Satellite soil moisture estimation method using organic matter time series variability modeling

Abstract

A satellite soil moisture estimation method using organic matter time series variability modeling according to the present invention comprises the steps of: generating a time series dataset on the basis of a soil organic matter map without time variability; identifying missing dates by year by interpolating the dataset; training a machine learning model to have dynamic characteristics having a plurality of delay times using the interpolated dataset; matching each observation point with variable input data having each delay time obtained through the machine learning model; parameterizing the delay time by regression analysis of a delay time model; parameterizing a delay temperature by deriving a temperature function directly related to organic matter variability through the parameterized delay time; parameterizing variable organic matter by giving variability to an organic matter map through the parameterized delay temperature; and improving a pre-stored dielectric constant model through the parameterized variable organic matter.

Description

Satellite Soil Moisture Estimation Method via Organic Matter Time Series Variability Modeling

[0001] The present invention relates to a method for estimating satellite soil moisture through organic time series variability modeling, and more specifically, to a method for estimating satellite soil moisture through organic time series variability modeling that takes organic data without time variability as input, models the characteristics of time variability, and then improves the accuracy of soil moisture estimation through machine learning to estimate the time variability of organic matter by grid.

[0002] Figure 1 is a diagram showing a cross-section of soil organic matter and soil moisture in the Alaska region as a latitude and longitude superposition, and Figure 2 is a diagram explaining the error between the predicted soil moisture variability and the actual observed value depending on whether soil organic matter variability is taken into account.

[0003] While organic matter information is a crucial factor in estimating satellite soil moisture based on permittivity models, conventional organic matter information is provided to radiative transfer models and machine learning models in the form of organic matter maps.

[0004] However, when organic matter information is provided in the form of a map with no temporal variability, errors occur in soil moisture estimation.

[0005] For example, as shown in Figure 1, the temporal variability of satellite-based soil moisture estimates in Alaska, where soil organic matter is high, appears unrealistically low compared to field-observed soil moisture values.

[0006] In addition, Figure 2 (a) illustrates the case where the variability of soil organic matter is not considered, (b) illustrates the case where only the spatial variability of soil organic matter is considered, and (c) illustrates the case where both the temporal and spatial variability of soil organic matter are considered. As shown in Figure 2, when the variability of soil organic matter content is not considered in soil moisture sensors and satellite observation systems using a permittivity model, the satellite-based soil moisture estimate appears significantly lower compared to the actual change in soil moisture.

[0007] Soil organic matter acts as a function that varies with temperature. While most literature uses temperature data from the same time period for the parameterization of organic matter, variability, such as organic matter decomposition, is accompanied by a time lag in response to temperature changes.

[0008] Therefore, to improve the accuracy of soil moisture estimation, soil moisture variability and organic matter variability from sensors must be considered together, and organic matter variability can be parameterized into temperature variability.

[0009] Furthermore, since there is a time lag between temperature and organic matter, the dynamic information of organic matter used in machine learning must consider all possible temperature lags as a function of temperature.

[0010] According to the present invention, the purpose is to provide a satellite soil moisture estimation method through organic matter time series variability modeling, which takes organic matter data without time variability as input, models the characteristics of time variability, and then improves the accuracy of soil moisture estimation through machine learning and estimates the time variability of organic matter by grid.

[0011] According to one embodiment of the present invention for achieving such technical challenges, a method for estimating satellite soil moisture through organic matter time series variability modeling comprises: generating a time series dataset based on a soil organic matter map without time variability; identifying missing yearly dates by interpolating the dataset; training a machine learning model with dynamic characteristics having multiple lag times using the interpolated dataset; matching each observation point with variability input data having each lag time obtained through the machine learning model; parameterizing the lag time by performing regression analysis on the lag time model; parameterizing the lag temperature by deriving a temperature function directly related to organic matter variability through the parameterized lag time; parameterizing the variable organic matter by imparting variability to the organic matter map through the parameterized lag temperature; and improving a previously stored permittivity model through the parameterized variable organic matter.

[0012] As such, according to the present invention, the temporal variability of organic matter is generated as a temperature-delayed time series, and by learning this with a random forest model to model the optimal variability, the uncertainty problem can be solved through machine learning by considering fixed map-shaped input data as characteristics that change over time.

[0013] In addition, the high-precision soil moisture information estimated by the present invention and the learned organic matter variability information provide important information for smart and sustainable carbon agriculture, which can contribute to technological development in multiple fields such as agriculture, climate change, and environmental monitoring.

[0014] Figure 1 is a cross-section of soil organic matter and soil moisture in the Alaska region, superimposed with latitude and longitude.

[0015] Figure 2 is a diagram illustrating the error between the predicted soil moisture variability and the actual observed value depending on whether soil organic matter variability is considered.

[0016] FIG. 3 is a diagram illustrating the flow of a method for estimating satellite soil moisture through organic time series variability modeling according to one embodiment of the present invention.

[0017] FIG. 4 is a diagram illustrating an example of configuring a time series dataset according to an embodiment of the present invention.

[0018] FIG. 5 is a diagram illustrating the matching of variable input data with each delay time and each observation point according to an embodiment of the present invention.

[0019] FIG. 6 is a diagram illustrating organic variability reflecting weighted average potential dynamic characteristics according to one embodiment of the present invention.

[0020] FIG. 7 is a diagram illustrating the results of improving the predictive ability of machine learning according to one embodiment of the present invention.

[0021] FIG. 8 is a diagram illustrating an example of a soil organic matter estimation result according to one embodiment of the present invention.

[0022] Preferred embodiments according to the present invention will be described in detail below with reference to the attached drawings. In this process, the thickness of lines or the size of components shown in the drawings may be exaggerated for clarity and convenience of explanation.

[0023] Furthermore, the terms described below are defined in consideration of their functions within the present invention, and these may vary depending on the intent or practice of the user or operator. Therefore, the definitions of these terms should be based on the content throughout this specification.

[0024] Hereinafter, a method for estimating satellite soil moisture through organic matter time series variability modeling according to an embodiment of the present invention will be described in detail with reference to FIGS. 3 to 5.

[0025] FIG. 3 is a diagram illustrating the flow of a method for estimating satellite soil moisture through organic matter time series variability modeling according to an embodiment of the present invention, FIG. 4 is a diagram illustrating an example of constructing a time series dataset according to an embodiment of the present invention, and FIG. 5 is a diagram for explaining the matching of variability input data with each delay time and each observation point according to an embodiment of the present invention.

[0026] At this time, the satellite soil moisture estimation method through organic time series variability modeling shown in Fig. 3 can be performed by a processor.

[0027] As illustrated in FIG. 3, a method for estimating satellite soil moisture through organic matter time series variability modeling according to one embodiment of the present invention first generates a time series dataset to reflect the variability of soil organic matter over time based on a soil organic matter map that has no time variability (S310).

[0028] At this time, the generated dataset includes temperature, organic matter map, n-day delayed time variation, clay content, microwave permittivity, output value, and ground soil moisture, as shown in Fig. 4, and is implemented within the dimension of [observation time series t * observation point s].

[0029] Next, the generated dataset is interpolated to identify the missing year-by-year date (DoY) (S320).

[0030] More specifically, the linear interpolated value y of the numerical variable for missing dates in the microwave permittivity dataset from microwave satellites (e.g., SMAP) and soil moisture sensors (e.g., TDR, FDR) is the missing date x out When this is located between two known dates x1 and x2, it can be calculated as shown in Equation 1 below using the known values ​​y1 and y2 of the corresponding variable.

[0031]

[0032] For example, organic map data is generated by repeatedly applying an organic map for each observation point for 365 days, and n-day delayed time variation data is generated by applying an n-day delayed temperature change function to all observation points.

[0033] For dates where observations do not exist, x1 and x2 are the dates on which data exists (e.g., dates on which measurements were taken) and the values ​​of numerical variables y1 and y2 (e.g., soil temperature, dielectric constant, soil moisture) at those dates, and x of the missing dates. out The missing y values ​​(e.g., soil temperature, dielectric constant, soil moisture) can be estimated.

[0034] This interpolation method can cover all days of the year (1 to 365 days) with missing dates and can be applied to datasets that are revisited every 3 days, such as SMAP, or datasets with severely insufficient time-series observations, such as SMAPVEX12, to generate observations with time continuity.

[0035] Through this, by generating time-series data that includes all realistic latency, it prevents performance degradation of machine learning models that effectively learn temporal elements and significantly increases the accuracy of soil moisture prediction.

[0036] Next, using a time series dataset generated through a machine learning model, dynamic features with multiple latency are trained on the machine learning model (S330).

[0037] In this case, the machine learning model can be trained using the Random Forest technique, but it can also be trained using other machine learning models.

[0038] Next, each observation point is matched with the variable input data having each delay time (S340).

[0039] More specifically, since organic values ​​that appear frequently in the organic map have a wide area of ​​effect, time-varying input data of high importance can be matched in order of frequency of organic value distribution.

[0040] Accordingly, unique time variability can be assigned to each grid in the organic map.

[0041] Next, the delay time model is regression-analyzed to parameterize the delay time (S350).

[0042] In this context, the time lag model is a model that quantifies the temporal delay in which the effect of temperature change on changes in soil organic matter does not appear immediately, but rather manifests after a certain period of time.

[0043] Specifically, as temperature changes increase, the metabolic activity of microorganisms in the soil increases; however, the activation of these microorganisms and the decomposition of organic matter do not occur immediately, and the reactions appear only after a certain period of time has passed.

[0044] The time delay model of the present invention models this temporal delay and is based on the concept that organic matter variability observed at the present time is a result of past temperature changes. Through this, the delay between temperature change and organic matter variability can be quantitatively explained.

[0045] This explains in more detail the phenomenon of microbial activity, temperature, and the delay in reaching the final organic matter decomposition process.

[0046] The activity of microorganisms in the soil is greatly affected by temperature.

[0047] When temperature rises, the metabolic rate of microorganisms increases, accelerating the decomposition of organic matter, but there is a physiological delay in this process.

[0048] For microorganisms to become active, processes such as intracellular enzyme synthesis and the regulation of metabolic pathways are required, and these processes take a certain amount of time. In addition, each microbial species has an optimal temperature range, and microbial activity may be reduced or delayed if it falls outside this range.

[0049] Microbial activity responds to temperature changes according to a sigmoid curve, characterized by minimal activity initially followed by a rapid increase once the optimal temperature is reached.

[0050] However, since this activation process takes time, immediate decomposition of organic matter due to temperature changes does not occur, and a delay time is required.

[0051] The decomposition process of soil organic matter can also be delayed over several stages.

[0052] Soil organic matter consists of complex chemical structures such as lignin and cellulose, and it takes time to decompose these materials.

[0053] The decomposition of organic matter proceeds through chemical reaction steps such as hydrolysis, oxidation, and dehydrogenation, and each step occurs through the action of microbial enzymes.

[0054] Therefore, time is required for these chemical reactions to occur, and the process of microorganisms synthesizing and secreting enzymes to decompose organic matter also causes delays.

[0055] All of these processes cause temporal variability in soil organic matter and are the reason why temperature changes do not show an immediate effect on organic matter decomposition.

[0056] Based on this background, the delay time model of the present invention accurately reflects the delay time occurring during the microbial activation and organic matter decomposition processes following a temperature change, thereby considering that current organic matter variability is a result of past temperature changes.

[0057] This allows for the quantification of the time gap between temperature changes and organic matter decomposition, and enables more precise soil moisture estimation.

[0058] Next, the delay temperature is parameterized by deriving a temperature function directly related to organic variability through the parameterized delay time (S360).

[0059] At this time, a temperature function directly related to organic matter variability can be derived by setting the seasonally fluctuating temperature function as an exponential-cosine composite function as shown in Equation 2 below.

[0060]

[0061] At this time, T lag is the delay temperature, and T max is the annual maximum temperature, T min is the annual minimum temperature, t is the number of days in the year, is the number of days of delay.

[0062] In this case, the temperature function is set to an exponential-cosine composite function to apply an average temperature fluctuation function instead of real-time temperature, thereby preventing organic matter from being simulated with excessive sensitivity due to daytime and nighttime temperature changes, noise in temperature observations, etc.

[0063] In addition, the Shared Latent Dynamic Feature of the organic material can be derived from the induced delayed temperature function as shown in Equation 3 below.

[0064]

[0065] If it is estimated by location (e.g., observation point or grid cell) as a delay of days, It can be written as.

[0066] in other words, is a phase shift parameter that can be individually specified (or learned) for each spatial position s, where at that position Determines the point in time when the maximum and minimum values ​​occur.

[0067] As increases, the seasonal curve of the same shape shifts parallel to the left and right along the DoY axis, while the shape of the curve itself is maintained. Therefore, in actual application, by position It is not fixed in advance, but estimated from the data during the learning process.

[0068] one side, is a parameter that controls the shape of the curve, As the value increases, the seasonal peaks become sharper and the rising and falling phases become more abrupt, whereas, The smaller the value, the gentler and wider the peaks become, and the smoother the changes throughout the year.

[0069] For example, if the concentration of organic matter (SOM) shows a pattern of rapidly increasing followed by rapidly decreasing at a specific time, large The value effectively simulates this.

[0070] FIG. 5 is a diagram to visually explain this relationship, the left section In the same DoY By changing only the sharpness difference of the peaks, the right section The same in at It shows that only the peak point moves while maintaining the shape of the curve by changing only.

[0071] In other words, the phase shift of the sine input It changes the point in time when the maximum value occurs, and as a result, the entire curve shifts parallel along the time axis.

[0072] thus, is “phase alignment”, Since it is responsible for “shape tuning,” the shared latent dynamics having different seasonal variation patterns by region and year are obtained through the combined optimization of the two parameters. It can precisely track and reproduce.

[0073] This allows for the flexible capture of various seasonal patterns (e.g., sharp unipeaks, gentle broads, etc.) within the same model structure, while mitigating excessive sensitivity caused by day-night temperature fluctuations and observation noise by using an exponential-cosine composite function (Equation 2) that reflects average seasonal variation instead of real-time temperature.

[0074] In one embodiment It can be set to 5, but this is merely an example for illustrative purposes and can be changed depending on data characteristics and model objectives. Also, It is not limited to predefined values ​​and can be automatically estimated by location through a learning process.

[0075] In addition, as shown in Equation 4 below, weights according to latent dynamic characteristics by latency through importance normalization of Random Forest Calculates.

[0076]

[0077] Here, I is the importance of the random forest.

[0078] Subsequently, as shown in Mathematical Formula 5 below, for each delay time calculated in this way and Weighted average latent dynamic characteristic F, which is simulated closer to reality through normalization of the sum of products. t Produces.

[0079]

[0080] Here, L is the number of days of delay.

[0081] Next, the variable organic matter is parameterized by imparting variability to the organic matter map through the parameterized delayed temperature (S370).

[0082] More specifically, the induced weighted average latent dynamic characteristics and the organic map (OM map)'s maximum(OM max ) and minimum(OM min By introducing variability as ), an organic function can be derived as shown in Equation 6 below.

[0083]

[0084] Finally, the stored permittivity model is improved through parameterized fluctuating organic matter (S370).

[0085] Through this, the machine learning-based accuracy of soil moisture can be further improved.

[0086] Furthermore, by utilizing the organic matter lag information and temperature function obtained from machine learning to apply fluctuating organic matter input data to the permittivity model, which is a stored physical model, the existing static input data, the organic matter map (OM) map The accuracy of soil moisture estimation can be improved compared to when ) is applied. Furthermore, when simultaneously estimating organic matter, optimal estimation becomes possible based on more accurate initial organic matter values.

[0087] FIG. 6 is a diagram illustrating organic variability reflecting weighted average potential dynamic characteristics according to one embodiment of the present invention.

[0088] More specifically, FIG. 6 presents the results of comparing and verifying at two sites (D08.TALL, D02.BLAN) a SOC time series (solid line) reconstructed using a weighted average latent dynamic having positional phase and shape according to an embodiment of the present invention, and a physics-based optimal estimation result (dotted line) calculated by inversely calculating the SOC at each time point by the physical coupling equation of permittivity-moisture-organic matter.

[0089] The two curves show high agreement in terms of the timing of peaks (maximum values) and troughs (minimum values), the slopes of rising and falling sections, and seasonal amplitudes, confirming that consistent estimation is secured despite methodological independence (physics law-based correction vs. data-based learning).

[0090] In particular, the machine learning-based configuration of the present invention is implemented as end-to-end learning that directly minimizes soil moisture prediction errors, and automatically obtains the seasonal combined structure of SOC-SM (soil moisture) from data without physics-based inversion by co-learning SOC as a latent driving term.

[0091] In addition, it is operationally robust against observation noise, missing values, and anomaly events through the fusion of multi-source data (satellite, ground, and weather) and a learning pipeline that includes robust loss.

[0092] By observing the same trend at different NEON points of the left and right panels, it is proven that the proposed method has universality and generalizability applicable regardless of point characteristics.

[0093] Therefore, the weighted average latent dynamic feature model of the present invention demonstrates that (i) the causal interpretability provided by the physical model and (ii) the data fitting capability provided by machine learning converge to the same dynamic feature, and can accurately reproduce SOC seasonal variation even in the presence of observation gaps and noise.

[0094] As a result, it effectively alleviates the constraints on assumptions, identifiability, boundary conditions, and computational costs inherent in physics-based optimal estimation, while providing practical and concrete technical effects in terms of prediction accuracy, generalization, and robustness.

[0095] FIG. 7 is a diagram illustrating the results of improving the predictive ability of machine learning according to an embodiment of the present invention, FIG. 7 (a) shows soil moisture estimated using a random forest method with an organic matter map as input data, and FIG. 7 (b) shows soil moisture estimated by machine learning with a random forest method together with an organic matter map using input data to which a plurality of organic matter variability is applied as a soil temperature lag function.

[0096] At this time, information on ground soil moisture, microwave brightness temperature, soil temperature, clay, NDVI, and soil roughness is commonly applied to the learning process.

[0097] In addition, according to one embodiment of the present invention, as shown in FIG. 8, machine learning prediction ability can be improved by using input data without dynamic variability and input data with artificially generated variability.

[0098] FIG. 8 is a diagram illustrating an example of a soil organic matter estimation result according to an embodiment of the present invention, FIG. 8 (a) illustrates an example of optimal organic matter estimation with an organic matter map as the initial value, and FIG. 8 (b) illustrates an example of optimal organic matter estimation with variability in the organic matter map using a delayed temperature as the initial value.

[0099] As a result, as shown in Fig. 8, even when information on organic matter is given only in the form of a map, by incorporating multiple temporal variability into machine learning, the machine learning model can provide more accurate and realistic predictions.

[0100] As such, according to the present invention, the temporal variability of organic matter is generated as a temperature-delayed time series, and by learning this with a random forest model to model the optimal variability, the uncertainty problem can be solved through machine learning that considers fixed map-shaped input data as characteristics that change over time.

[0101] In addition, the high-precision soil moisture information estimated by the present invention and the learned organic matter variability information provide important information for smart and sustainable carbon agriculture, which can contribute to technological development in multiple fields such as agriculture, climate change, and environmental monitoring.

[0102] The present invention has been described with reference to the embodiments illustrated in the drawings, but this is merely illustrative, and those skilled in the art will understand that numerous modifications and equivalent alternative embodiments are possible therefrom. Accordingly, the true technical scope of protection of the present invention should be determined by the technical spirit of the following claims.

Claims

1. A step of generating a time series dataset based on a soil organic matter map with no temporal variability; A step of identifying missing year-by-year dates by interpolating the above dataset; A step of training a machine learning model to have dynamic features with multiple latency times using the interpolated dataset above; A step of matching each observation point with variable input data having each delay time obtained through the machine learning model above; Step of parameterizing the delay time by performing regression analysis on the delay time model; A step of deriving a temperature function directly related to organic matter variability through the above-mentioned parameterized delay time to parameterize the delay temperature; A step of parameterizing the fluctuating organic matter by imparting variability to the organic matter map through the above parameterized delayed temperature; and A method for estimating satellite soil moisture through organic time series variability modeling, comprising the step of improving a previously stored permittivity model through the parameterized fluctuating organic matter.

2. In Paragraph 1, The above time series dataset is, A method for estimating satellite soil moisture through organic time series variability modeling, which includes temperature, organic matter map, n-day delayed time variation, clay content, microwave permittivity, output value, and ground soil moisture, and is implemented within the dimension of [observation time series t * observation point s].

3. In Paragraph 1, The step of identifying missing year-by-year dates by interpolating the above dataset is: The linear interpolation value y of the numerical variable for missing dates in the microwave permittivity dataset from microwave satellites and soil moisture sensors is the missing date x out When this is located between two known dates x1 and x2, using the known values ​​y1 and y2 of the corresponding variable, the following mathematical formula A method for estimating satellite soil moisture through organic matter time series variability modeling calculated as follows.

4. In Paragraph 3, The step of parameterizing the above delay temperature is, The following mathematical formula As shown above, by setting the seasonally fluctuating temperature function as a composite exponential-cosine function to derive a temperature function directly related to organic matter variability, the lagged temperature is parameterized, and T lag is the delay temperature, and T max is the annual maximum temperature, T min is the annual minimum temperature, t is the number of days in the year, is a satellite soil moisture estimation method through organic time series variability modeling with a lag of days.

5. In Paragraph 4, The Shared Latent Dynamic Feature of the organic material in the above temperature function is given by the following mathematical formula Derived as in, The following mathematical formula Weights based on latent dynamic characteristics by latency using importance normalization of Random Forest as shown Calculate, The following mathematical formula By delay time calculated as follows and Weighted average latent dynamic characteristic F through normalization of the sum of products t It produces, A method for estimating satellite soil moisture through organic time series variability modeling, where is the sharpness, I is the random forest importance, and L is the number of lag days.

6. The step of parameterizing the above-mentioned fluctuating organic matter is, The weighted average latent dynamic characteristics derived above and the organic map (OM) map )'s maximum(OM max ) and minimum(OM min By introducing variability with ), the following mathematical formula A satellite soil moisture estimation method through organic time series variability modeling that derives an organic matter function, as shown above.

Images