fires and the drought index to develop a Fire Risk Index (FRI) suitable for the Korean environment. The existing Canadian Forest Fire Hazard Index (CFWI) calculation method is generated by multiplying a function for wind and FFMC, a function for DMC and DC, and a coefficient as follows.

CFWI = coefficient * f(wind,FFMC) * f(DMC,DC)

Therefore, the forest fire risk index (FRI) according to the present invention was developed using the revised FFMC (revised FFMC), drought idx, and forest fire frequent area map (frequency) as coefficients as follows, and considering the seasonal characteristics A process for giving a different weight (temporal weight) for each month has been added.

FRI = (frequency + 0.5) * (revised FFMC) * (1.5 - drought idx) * (temporal weight)

Here, the constant added to the forest fire frequency and drought idx is to prevent the situation in which the FRI becomes 0 when multiplied when each indices has a minimum value of 0. Through analysis and analysis of how many percent of all pixels on the same day belong to the top, the constant determined to best simulate the occurrence of a forest fire was finally used.

12 is a view for explaining a drought index used to obtain a forest fire risk index according to a preferred embodiment of the present invention.

12 is a comparison of the distribution of the previously developed forest fire-related drought indices for wildfire and nonfire pixels, and various previously developed drought indices were tested to develop a forest fire risk index, Among them, the distribution of drought indices that showed the most significant results is shown. After deriving the forest fire risk index using these drought indices, as a result of analyzing the CDF and upper percent, Scheme 2 among the drought indices was judged to be the most suitable for the development of the forest fire risk index. Accordingly, the drought index used to obtain the forest fire risk index according to the present invention was a drought index according to Scheme 2 among drought indices to which a weight was applied.

The method of giving the monthly weight (temporal weight) is the method of giving weight to have a larger FRI value as the number of monthly occurrences increases according to the ratio of monthly occurrences for all wildfires from 2014 to 2017 (FRI_M1) and monthly wildfires Two methods were tested (FRI_M2), in which only March to May with the highest number of occurrences were weighted.

To determine the suitability of FRI_M1 and FRI_M2, which were multiplied by monthly weights to consider seasonal characteristics, CDF analysis was performed similarly to comparing the existing FFMC and the modified FFMC. A probability that is less than or equal to a certain value. For comparison, after calculating the FFMC and FRI of all wildfire pixels and non-wildfire pixels for the days of wildfires from 2014 to 2017, the CDF values of the wildfire pixels were averaged and derived as shown in [Table 3] below.

	wildfire area	>= 50ha	< 50 ha	< 10ha	< 1 ha
existing FFMC	CDF average	0.769	0.791	0.575	0.429
Modified FFMC	CDF average	0.972	0.978	0.960	0.924
FRI_M1	CDF average	0.945	0.994	0.953	0.895
FRI_M2	CDF average	0.959	0.979	0.974	0.928

As a result, it was confirmed that the value was significantly improved compared to the existing FFMC, and when compared with the modified FFMC, the FRI showed similar or higher values to the modified FFMC, so it was found that the FRI value of the wildfire pixels was higher than that of the non-wildfire pixels. can be interpreted In order to understand the spatial simulation accuracy only on the day of the wildfire, not during the entire study period, the FFMC and FRI of each forest fire were analyzed to be among the top percent of the national forest fire risk index. As a result, as shown in [Table 4] below, compared to the existing FFMC, the proportion of forest fires within the top 5% of FRIs has increased nearly twice, and those within the top 40% increased from 58.55% to an average of 67.91%. did

difference%	<= 5%	<= 10%	<= 20%	<= 30%	<= 40%	<= 50%	> 50%	sum
existing FFMC	12.134	10.434	14.448	12.181	9.348	8.593	32.861	100
FRI_M1	20.608	10.161	16.144	11.728	9.497	7.028	24.834	100
FRI_M2	20.560	10.446	15.907	11.728	9.544	7.028	24.786	100

13 is a view for explaining a first monthly weight used to obtain a forest fire risk index according to a preferred embodiment of the present invention, Figure 14 is a second used to obtain a forest fire risk index according to a preferred embodiment of the present invention 2 It is a diagram for explaining monthly weights.

13 and 14 show the forest fire risk index (DWI) provided by the existing forest fire risk forecasting system of the Forestry Academy of the Korea Forest Service. As a result of comparing the weighted model (FRI_M1) and the weighted model (FRI_M2) for the period from March to May, when the number of forest fires occurred the most, with the number of wildfires in Korea from 2014 to 2017, both FRI_M1 and FRI_M2 were DWI Compared to , it showed a good agreement with the rising and falling time series pattern of the number of monthly wildfire occurrences. In particular, in the case of FRI_M1, the forest fire risk index (FRI) according to the present invention was higher than the existing forest fire risk index (DWI) according to the present invention. It can be seen that there is a better agreement with the time series shown by (see purple box). Therefore, as the monthly weight used to obtain the forest fire risk index according to the present invention, the first monthly weight weighted so that the FRI has a higher value as the number of monthly forest fires increases was used.

Then, with reference to FIGS. 15 to 19, the forest fire risk medium-term forecasting model according to a preferred embodiment of the present invention will be described in more detail.

15 is a view for explaining a forest fire risk medium-term forecast model according to a preferred embodiment of the present invention.

Referring to FIG. 15 , in the present invention, the weather forecast data of the forest fire risk index, drought index, and GDAPS described above were applied to machine learning to generate a forest fire risk medium-term forecast model. Referring to [Table 5] below, the input variables used in the forest fire risk medium-term forecasting model are the forest fire frequency index (that is, a map of the forest fire area), altitude, time series data of the 7-day forest fire risk index, and the drought index (Scheme 2) , air temperature, precipitation, relative humidity, surface temperature, and wind speed calculated from the Global Data Assimilation and Prediction System (GDAPS) were used. After 1 day (acc1), 2 days later (acc2), 3 days later (acc3), 4 days later (acc4), 5 days later (acc5), 6 days later (acc6), 7 days later (acc7) according to the number of days of using the GDAPS forecast guarantee predicted. As GDAPS became available from July 1, 2016, the total study period used for development was from July 1, 2016 to December 31, 2018. In the present invention, a mid-term forest fire risk forecasting model was developed through three different machine learning methods: random forest (RF), support vector regression analysis (SVR), and deep neural network (DNN) (number of prediction days * machine learning) Type = 21 models in total).

Variable Abbreviations		variable name
input variable	before 7	Short-term forecast index value 7 days ago (forest fire risk index 7 days ago)
	before 6	Short-term forecast index value 6 days ago (forest fire risk index 6 days ago)
	before 5	Short-term forecast index value 5 days ago (forest fire risk index 5 days ago)
	before 4	Short-term forecast index value 4 days ago (forest fire risk index 4 days ago)
	before 3	Short-term forecast index value 3 days ago (3 days ago forest fire risk index)
	before 2	Short-term forecast index value 2 days ago (wild fire risk index 2 days ago)
	before 1	Short-term forecast index value one day ago (one day ago forest fire risk index)
	dem	SRTM DEM (altitude)
	fire	Wildfire frequency index (map of wildfire hot spots)
	air temp	GDAPS air temperature (average of 1 to 7 days)
	precipitation	GDAPS precipitation (cumulative from 1 to 7 days)
	RH	GDAPS Relative Humidity (average of 1 to 7 days)
	surface temperature	GDAPS surface temperature (1 to 7 day average)
	wind	GDAPS wind speed (average from 1 to 7 days)
dependent variable	target	Medium-term forecast index value after 1 to 7 days

16 is a view for explaining the accuracy comparison results of the forest fire risk medium-term forecasting model according to a preferred embodiment of the present invention.

The forest fire risk medium-term forecasting model was developed by dividing it into an offline model and a real-time model. The offline model performs forest fire risk forecasting using a fixed model according to each forecast period, and daily data from July 1, 2016 to April 30, 2017 were used as training data. For RF, 500 trees were used, and for SVR, a Gaussian kernel was used. The DNN consisted of three hidden layers, each with 5, 4, and 3 neurons through several tests. And, the training function used Levenberg-Marquardt and the performance function used cross-entropy. 16 shows R, RMSE, Slope, and Bias for each forecast period of the offline model, "acc + number" means a predictive model as many as "number", for example, "acc1" is a forecast model one day later. , "acc7" means the 7-day forecast model. All models showed significant results with R above 0.9, RSME below 8%, slope above 0.8, and bias below 5. Among the three models, RF calibration and validation results showed high R and low RMSE, a slope close to 1, and low bias. All three models showed a tendency to decrease in accuracy as the forecast period increased, and among the three models, the RF error increased the lowest. Accordingly, the forest fire risk medium-term forecast model according to the present invention was generated using RF.

17 is a diagram for explaining the importance of input variables of a mid-term forest fire risk forecasting model generated based on a random forest according to a preferred embodiment of the present invention.

As shown in FIG. 17 , looking at the importance of input variables of the RF-based forest fire risk medium-term forecasting model, precipitation, atmospheric temperature, relative humidity, and surface temperature of GDAPS were found to be important variables, and precipitation was the most important among them. In GDAPS, the importance of precipitation tends to decrease as the forecast period becomes longer. is judged Therefore, if the accuracy of GDAPS is increased, the performance of the medium-term forest fire risk forecasting model will also be improved. In the FRI time series variables (before 1 to 7), the effect of the previous day was found to be the greatest.

18 is a view for explaining the results of a forest fire risk medium-term forecast model generated based on a random forest according to a preferred embodiment of the present invention.

18 is a comparison of the map calculated using the RF-based forest fire risk medium-term forecast model (4-day forecast model) when the risk of forest fire on May 6, 2017 was calculated by FRI, and compared with the reference, FRI had an R value of 0.66 and an RMSE of 24.45%, and RF showed an R of 0.85 and an RMSE of 4.86%, so RF showed better results. In terms of spatial distribution, FRI showed an overall inconsistent tendency, but the range of values was generally consistent in the RF model in Gangwon-do, Chungcheong-do, and Gyeongsangbuk-do.

19 is a view for explaining the accuracy of the forest fire risk medium-term forecast model generated in real time according to a preferred embodiment of the present invention.

An attempt was made to develop a real-time model to correct the output of GDAPS in real time according to the version improvement of GDAPS. All three machine learning techniques introduced above were tried, but in the case of SVR and DNN, it was judged to be inappropriate because it takes a lot of time to apply to real systems. Therefore, a model was developed by training the past 30 days using RF and forecasting 1 to 7 days later. It took about 3 minutes to train the data and produce the forest fire risk map, and as shown in FIG. 19 , it exhibited a relatively low RMSE compared to the offline model. Therefore, it was determined that the real-time model would be more efficient if the GDAPS model was continuously updated.

Then, with reference to FIG. 20, a medium-term forest fire risk forecasting method according to a preferred embodiment of the present invention will be described.

20 is a flowchart for explaining a medium-term forest fire risk forecasting method according to a preferred embodiment of the present invention.

Referring to FIG. 20 , the forest fire risk medium-term forecasting apparatus 100 may obtain a soil moisture index downscaled to a grid size of 1 km based on a machine learning algorithm ( S110 ).

That is, the forest fire risk medium-term forecasting device 100 is based on a machine learning algorithm, with TRMM precipitation data, ASCAT soil moisture data, NDVI, LST, and DEM as input variables, and soil moisture with GLDAS soil moisture data as output variables. Create an exponential downscaling model. Here, the machine learning algorithm may use one of random forest (RF), support vector regression (SVR), and artificial neural network (ANN). In particular, the present invention can generate a soil moisture index downscaling model using a random forest (RF).

At this time, the forest fire risk medium-term forecasting device 100 uses, as training data, the upscaling data to a grid size of 25 km, which is the same as the grid size of the GLDAS soil moisture data, as training data for the unity of spatial resolution between data, and the soil moisture index downscaling model create Then, the forest fire risk medium-term forecasting apparatus 100 obtains the soil moisture index downscaled to the grid size of 1 km by inputting the input variable converted to the grid size of 1 km into the soil moisture index downscaling model.

In more detail, the forest fire risk medium-term forecasting device 100 uses the data upscaled to a grid size of 25 km as training data, TRMM precipitation data, ASCAT soil moisture data, NDVI, LST and DEM as input variables, Using the GLDAS soil moisture data as an output variable, a first soil moisture index downscaling model may be generated. And, the forest fire risk medium-term forecasting device 100 uses the data upscaled to a grid size of 25 km as training data, TRMM precipitation data, NDVI, LST, and DEM as input variables, and GLDAS soil moisture data as output variables. , it is possible to generate a second soil moisture index downscaling model. Then, the forest fire risk medium-term forecasting device 100 inputs the input variable converted to the grid size of 1 km into the first soil moisture index downscaling model to obtain the first soil moisture index downscaled to the grid size of 1 km, The input variable converted to a grid size of 1 km is input to the second soil moisture index downscaling model to obtain a second soil moisture index downscaled to a grid size of 1 km, and the missing part in the first soil moisture index is the second By substituting the soil moisture index, the final soil moisture index downscaled to a grid size of 1 km can be obtained.

Then, the forest fire risk medium-term forecasting apparatus 100 may obtain a drought index by using the soil moisture index, NDWI, and TCI downscaled to a grid size of 1 km (S130).

That is, the forest fire risk medium-term forecasting apparatus 100 may obtain the drought index through the equation 0.4 * (downscaled soil moisture index) + 0.3 * (NDWI) + 0.3 * (TCI).

Then, the forest fire risk medium-term forecasting apparatus 100 may obtain a forest fire risk index by using the wildfire frequent area map, the modified FFMC, the drought index, and the monthly weight (S150).

That is, the forest fire risk medium-term forecasting apparatus 100 may obtain the forest fire risk index through the equation (forest fire frequent area map + 0.5) * (modified FFMC) * (1.5 - drought index) * (monthly weight).

Here, the monthly weight may be given a greater weight as the number of monthly forest fires increases, using the ratio of the number of monthly wildfires obtained based on all forest fires that have occurred in the past.

And, the modified FFMC is the conventional FFMC that calculates the FFMC based on the equilibrium moisture content, drying rate and minimum moisture content of the day calculated using the relative humidity, precipitation measured at noon, temperature, and wind speed. say FFMC. That is, in the modified FFMC, the coefficient of the formula for calculating the FFMC initial value, the precipitation reference value, the FFMC reference value and the drying rate is modified in the conventional FFMC, the coefficient of the formula calculating the moisture content of the previous day is modified, and the minimum moisture content is used It refers to an FFMC in which the coefficient of the equation for calculating the FFMC value is modified, the accumulated precipitation per day is used instead of the precipitation measured at noon, and the precipitation reference value is larger than that of the conventional FFMC.

Then, the forest fire risk medium-term forecasting apparatus 100 generates, based on the machine learning algorithm, a forest fire risk medium-term forecast model for each medium-term forecast day (S170).

That is, the forest fire risk medium-term forecasting device 100 is based on a machine learning algorithm, and the weather forecast data obtained from the forest fire frequent area map, altitude, forest fire risk index, drought index, and GDAPS for the past 7 days as input variables, A forest fire risk medium-term forecast model using the medium-term forecast index value as an output variable is generated for each medium-term forecast date.

Here, the machine learning algorithm may use one of random forest (RF), support vector regression (SVR), and deep neural network (DNN). In particular, the present invention can generate a forest fire risk medium-term forecast model using a random forest (RF).

And, the weather forecast data obtained from GDAPS includes air temperature, precipitation, relative humidity, surface temperature and wind speed, and the precipitation is an accumulated value for the past 7 days, and the remainder (air temperature, relative humidity, surface temperature, wind speed) ) is the average value for the past 7 days.

In addition, the medium-term forecast date refers to the weekly forecast such as 1 day later, 2 days later, 3 days later, 4 days later, 5 days later, 6 days later, and 7 days later.

In this case, the forest fire risk medium-term forecasting apparatus 100 may generate a forest fire risk medium-term forecasting model for each medium-term forecast day by using daily data of a preset past period as training data. On the other hand, the forest fire risk medium-term forecasting apparatus 100 uses daily data of input variables for a preset past period with the forecast date as the reference date as training data to generate a forest fire risk medium-term forecast model for each medium-term forecast date in real time. may be

Thereafter, the forest fire risk medium-term forecasting apparatus 100 obtains a medium-term forecast index value for each medium-term forecast date for the forest fire risk with the forecast date as the reference date based on a plurality of forest fire risk medium-term forecast models generated for each medium forecast date (S190) ).

At this time, when the forest fire risk medium-term forecast model is generated in real time, the forest fire risk medium-term forecasting device 100 is based on a plurality of forest fire risk medium-term forecast models generated in real time for each medium-term forecast date, the forest fire risk with the forecast date as the reference date It is possible to obtain the medium-term forecast index value for each medium-term forecast date.

Even though all the components constituting the embodiment of the present invention described above are described as being combined or operated in combination, the present invention is not necessarily limited to this embodiment. That is, within the scope of the object of the present invention, all the components may operate by selectively combining one or more. In addition, all of the components may be implemented as one independent hardware, but a part or all of each component is selectively combined to perform some or all of the functions of the combined hardware in one or a plurality of hardware program modules It may be implemented as a computer program having In addition, such a computer program is stored in a computer readable media such as a USB memory, a CD disk, a flash memory, etc., read and executed by a computer, thereby implementing an embodiment of the present invention. The recording medium of the computer program may include a magnetic recording medium, an optical recording medium, and the like.

The above description is merely illustrative of the technical idea of the present invention, and various modifications, changes, and substitutions are possible within the range that does not depart from the essential characteristics of the present invention by those of ordinary skill in the art to which the present invention pertains. will be. Accordingly, the embodiments disclosed in the present invention and the accompanying drawings are for explaining, not limiting, the technical spirit of the present invention, and the scope of the technical spirit of the present invention is not limited by these embodiments and the accompanying drawings . The protection scope of the present invention should be construed by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

< Explanation of symbols >

100: forest fire danger medium-term forecasting device,

110: soil moisture index fraction,

130: drought index acquisition unit,

150: forest fire risk index acquisition unit,

170: model generation unit;

190: forecast department

Claims (13)

1. Using daily data of a preset past period as training data, based on a machine learning algorithm, it is based on a map of wildfire hot spots, altitude, forest fire risk index for the past 7 days, drought index and GDAPS (Global Data Assimilation and Prediction System) from a model generator for generating a forest fire risk medium-term forecasting model for each medium-term forecast day, using the obtained weather forecast data as an input variable and a medium-term forecast index value as an output variable; and

   a forecasting unit configured to obtain a medium-term forecast index value for each medium-term forecast day for the forest fire risk based on the forecast date as a reference date, based on the plurality of medium-term forecast models of the forest fire risk generated for each medium-term forecast date;

   includes,

   The weather forecast data obtained from the GDAPS includes atmospheric temperature, precipitation, relative humidity, surface temperature and wind speed, wherein the precipitation is a cumulative value for the past 7 days, and the remainder is an average value for the past 7 days,

   Forest fire risk medium-term forecasting device.
2. In claim 1,

   The model generation unit,

   Using the daily data of the input variable for a preset past period with the forecast date as the reference date as training data, generating the forest fire risk medium-term forecast model for each medium-term forecast date in real time,

   The forecasting unit,

   Based on the plurality of forest fire risk medium-term forecast models generated in real time for each medium-term forecast date, obtaining a medium-term forecast index value for each medium-term forecast date for the forest fire risk based on the forecast date as a reference date,

   Forest fire risk medium-term forecasting device.
3. In claim 1,

   a forest fire risk index acquisition unit configured to acquire the forest fire risk index using the forest fire frequent area map, the modified Fine Fuel Moisture Code (FFMC), the drought index and monthly weights;

   further comprising,

   The monthly weight, using the ratio of the number of monthly wildfires obtained based on all forest fires that have occurred in the past, is given a greater weight as the number of monthly forest fires increases,

   Forest fire risk medium-term forecasting device.
4. In claim 3,

   The forest fire risk index acquisition unit,

   Obtaining the wildfire risk index through the formula (the wildfire hot area map + 0.5) * (the modified FFMC) * (1.5 - the drought index) * (the monthly weight),

   Forest fire risk medium-term forecasting device.
5. In claim 3,

   The modified FFMC is,

   In the conventional FFMC that calculates the FFMC based on the equilibrium moisture content, drying rate, and minimum moisture content for the day calculated using relative humidity, precipitation measured at noon, temperature, and wind speed, FFMC initial value, precipitation reference value, FFMC reference value and dry The coefficient of the equation for calculating the tuning is corrected, the coefficient of the equation for calculating the moisture content of the previous day is corrected, the coefficient of the equation for calculating the FFMC value using the minimum moisture content is corrected, and the cumulative precipitation per day, not the precipitation measured at noon is used, the precipitation reference value is larger than that of the conventional FFMC,

   Forest fire risk medium-term forecasting device.
6. In claim 3,

   a drought index acquisition unit configured to acquire the drought index using a soil moisture index downscaled to a grid size of 1 km, a Normalized Different Water Index (NDWI), and a Temperature Condition Index (TCI);

   Forest fire risk medium-term forecasting device further comprising.
7. In claim 6,

   The drought index acquisition unit,

   Obtaining the drought index through the formula 0.4 * (the downscaled soil moisture index) + 0.3 * (the NDWI) + 0.3 * (the TCI),

   Forest fire risk medium-term forecasting device.
8. In claim 6,

   Using the data upscaled to a grid size of 25 km as training data, based on a machine learning algorithm, TRMM (Tropical Rainfall Measuring Mission) precipitation data, ASCAT (Advanced SCATterometter) soil moisture data, NDVI (Normalized Difference Vegetation Index), A Soil Moisture Index downscaling model is created using Land Surface Temperature (LST) and Digital Elevation Model (DEM) as input variables and Global Land Data Assimilation System (GLDAS) soil moisture data as output variables, and a grid size of 1 km a soil moisture index acquisition unit for obtaining the downscaled soil moisture index by inputting the input variable converted to , into the soil moisture index downscaling model;

   Forest fire risk medium-term forecasting device further comprising.
9. In claim 8,

   The soil moisture index acquisition unit,

   Using the data upscaled to a grid size of 25 km as training data, the TRMM precipitation data, the ASCAT soil moisture data, the NDVI, the LST, and the DEM are input variables, and the GLDAS soil moisture data are output variables. to create a first soil moisture index downscaling model,

   Using the data upscaled to a grid size of 25 km as training data, the TRMM precipitation data, the NDVI, the LST, and the DEM are input variables, and the GLDAS soil moisture data are output variables, second soil moisture create an exponential downscaling model,

   The input variable converted to a grid size of 1 km is input to the first soil moisture index downscaling model to obtain a first soil moisture index downscaled to a grid size of 1 km, and the input variable converted to a grid size of 1 km is the input variable. A second soil moisture index downscaled to a grid size of 1 km is obtained by inputting the second soil moisture index downscaling model, and the missing part in the first soil moisture index is replaced with the second soil moisture index to perform the downscaling to obtain the soil moisture index,

   Forest fire risk medium-term forecasting device.
10. A medium-term forest fire risk forecasting method performed by a forest fire risk medium-term forecasting device, comprising:

    Using daily data of a preset past period as training data, based on a machine learning algorithm, it is based on a map of wildfire hot spots, altitude, forest fire risk index for the past 7 days, drought index and GDAPS (Global Data Assimilation and Prediction System) from generating a forest fire risk medium-term forecast model for each medium-term forecast day using the obtained weather forecast data as an input variable and a medium-term forecast index value as an output variable; and

    obtaining a medium-term forecast index value for each medium-term forecasting date for a forest fire risk based on a forecast date as a reference date based on the plurality of medium-term forecasting models for forest fire risk generated for each medium-term forecast date;

    includes,

    The weather forecast data obtained from the GDAPS includes atmospheric temperature, precipitation, relative humidity, surface temperature and wind speed, wherein the precipitation is a cumulative value for the past 7 days, and the remainder is an average value for the past 7 days,

    How to forecast forest fire risk in the medium term.
11. In claim 10,

    The forest fire risk medium-term forecast model creation step is,

    Using the daily data of the input variable for a preset past period with the forecast date as the reference date as training data, generating the forest fire risk medium-term forecast model for each medium-term forecast date in real time,

    The medium-term forecast index value acquisition step is,

    Based on the plurality of forest fire risk medium-term forecast models generated in real time for each medium-term forecast date, it consists of obtaining a medium-term forecast index value for each medium-term forecast date for the forest fire risk with the forecast date as a reference date,

    How to forecast forest fire risk in the medium term.
12. In claim 10,

    obtaining the forest fire risk index using the wildfire frequent area map, a modified Fine Fuel Moisture Code (FFMC), the drought index, and monthly weights;

    further comprising,

    The monthly weight, using the ratio of the number of monthly wildfires obtained based on all forest fires that have occurred in the past, is given a greater weight as the number of monthly forest fires increases,

    How to forecast forest fire risk in the medium term.
13. A computer program stored in a computer-readable recording medium in order to execute the medium-term forest fire risk forecasting method according to any one of claims 10 to 12 on a computer.

Images