머신러닝 기법의 산림 총일차생산성 예측 모델 비교

Predicting Forest Gross Primary Production Using Machine Learning Algorithms

Abstract

산림생태계에서 총일차생산성(Gross Primary Production, GPP)은 기후변화에 따른 산림의 생산성과 그에 영향을 미치는 식물계절, 건강성, 탄소 순환 등을 대표하는 지표이다. 총일차생산성을 추정하기 위해서는 에디공분산 타워 자료나 위성영상관측자료를 이용하기도 하고 물리지형적 한계나 기후변화 등을 고려하기 위해 기작기반모델링을 활용하기도 한다. 그러나 총일차생산성을 포함한 산림 탄소 순환의 기작기반 모델링은 식물의 생물, 생리, 화학적 기작들의 반응과 지형, 기후 및 시간 등과 같은 환경 조건들이 복잡하게 얽혀 있어 비선형적이고 유연성이 떨어져 반응에 영향을 주는 조건들을 모두 적용하기가 어렵다. 본 연구에서는 산림 생산성 추정 모델을 에디공분산 자료와 인공위성영상 정보를 사용하여 기계학습 알고리즘을 사용한 모델들로 구축해 보고 그 사용 및 확장 가능성을 검토해 보고자 하였다. 설명변수들로는 에디공분산자료와 인공위성자료에서 나온 대기기상인자들을 사용하였고 검증자료로 에디공분산 타워에서 관측된 총일차생산성을 사용하였다. 산림생산성 추정 모델은 1) 에디공분산 관측 기온($T_{air}$), 태양복사($R_d$), 상대습도(RH), 강수(PPT), 증발산(ET) 자료, 2) MODIS 관측 기온(T), 일사량($R_{sd}$), VPD 자료(개량식생지수 제외), 3) MODIS 관측 기온(T), 일사량($R_{sd}$), VPD, 개량식생지수(EVI) 자료를 사용하는 세 가지 경우로 나누어 구축하여 2006 - 2013년 자료로 훈련시키고 2014, 2015년 자료로 검증하였다. 기계학습 알고리즘은 support vector machine (SVM), random forest (RF), artificial neural network (ANN)를 사용하였고 단순 비교를 위해 고전적 방법인 multiple linear regression model (LM)을 사용하였다. 그 결과, 에디공분산 입력자료로 훈련시킨 모델의 예측력은 피어슨 상관계수 0.89 - 0.92 (MSE = 1.24 - 1.62), MODIS 입력자료로 훈련시킨 모델의 예측력은 개량식생지수 제외된 모델은 0.82 - 0.86 (MSE = 1.99 - 2.45), 개량식생지수가 포함된 모델은 0.92 - 0.93(MSE = 1.00 - 1.24)을 보였다. 이러한 결과는 산림총일차생산성 추정 모델 구축에 있어 MODIS인공위성 영상 정보 기반으로 기계학습 알고리즘을 사용하는 것에 대한 높은 활용가능성을 보여주었다.

Terrestrial Gross Primary Production (GPP) is the largest global carbon flux, and forest ecosystems are important because of the ability to store much more significant amounts of carbon than other terrestrial ecosystems. There have been several attempts to estimate GPP using mechanism-based models. However, mechanism-based models including biological, chemical, and physical processes are limited due to a lack of flexibility in predicting non-stationary ecological processes, which are caused by a local and global change. Instead mechanism-free methods are strongly recommended to estimate nonlinear dynamics that occur in nature like GPP. Therefore, we used the mechanism-free machine learning techniques to estimate the daily GPP. In this study, support vector machine (SVM), random forest (RF) and artificial neural network (ANN) were used and compared with the traditional multiple linear regression model (LM). MODIS products and meteorological parameters from eddy covariance data were employed to train the machine learning and LM models from 2006 to 2013. GPP prediction models were compared with daily GPP from eddy covariance measurement in a deciduous forest in South Korea in 2014 and 2015. Statistical analysis including correlation coefficient (R), root mean square error (RMSE) and mean squared error (MSE) were used to evaluate the performance of models. In general, the models from machine-learning algorithms (R = 0.85 - 0.93, MSE = 1.00 - 2.05, p < 0.001) showed better performance than linear regression model (R = 0.82 - 0.92, MSE = 1.24 - 2.45, p < 0.001). These results provide insight into high predictability and the possibility of expansion through the use of the mechanism-free machine-learning models and remote sensing for predicting non-stationary ecological processes such as seasonal GPP.

Fig. 1. Schematic diagram showing data flow and analysis. Models were trained in the testing period (2006 - 2013) by three different types of input data sets as follows, Type 1: air temperature (T_air), relative humidity (RH), daily net radiation (R_d), precipitation (PPT), and evapotranspiration (ET) from eddy flux measurement, Type 2: air temperature (T), daily shortwave radiation (R_sd), and vapor pressure deficit (VPD) from MODIS, and Type 3: air temperature (T), daily shortwave radiation (R_sd), and vapor pressure deficit (VPD), and EVI from MODIS. Gross Primary Production (GPP) calculated based on multiple linear regression model (LM), support vector machine (SVM), random forest (RF), and artificial neural network (ANN) and evaluated with eddy covariance GPP in 2014 and 2015.

Fig. 2. Daily GPP prediction obtained with linear regression model (LM), Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Network (ANN) (black opened circle) and GPP obtained eddy covariance (EC) measurement (gray closed circle) using EC measurement datasets in 2014 and 2015.

Fig. 3. Comparisons of Eddy Covariance (EC) measurement GPP and modeled GPP from the trained models by EC measurement datasets in 2014 and 2015. LM = linear regression model, SVM = Support Vector Machine, RF = Random Forest, ANN = Artificial Neural Network.

Fig. 4. Daily GPP prediction obtained with linear regression model (LM), Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Network (ANN) (black opened circle) and GPP obtained eddy covariance (EC) measurement (gray closed circle) using MODIS datasets (without Enhanced Vegetation Index, EVI) in 2014 and 2015.

Fig. 5. Comparisons of Eddy Covariance (EC) measurement GPP and modeled GPP from the trained models by MODIS datasets (without Enhanced Vegetation Index, EVI) in 2014 and 2015. LM = linear regression model, SVM = Support Vector Machine, RF = Random Forest, ANN = Artificial Neural Network.

Fig. 6. Daily GPP prediction obtained with linear regression model (LM), Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Network (ANN) (black opened circle) and GPP obtained eddy covariance (EC) measurement (gray closed circle) using MODIS datasets (with EVI) in 2014 and 2015.

Fig. 7. Comparisons of Eddy Covariance (EC) measurement GPP and modeled GPP from the trained models by MODIS datasets (with EVI) in 2014 and 2015. LM = linear regression model, SVM = Support Vector Machine, RF = Random Forest, ANN = Artificial Neural Network.

Table 1. Summary of MODIS Land Products used in explanatory variables from 2006 to 2015

Table 2. Summary of MODIS Atmosphere Products used in explanatory variables from 2006 to 2015

Table 3. Table Summary of statistics for the different algorithms based on the three different input sets. R, RMSE, STD and MSE denote correlation coefficient, root mean square error, standard deviation and mean squared error, respectively