Satellite-based high-resolution gross primary productivity calculation system, and calculation method using same

Abstract

The present invention relates to a gross primary productivity calculation system using deep learning, and a calculation method using same, wherein gross primary productivity can be estimated on the basis of Sentinel-2 reflectance and photosynthetically active radiation (PAR). In particular, by applying a deep learning model that does not require other weather variables and uses only "Nadir BRDF-adjusted reflectance (NBAR)" of Sentinel-2 and the PAR as input data, accuracy similar to or higher than that of existing models can be achieved, even though the gross primary productivity is derived using deep learning with significantly simplified input data.

Description

Satellite-based high-resolution gross primary production calculation system and calculation method using the same

[0001] The present invention relates to a system for calculating total primary production using deep learning and a calculation method using the same.

[0002] Gross Primary Productivity (GPP) refers to the total amount of organic matter produced through photosynthesis in an ecosystem. It measures the total energy produced by photosynthetic organisms such as plants, algae, and bacteria when they use sunlight to create sugars from carbon dioxide and water.

[0003] GPP is used as one of the important indicators for evaluating the productivity of an ecosystem, and it includes several characteristics:

[0004] GPP reflects total photosynthetic activity, indicating the total amount of photosynthesis occurring in plants' leaves, stems, and roots. It also includes net primary productivity (NPP), which is the net difference between GPP and the energy consumed for respiration. NPP can provide information on the amount of energy remaining in an ecosystem and available for other organisms to utilize. Furthermore, GPP is used as an important indicator of ecosystem health and function, with a high GPP typically indicating a vibrant and healthy ecosystem.

[0005] For example, tropical rainforests have very high GPP, meaning that plants in those areas are actively photosynthesizing, generating large amounts of organic matter. This may indicate limited photosynthetic activity.

[0006] GPP is an important concept used in various fields such as environmental science, ecology, and climate change studies, and it helps researchers understand and manage important issues such as energy flow in ecosystems, carbon cycle, and biodiversity conservation.

[0007] In this way, information on Gross Primary Productivity (GPP) is used as very important information, and recently, satellite-based Earth observation systems have been useful in deriving Gross Primary Productivity (GPP) for the entire globe.

[0008] In particular, the Sentinel-2 data, freely provided by the European Space Agency (ESA), is widely used. However, existing methodologies for estimating gross photosynthetic potential (GPP) rely on mechanistic models or semi-empirical light use efficiency (LUE) models. These models require a variety of meteorological variables, require diverse empirical data, and vary in performance across ecosystem types. There is a growing need for technologies that overcome these limitations and enable rapid and easy estimation of gross primary production (GPP) in ecosystems.

[0009] The present invention has been made to solve the above-described needs, and the purpose of the present invention is to provide a system that can estimate total primary production based on Sentinel-2 reflectance and photosynthetically active radiation, and in particular, can provide a system that can achieve accuracy similar to or higher than existing models despite deriving total primary production using deep learning and much simpler input data by applying a deep learning model that uses only the 'Nadir BRDF-Adjusted Reflectance (NBAR)' and photosynthetically active radiation (PAR) of Sentinel-2 as input data without requiring other meteorological variables.

[0010] As a means for solving the above-described problem, an embodiment of the present invention provides a satellite-based high-resolution gross primary production (GPP) estimation system, including a data collection unit (110) that collects Sentinel-2 reflectance data and photosynthetically active radiation (PAR) data stored in an external server, as illustrated in FIGS. 1 to 5; a data preprocessing unit (120) that preprocesses the data collected by the data collection unit (110) and calculates an average NABR (Nadir BRDF-Adjusted Reflectance) using the collected data; and a data learning unit (130) that has a GPP calculation model that calculates GPP through deep learning using the average NABR and photosynthetically active radiation (PAR) calculated by the data preprocessing unit (120) as input values.

[0011] In particular, in another embodiment of the present invention, a satellite-based high-resolution gross primary production calculation method is provided, including: a first step of collecting data provided from a database of a server having satellite Sentinel-2 NBAR data, EuroFlux and AmeriFlux in-situ photosynthetically active radiance (PAR) data by using the satellite-based high-resolution gross primary production calculation system described above in a data collection unit (110); a second step of preprocessing the collected data in a data preprocessing unit (120) and calculating an average NABR (Nadir BRDF-Adjusted Reflectance) using the collected data; and a third step of calculating gross primary production (GPP) through deep learning using the calculated average NABR and photosynthetically active radiance (PAR) as input values ​​in a data learning unit (130).

[0012] Furthermore, in an embodiment of the present invention, a computer-readable recording medium having recorded thereon a program for performing the above-described satellite-based high-resolution total primary production calculation method is provided.

[0013] According to an embodiment of the present invention, the total primary production can be estimated based on Sentinel-2 reflectance and photosynthetically active radiation, and in particular, a deep learning model that does not require other meteorological variables and uses only the 'Nadir BRDF-Adjusted Reflectance (NBAR)' and Photosynthetically Active Radiation (PAR) of Sentinel-2 as input data is applied, thereby deriving the total primary production using deep learning and much simplified input data, and yet has the effect of implementing an accuracy similar to or higher than that of existing models.

[0014] The present invention is a technology for estimating gross primary production based on Sentinel-2 reflectance and photosynthetically active radiation. It can be applied to a wide range of fields, including science, engineering, and technology, such as ecology, computational vision, remote sensing, deep learning, and machine learning. It can also be applied to technologies that estimate not only carbon dioxide but also methane absorption in ecosystems.

[0015] The present invention can be used to manage and monitor the gross primary production (GPP) of ecosystems, which is difficult to access. It can also be utilized in the development of software for local governments to calculate ecosystem carbon credits. Furthermore, the present invention is a technology that quickly and easily estimates GPP generated from ecosystems. It is deeply connected to ecosystem services, the growing satellite utilization market, and the carbon credit market, and can be used as a basis for land use planning and decision-making.

[0016] Furthermore, the present invention can be extended to a technology for estimating not only carbon dioxide but also methane absorbed in an ecosystem.

[0017] FIG. 1 is a block diagram illustrating the configuration of a satellite-based high-resolution total primary production calculation system (hereinafter referred to as “the present invention”) according to one embodiment of the present invention.

[0018] Figure 2 lists examples of data provision sites that can be utilized by the data collection unit of the present invention.

[0019] FIG. 3 illustrates a preprocessing process when the data collection path of the present invention is Sentinel-2, and FIG. 4 illustrates an eddy covariance result applied to the footprint area used in FIG. 3.

[0020] Figure 5 illustrates the structure of the GPP production model (GPP-net architecture) of the present invention.

[0021] Figure 6 is a flowchart illustrating the application process of the present invention.

[0022] Figure 7 is a conceptual diagram for explaining the process of collecting data from EuroFlux tower data and calculating NBAR in Figure 6.

[0023] Figure 8 is a conceptual diagram for explaining the process of collecting AmeriFlux tower data and calculating NBAR in Figure 6.

[0024] Figure 9 is a conceptual diagram for explaining the 'Five-fold cross validation' method used to evaluate the model performance of the GPP production model (GPP-net) described above in Figure 6.

[0025] Figure 10 is a graph comparing values ​​calculated using the GPP calculation model (GPP_net) according to the present invention and the existing GPP calculation method.

[0026] [Explanation of symbols]

[0027] 100: GPP production module

[0028] 110: Data Collection Department

[0029] 120: Data preprocessing unit

[0030] 130: Data Learning Department

[0031] The present invention is susceptible to various modifications and embodiments. Specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, but rather to encompass all modifications, equivalents, and alternatives falling within the spirit and technical scope of the present invention.

[0032] However, when describing embodiments, if a detailed description of a related known function or configuration is judged to unnecessarily obscure the gist of the present invention, a detailed description thereof will be omitted. Furthermore, the sizes of each component in the drawings may be exaggerated for illustrative purposes and do not necessarily represent the sizes actually applied.

[0033] Additionally, throughout the specification, when a component is referred to as being "connected" or "connected" to another component, it should be understood that the component may be directly connected or connected to the other component, but may also be connected or connected via another component in between, unless otherwise specifically stated. Additionally, throughout the specification, when it is said that a part "includes" a component, this does not exclude other components, but rather includes other components, unless otherwise specifically stated.

[0034] FIG. 1 is a block diagram illustrating the configuration of a satellite-based high-resolution total primary production calculation system (hereinafter referred to as “the present invention”) according to one embodiment of the present invention.

[0035] Referring to FIG. 1, the present invention may be configured to include a data collection unit (110) that collects Sentinel-2 reflectance data and photosynthetically active radiation (PAR) data stored in an external server, a data preprocessing unit (120) that preprocesses the data collected by the data collection unit (110) and calculates an average Nadir BRDF-Adjusted Reflectance (NBAR) using the collected data, and a data learning unit (130) that has a GPP calculation model that calculates gross primary production (GPP) through deep learning using the average NBAR and photosynthetically active radiation (PAR) calculated by the data preprocessing unit (120) as input values.

[0036] NBAR (Nadir BRDF-Adjusted Reflectance) used in the present invention refers to adjusted reflectance, a method for accurately evaluating the reflectance characteristics of the Earth's surface by standardizing the reflectance observed from satellites. In this case, BRDF (Bidirectional Reflectance Distribution Function) means that the reflectance characteristics at a specific point are measured from multiple angles.

[0037] Since satellite images are taken from various angles, the reflectivity of the same point may vary depending on the shooting angle. NBAR can compensate for these angular differences to provide consistent reflectivity values, and when comparing and analyzing satellite images taken from multiple points of view, it can provide standardized data that minimizes reflectivity changes due to angular differences and enables accurate comparison.

[0038] Additionally, the photosynthetically active radiation (PAR) in the present invention refers to the range of light used for photosynthesis in water. In particular, PAR encompasses wavelengths between 400 and 700 nanometers (nm) within the solar spectrum, and light in this range is absorbed by chlorophyll and other photosynthetic pigments in plants, thereby promoting the photosynthetic process.

[0039] In addition, the database (10 to 30) of the external server in FIG. 1 of the present invention preferably means a database of a server having photosynthetically active radiation (PAR) data of artificial satellite Sentinel-2 NBAR data, EuroFlux and AmeriFlux in-situ, and performs a function of collecting data provided therein.

[0040] EuroFlux is a network that measures carbon, water, and energy fluxes with high precision across diverse ecosystems in Europe. It conducts field observations to study climate change, ecosystem productivity, and the carbon cycle. EuroFlux collects data through towers and equipment installed in various locations across Europe, and these data are used for ecosystem and climate modeling, environmental policymaking, and more. The EuroFlux network is structured around towers across Europe, representing diverse ecosystems. Towers are typically 10-30 m high, with some towers or ground-based sensors incorporated into the network. The collected data is stored in the EuroFlux database and made available to researchers and the public. The data stored and provided in this database are provided in a standardized format, making them available for various research and analysis. The present invention makes this data available.

[0041] AmeriFlux in-situ is a network of in-situ observations designed to measure carbon dioxide, water, and energy fluxes with high precision in diverse ecosystems across the North American continent. AmeriFlux is a large-scale research project that began in 1996 to study climate change, ecosystem productivity, and the carbon cycle, and includes diverse ecosystems in the United States, Canada, and Mexico.

[0042] AmeriFlux in-situ observations are based on field observations that directly measure gas exchange between the atmosphere and the ground through towers installed at specific locations. These towers are primarily installed in various ecosystems such as forests, grasslands, wetlands, and agricultural lands.

[0043] The AmeriFlux network consists of a network of towers, over 100 of which are installed across North America. Each tower represents a specific ecosystem. Towers are typically 10 to 30 meters tall, with some even taller towers or ground-level sensors. The collected data is stored in the AmeriFlux database and made available to researchers and the public. The data is provided in a standardized format for use in various research and analyses. The present invention facilitates the collection and use of data provided in the AmeriFlux database.

[0044] Additionally, the Sentinel-2 satellite in the present invention is an Earth observation satellite operated by the European Space Agency (ESA). Two satellites (2A and 2B) observe the Earth and provide high-resolution multispectral images. In particular, the present invention utilizes data provided by the Copernicus program, an Earth observation program led by the European Union (EU) that supports various environmental and security applications, namely Sentinel-2 COPERNICUS / S2_SR_HARMONIZED data.

[0045] The Sentinel-2 COPERNICUS / S2_SR_HARMONIZED NBAR data is a product based on data collected by the Copernicus program's Sentinel-2 satellite. This data can be used in various applications, and is particularly important for standardizing and providing consistent ground surface reflectance characteristics. NBAR here stands for "Nadir BRDF-Adjusted Reflectance," and refers to a method for providing consistent ground surface reflectance characteristics regardless of angular differences.

[0046] The data collection unit (110) of the present invention collects data provided from the above-described artificial satellite Sentinel-2 NBAR data, EuroFlux and AmeriFlux in-situ, and in particular, the data may be configured to include photosynthetically active radiation (PAR) data.

[0047] In particular, in order to train and verify the gross primary production (GPP) calculation model that calculates GPP through deep learning described below, the present invention uses EuroFlux, AmeriFlux in-situ data, and Sentinel-2 NBAR data. EuroFlux and AmeriFlux in-situ data can be used as input photosynthetically active radiation (PAR) data, and GPP data can be used as training reference data. Sites that utilize such data may include the sites illustrated in Fig. 2.

[0048] Next, the data preprocessing unit (120) of the present invention performs a function of preprocessing data collected from the above-described data collection unit (110) and calculating an average NBAR (Nadir BRDF-Adjusted Reflectance) using the collected data.

[0049] The preprocessing process of data varies depending on the path of data being collected, and as a preferred example, in the case of satellite Sentinel-2 NBAR data, as shown in Fig. 3, a) Sentinel-2 COPERNICUS / S2_SR_HARMONIZED NBAR data is used, and the cloud percentage is set to 10%, b) all useful images are searched within the data period based on the image center coordinates on the site to match the location of the site data, and c) field observations are extracted according to the Sentinel-2 acquisition time to match the time of the field data and Sentinel-2 NBAR.

[0050] In addition, for the satellite Sentinel-2 NBAR data, for the data that performed the process of step c), the eddy covariance footprint is applied to calculate the average Sentinel-2 NBAR, but is applied using 80% of the cumulative footprint area, and the average Sentinel-2 NBAR can be calculated using {Formula 1} below.

[0051] {Formula 1}

[0052]

[0053] Specifically, to calculate the average Sentinel-2 NBAR, the 80% cumulative footprint area calculated from the eddy covariance results shown in Fig. 4 is used, and the weighting factors are divided into 0-50%, 50-60%, 60-70%, and 70-80%. Then, the Average Sentinel-2 NBAR is calculated using {Equation 1}. CL represents a contour line. For example, CL50% means cumulative 50%.

[0054]

[0055] Alternatively, if the data is collected along a path provided by a EuroFlux tower, preprocessing can be performed to calculate a footprint based on the data provided, thereby estimating the average NBAR. In this case, the footprint represents the area of ​​the Earth's surface affected by the exchange between the atmosphere and the ground measured by the flux tower, and spatially defines the region contributing to the measured flux. This footprint allows researchers to analyze how the topography, vegetation, and land use surrounding the tower influence the flux measurements.

[0056] Specifically, a footprint refers to the spatial range that indicates how much the gas or energy flux (e.g., carbon dioxide, water vapor, heat, etc.) measured at the location of a flux tower is influenced by the surrounding area. In other words, it indicates to what extent the flux data measured by a EuroFlux tower actually originates from the surrounding environment. The footprint is affected by various environmental factors such as wind speed and direction, terrain, and vegetation conditions. Therefore, in the present invention, "calculating a footprint" can be defined as a process of determining from which area the data measured by a flux tower at a specific time mainly originated, reflecting how much a specific area around the tower contributed to the flux data and enabling more accurate interpretation of this data.

[0057] Such calculation of the footprint can be performed using various known program models, such as the Kormann and Meixner model, the Hsieh model, and the Kljun model. In an embodiment of the present invention, the footprint can be calculated using the Hsieh model using Python.

[0058] In the case of the AmeriFlux tower, the data collection path does not allow for the calculation of a footprint based on the provided data. This is because the data collected from the AmeriFlux tower does not, on its own, provide sufficient information for calculating a footprint. This is because, in order to calculate a footprint, various additional information such as atmospheric turbulence, wind direction and speed, temperature, humidity, and topography and vegetation characteristics around the tower are required, but the dataset provided by the AmeriFlux tower does not include such additional information or is insufficient. Therefore, in the present invention, in the case of the AmeriFlux tower, the data collection path is preprocessed so that a 3x3 grid average NBAR can be calculated centered on the tower.

[0059] Next, the present invention can be configured to further include a data learning unit (130) having a GPP calculation model that calculates gross primary production (GPP) through deep learning using the average NBAR and photosynthetically active radiation (PAR) calculated in the data preprocessing unit (120) as input values.

[0060] In this case, as illustrated in FIG. 5, the GPP production model of the data learning unit (130) is trained and verified using photosynthetically active radiation (PAR) data as input and GPP data as learning reference data, and may be configured as a full 1-D convolutional encoder-decoder network consisting of 11 1-D convolutional layers and 1 fully connected layer.

[0061] The above GPP production model (GPP-net architecture) can be implemented with the structure shown in Fig. 5.

[0062] Referring to FIG. 5, the GPP output model (GPP-net architecture) may be configured as a full 1-D convolutional encoder-decoder network consisting of one 1-D convolutional layer and one fully connected layer.

[0063] In this case, 'Conv1d' represents a 1-D convolutional layer, and BatchNorm1d represents a 1-D batch normalization operation. Additionally, the LeakyReLU activation function is utilized. Up2x is an upsampling function. FC is a fully connected layer.

[0064] In particular, the 'green line' in Fig. 5 represents a skip connection architecture that can transfer features from superficial layers to deep layers for robust feature representation.

[0065] In the GPP production model (GPP-net architecture) of the present invention configured in this way, BAR*PAR is input to the network, and the network can output GPP in 30-minute units.

[0066] That is, EuroFlux, AmeriFlux in-situ data, and Sentinel-2 NBAR data are used to train and validate the model proposed in the present invention. The EuroFlux and AmeriFlux in-situ data can be used as input for photosynthetically active radiation (PAR) data, and GPP data can be used as training reference data. Figure 6 is a block diagram that summarizes the entire process of this embodiment of the present invention.

[0067] That is, as illustrated in FIG. 6, the present invention collects EuroFlux data, AmeriFLux data, and Sentinel-2 Nadir BRDF Adjusted Reflectance (NBAR) data from the data collection unit (110).

[0068] Afterwards, for each collected data, in the case of the EuroFlux tower, the footprint can be calculated based on the provided data, and the average NBAR is calculated based on this. In the case of the AmeriFlux tower, the footprint cannot be calculated based on the provided data. In this case, the average NBAR of a 3x3 grid centered on the tower is calculated. Furthermore, to calculate the average Sentinel-2 NBAR, the Average Sentinel-2 NBAR is calculated using {Equation 1} described above. The calculated result values ​​go through a process of matching on a map (Georeferencing). This forms a dataset by matching the flux tower data and Sentinel-2 NBAR data.

[0069] Furthermore, by applying the dataset, temporally and spatially consistent field PAR, GPP, and Sentinel-2 NBAR can be used to train and validate the GPP production model (GPP-net). For model validation, a 5-fold cross-validation strategy is used. The entire dataset is randomly split into five equal parts. The training and validation processes will be repeated five times, with a different validation set at each stage.

[0070] Figure 7 is a conceptual diagram for explaining the process of collecting data from the EuroFlux tower data and calculating NBAR in Figure 6 described above.

[0071] Referring to Figure 7, for example, the EuroFlux data collection allows for the collection of tower flux data at 33 sites covering seven vegetation functional types: farmland, wetland, grassland, evergreen coniferous forest, mixed forest, evergreen broadleaf forest, and deciduous broadleaf forest.

[0072] Using these tower flux data, we can determine the presence and location of a footprint. Then, based on geographic information, we match the flux tower data with the Sentinel-2 NBAR footprint.

[0073] The average NBAR can be calculated using footprint weights. For weight calculations, 80% of the cumulative footprint area is considered, and the weighting factors are divided into 0-50%, 50-60%, 60-70%, and 70-80%.

[0074] Figure 8 shows an example of collecting AmeriFlux data, covering 31 sites and 6 vegetation functional types, namely farmland, wetland, grassland, evergreen coniferous forest, mixed forest, and deciduous broadleaf forest.

[0075] For the AmeriFlux data in question, as described above, the data is in a form that lacks information to calculate the footprint, so the average NBAR is calculated within a 3x3 grid.

[0076] Figure 9 is a conceptual diagram illustrating the "Five-fold cross validation" method used to evaluate the model performance of the GPP production model (GPP-net) described in Figure 6. Specifically, in the present invention, a five-fold cross validation method was used to evaluate the performance of the model. This method divided the dataset into five subsets, four of which were used for training, and the remaining one for validation. This process was repeated five times, and the results of these five validations were combined to calculate an accuracy index.

[0077]

[0078] The present invention (a high-resolution gross primary production estimation technology based on the Sentinel-2 satellite that is independent of the ecosystem type) is an algorithm that uses only the 'Nadir BRDF-Adjusted Reflectance (NBAR)' and Photosynthetically Active Radiation (PAR) of Sentinel-2 as input data without requiring other meteorological variables. In order to regress the relationship between NBAR* PAR and GPP, a network (CNN) with a one-dimensional convolutional encoder and decoder was configured.

[0079]

[0080] Traditional methods for estimating GPP require a significant amount of data. Furthermore, because these methods rely on complex models to estimate GPP, the range of error varies widely depending on the ecosystem type. Understanding these models requires a deep understanding of ecosystem carbon exchange, significantly limiting their usability.

[0081] The present invention develops a technology that uses deep learning and significantly simplified input data to derive gross primary production while demonstrating accuracy similar to or higher than existing models, thereby making it available to a wide range of users.

[0082] The method for estimating total primary production using the system according to the present invention can proceed in the following order.

[0083] First, the first step of collecting data provided from the database of a server having photosynthetically active radiation (PAR) data from the satellite Sentinel-2 NBAR data, EuroFlux and AmeriFlux in-situ can be performed in the data collection unit (110).

[0084] Next, in the data preprocessing unit (120), the collected data is preprocessed so that the second step of calculating the average NBAR (Nadir BRDF-Adjusted Reflectance) using the collected data can be performed.

[0085] And, in the data learning unit (130), the third step of calculating gross primary production (GPP) through deep learning using the calculated average NBAR and photosynthetically active radiation (PAR) as input values ​​can be performed.

[0086] In particular, in this case, the above 2 steps are, in the case of satellite Sentinel-2 NBAR data,

[0087] a) Using Sentinel-2 COPERNICUS / S2_SR_HARMONIZED NBAR data, the cloud percentage is set to 10%,

[0088] b) Match the location of the site data by retrieving all useful images within the data period based on the image center coordinates on the site, and

[0089] c) To synchronize the time of field data and Sentinel-2 NBAR, a process of extracting field observations according to the Sentinel-2 acquisition time is performed.

[0090] d) For data that has performed the process of step c), the method may be configured to include a step of applying an eddy covariance footprint to calculate an average Sentinel-2 NBAR using 80% of the cumulative footprint area, and calculating an average Sentinel-2 NBAR using the above-described {Formula 1}.

[0091] In addition, the configuration of the above two steps may further include a step of calculating an average NBAR by calculating a footprint based on the provided data in the case of data provided from a EuroFlux tower, and calculating an average NBAR of a 3×3 grid centered on the tower in the case of data provided from an AmeriFlux tower.

[0092]

[0093] Figure 10 is a graph comparing values ​​calculated using the GPP calculation model (GPP_net) according to the present invention and the existing GPP calculation method.

[0094] The data applied were collected from EuroFlux and AmeriFlux data, and are data including 31 sites (sites described in FIGS. 7 and 8) and 6 plant functional types applied in the above-described embodiment of the present invention. This means information on cropland (CRO), grassland (GRA), wetland (WET), evergreen needleleaf forest (ENF), mixed forest (MF), evergreen broadleaf forest (EBF), and deciduous broadleaf forest (DBF).

[0095] The results shown in Figure 10(a) illustrate the results of calculating GPP for each ecosystem type using NIRvP. NIRvP is an index designed to more accurately estimate plant photosynthesis and productivity. It is a variant of the Normalized Difference Vegetation Index (NDVI), and is designed to more accurately estimate plant photosynthesis and productivity.

[0096] In particular, NIRv is one of the photosynthetic vegetation indices, reflecting the physiological activity of plants through a combination of Near-Infrared Reflectance and NDVI, and NIRvP is used as an indicator of the photosynthetic capacity (potential) of plants based on NIRv.

[0097] In particular, NIRvP is a widely used indicator for estimating GPP, and is calculated by multiplying NIRv (NDVI * NIR reflectance) and PAR (solar irradiance).

[0098] The result shown in Fig. 10 (b) shows the result of calculating GPP by PLS (Partial Least Squares), which is one of the multiple regression analysis methods, and is used especially when analyzing the relationship between an independent variable (X) and a dependent variable (Y), and is a technique mainly used when using various reflectances in a way that maximizes the covariance between X and Y.

[0099] The result shown in Fig. 10 (c) shows the GPP produced by applying the GPP-Net of the present invention.

[0100] When comparing the three methods above, it can be confirmed that the RMSE (Root Mean Square Error) in each analysis technique is 5.23 for the RMSE calculated by applying the GPP-Net of the present invention, which is significantly lower than the RMSE of 7.44 for the GPP calculated by the existing NIRvP and 5.78 for the GPP calculated by PLS. This confirms that the accuracy of the value calculated by applying the GPP calculation model according to the present invention produces a result value with higher reliability.

[0101] Accordingly, in the case of the GPP calculation model according to the present invention, it is economical by using satellite data, which is free data, and it greatly simplifies the existing deep learning training data, so that GPP can be calculated by applying only the radiation amount and satellite reflectivity values, thereby increasing its efficiency.

[0102] In addition, in the case of a general GPP output model, learning is performed by segmenting the ecosystem type for output, whereas in the present invention, learning is performed using an input data set for the diversity of ecosystem types, which has the advantage of innovatively reducing deviation in output values.

[0103] As described above, the functional configuration and execution operations applicable to performing the satellite-based high-resolution gross primary yield calculation method of the present invention can be represented by functional block configurations and various processing steps. These functional blocks can be implemented by various numbers of hardware and / or software configurations that execute specific functions. For example, the present invention can employ direct circuit configurations such as memory, processing, logic, and look-up tables that can execute various functions under the control of one or more microprocessors or other control devices.

[0104] Just as the components of the present invention can be implemented as software programming or software elements, the present invention can be implemented in a programming or scripting language such as Python, C, C++, Java, assembler, etc., including various algorithms implemented as a combination of data structures, processes, routines, or other programming constructs. Functional aspects can be implemented as algorithms executed on one or more processors. Furthermore, the present invention can employ conventional techniques for electronic configuration, signal processing, and / or data processing. Terms such as “module,” “part,” “mechanism,” “element,” “means,” and “configuration” can be used broadly and are not limited to mechanical and physical components. The above terms can include the meaning of a series of software processes (routines) in connection with a processor, etc.

[0105] The detailed description of the present invention, as described above, has described specific embodiments. However, various modifications are possible without departing from the scope of the present invention. The technical spirit of the present invention should not be limited to the aforementioned embodiments, but should be defined not only by the claims but also by equivalents thereof.

Claims

1. In the satellite-based high-resolution gross primary production calculation system, It includes a data collection unit (110) that collects Sentinel-2 reflectance data and photosynthetically active radiance (PAR) data stored in an external server; a data preprocessing unit (120) that preprocesses the data collected by the data collection unit (110) and calculates an average NBAR (Nadir BRDF-Adjusted Reflectance) using the collected data; and a data learning unit (130) that has a GPP calculation model that calculates gross primary production (GPP) through a deep learning model composed of a fully convolutional encoder-decoder network, using the average NBAR and photosynthetically active radiance (PAR) calculated by the data preprocessing unit (120) as input values. The above data preprocessing unit (120) is, in the case of artificial satellite Sentinel-2 NBAR data, a) Using Sentinel-2 COPERNICUS / S2_SR_HARMONIZED NBAR data, the cloud percentage is set to 10%, b) Match the location of the site data by retrieving all useful images within the data period based on the image center coordinates on the site, and c) To synchronize the time of field data and Sentinel-2 NBAR, a process of extracting field observations according to the Sentinel-2 acquisition time is performed. Satellite-based high-resolution gross primary production calculation system.

2. In claim 1, The above data collection unit (110) Collecting data provided by a database on a server containing satellite Sentinel-2 NBAR data, EuroFlux and AmeriFlux in-situ photosynthetic active radiance (PAR) data, Satellite-based high-resolution gross primary production calculation system.

3. In claim 2, The above data preprocessing unit (120) For satellite Sentinel-2 NBAR data, c) For the data that has undergone the process of step 1, apply the eddy covariance footprint to calculate the average Sentinel-2 NBAR, but apply it using 80% of the cumulative footprint area. Calculate the average Sentinel-2 NBAR using {Equation 1} below. Satellite-based high-resolution gross primary production calculation system. {Formula 1} (CL stands for contour line, NBAR stands for adjusted reflectance value.) 4. In claim 3, The above data preprocessing unit (120) For data provided by EuroFlux towers, the average NBAR is derived by calculating the footprint based on the provided data. For data provided by AmeriFlux towers, we compute the NBAR averaged over a 3x3 grid centered on the tower. Satellite-based high-resolution gross primary production calculation system.

5. In claim 4, The GPP production model of the above data learning unit (130) is Photosynthetically active radiation (PAR) data is used as input, and GPP data is used as learning reference data to train and verify. It consists of a fully 1-D convolutional encoder-decoder network consisting of 11 1-D convolutional layers and 1 fully connected layer. Satellite-based high-resolution gross primary production calculation system.

6. In a method for calculating total primary production using a satellite-based high-resolution total primary production calculation system according to claim 1, Step 1: Collecting data provided from a database of a server having photosynthetically active radiation (PAR) data from satellite Sentinel-2 NBAR, EuroFlux and AmeriFlux in-situ in the data collection unit (110); In the data preprocessing unit (120), the second step is to preprocess the collected data and calculate the average NBAR (Nadir BRDF-Adjusted Reflectance) using the collected data; In the data learning unit (130), the calculated average NBAR and photosynthetically active radiation (PAR) are input values, and the 3rd step of calculating the gross primary production (GPP) through a deep learning model consisting of a fully convolutional encoder-decoder network is included. In the above 2nd step, the data preprocessing unit (120) is, in the case of the artificial satellite Sentinel-2 NBAR data, a) Using Sentinel-2 COPERNICUS / S2_SR_HARMONIZED NBAR data, the cloud percentage is set to 10%, b) Match the location of the site data by retrieving all useful images within the data period based on the image center coordinates on the site, and c) To synchronize the time of field data and Sentinel-2 NBAR, the process of extracting field observations according to the Sentinel-2 acquisition time is performed. Satellite-based high-resolution gross primary production calculation method.

7. In claim 6, The above two steps are: For satellite Sentinel-2 NBAR data, a) Using Sentinel-2 COPERNICUS / S2_SR_HARMONIZED NBAR data, the cloud percentage is set to 10%, b) Match the location of the site data by retrieving all useful images within the data period based on the image center coordinates on the site, and c) To synchronize the time of field data and Sentinel-2 NBAR, a process of extracting field observations according to the Sentinel-2 acquisition time is performed. d) For the data that has undergone the process of step c), apply the eddy covariance footprint to calculate the average Sentinel-2 NBAR, but apply it using 80% of the cumulative footprint area. A method for calculating satellite-based high-resolution gross primary production, comprising: calculating an average Sentinel-2 NBAR using the following {Formula 1}. {Formula 1} (CL stands for contour line, NBAR stands for adjusted reflectance value.) 8. In claim 7, The above two steps are: e) For data provided by EuroFlux towers, the average NBAR is derived by calculating the footprint based on the provided data. For data provided by AmeriFlux towers, a further step is included: calculating a 3×3 grid average NBAR centered on the tower; Satellite-based high-resolution gross primary production calculation method.

9. A computer-readable recording medium having recorded thereon a program for performing a satellite-based high-resolution total primary production calculation method according to claim 8.

Images