Completing the machine learning saga in fractional snow cover estimation from MODIS Terra reflectance data: Random forests versus support vector regression

Kuter, Semih

doi:10.1016/j.rse.2021.112294

Cited by 68 publications

(33 citation statements)

References 89 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Compared to alternative methods such as artificial neural networks (like the GRNN model used here), SVR can yield comparable accuracy using a much smaller training sample size (Mountrakis et al, 2011). This is in line with the "support vector" concept that relies only on a few data points to define the position of the decision surface (Huang and Zhao, 2018;Kuter, 2021;Mountrakis et al, 2011). In light of this research, the SVR model had the ability to accurately modeling δ 15 N-NO 3 − values in surface water relying on commonly measured hydro-chemical variables, and thereby serves as an indirect, rapid, and convenient tool for environmental stable isotope prediction.…”

Section: Discussionsupporting

confidence: 64%

A support vector regression model to predict nitrate-nitrogen isotopic composition using hydro-chemical variables

Yang

Shang

Chen

et al. 2021

Journal of Environmental Management

View full text Add to dashboard Cite

Nitrate is a prominent pollutant in surface and groundwater bodies worldwide. Isotopes in nitrate provide a powerful approach for tracing nitrate sources and transformations in waters. Given that analytical techniques for determining isotopic compositions are generally time-consuming, laborious and expensive, alternative methods are warranted to supplement and enhance existing approaches. Hence, we developed a support vector regression (SVR) model and explored its feasibility to predict nitrogen isotopic composition of nitrate (δ 15 N-NO 3 − ) in a ruralurban river system in Southeastern China. A total of 16 easily obtained hydro-chemical variables were measured in the wet season (September 2019) and dry season (January 2020) and used to develop the SVR prediction model. The grading method utilized ~75% (35) of the samples for model building while the remaining 11 samples assessed model performance. Principal component analysis (PCA) extracted 7 principal components for SVR model inputs as PCA reduces superfluous variables. We optimized tuning parameters in the SVR model using a grid search technique coupled with V-fold cross-validation. The optimized SVR model provided accurate δ 15 N-NO 3 − predictions with a determination coefficient (R 2 ) of 0.88, Nash-Sutcliffe (NS) of 0.87, and mean square error (MSE) of 0.53‰ in the testing step, and performed much better than the corresponding multivariate linear regression model (R 2 = 0.60, NS = 0.58 and MSE = 1.76‰) and general regression neural network model (R 2 = 0.66, NS = 0.65 and MSE = 1.45‰). Overall, the SVR model provides a potential indirect method to predict environmental isotope values for water quality management that will complement and enhance the interpretation of direct measurements of δ 15 N-NO 3 − .

show abstract

Section: Discussionsupporting

confidence: 64%

A support vector regression model to predict nitrate-nitrogen isotopic composition using hydro-chemical variables

Yang

Shang

Chen

et al. 2021

Journal of Environmental Management

View full text Add to dashboard Cite

show abstract

“…In the last step, a linear model is built by the new derived feature space to minimize the errors [73]. The SVM algorithm based on the Radial Basis Function (RBF) (the most popular choice in the literature) had two hyper-parameters: the penalty factor C aiming to find a trade-off between the fitting error and the model "complexity", and the kernel width gamma [74]. The SVM hyper-parameters were tuned with a grid search method.…”

Section: Support Vector Machine (Svm)mentioning

confidence: 99%

Cereal Yield Forecasting with Satellite Drought-Based Indices, Weather Data and Regional Climate Indices Using Machine Learning in Morocco

et al. 2021

View full text Add to dashboard Cite

Accurate seasonal forecasting of cereal yields is an important decision support tool for countries, such as Morocco, that are not self-sufficient in order to predict, as early as possible, importation needs. This study aims to develop an early forecasting model of cereal yields (soft wheat, barley and durum wheat) at the scale of the agricultural province considering the 15 most productive over 2000–2017 (i.e., 15 × 18 = 270 yields values). To this objective, we built on previous works that showed a tight linkage between cereal yields and various datasets including weather data (rainfall and air temperature), regional climate indices (North Atlantic Oscillation in particular), and drought indices derived from satellite observations in different wavelengths. The combination of the latter three data sets is assessed to predict cereal yields using linear (Multiple Linear Regression, MLR) and non-linear (Support Vector Machine, SVM; Random Forest, RF, and eXtreme Gradient Boost, XGBoost) machine learning algorithms. The calibration of the algorithmic parameters of the different approaches are carried out using a 5-fold cross validation technique and a leave-one-out method is implemented for model validation. The statistical metrics of the models are first analyzed as a function of the input datasets that are used, and as a function of the lead times, from 4 months to 2 months before harvest. The results show that combining data from multiple sources outperformed models based on one dataset only. In addition, the satellite drought indices are a major source of information for cereal prediction when the forecasting is carried out close to harvest (2 months before), while weather data and, to a lesser extent, climate indices, are key variables for earlier predictions. The best models can accurately predict yield in January (4 months before harvest) with an R2 = 0.88 and RMSE around 0.22 t. ha−1. The XGBoost method exhibited the best metrics. Finally, training a specific model separately for each group of provinces, instead of one global model, improved the prediction performance by reducing the RMSE by 10% to 35% depending on the provinces. In conclusion, the results of this study pointed out that combining remote sensing drought indices with climate and weather variables using a machine learning technique is a promising approach for cereal yield forecasting.

show abstract

“…(2) the multicollinearity problem between features can be effectively avoided, supporting high-dimensional features [114,115]; (3) the hierarchical relationships can be characterized, showing good stability in regard to missing and unbalanced data [116]; (4) it has strong robustness to the outliers in the variables, and handles these outliers through the binding method [117]; (5) they can also provide the importance of explanatory variables to determine their contributions for the model and the main predictive factors [117,118]. Moreover, the RF algorithm also has significant advantages in terms of OOB error, and it is generally considered a good estimate of the expected error of unseen data [119].…”

Section: Applicability Of Machine Learning Algorithmsmentioning

confidence: 99%

Mapping Regional Soil Organic Matter Based on Sentinel-2A and MODIS Imagery Using Machine Learning Algorithms and Google Earth Engine

Zhang

Yang

et al. 2021

Remote Sensing

View full text Add to dashboard Cite

Many studies have attempted to predict soil organic matter (SOM), whereas mapping high-precision and high-resolution SOM maps remains a challenge due to the difficulty of selecting appropriate satellite data sources and prediction algorithms. This study aimed to investigate the influence of different remotely sensed images and machine learning algorithms on SOM prediction. We constructed two comparative experiments, i.e., full-band and common-band variable datasets of Sentinel-2A and MODIS images using Google Earth Engine (GEE). The predictive performances of random forest (RF), artificial neural network (ANN), and support vector regression (SVR) algorithms were evaluated, and the SOM map was generated for the Songnen Plain. Results showed that the model based on the full-band Sentinel-2A dataset achieved the best performance. The application of Sentinel-2A data resulted in mean relative improvements (RIs) of 7.67% and 5.87%, respectively. The RF achieved a lower root mean squared error (RMSE = 0.68%) and a higher coefficient of determination (R2 = 0.67) in all of the predicted scenarios than ANN and SVR. The resultant SOM map accurately characterized the SOM spatial distribution. Therefore, the Sentinel-2A data have obvious advantages over MODIS due to their higher spectral and spatial resolutions, and the combination of the RF algorithm and GEE is an effective approach to SOM mapping.

show abstract

Completing the machine learning saga in fractional snow cover estimation from MODIS Terra reflectance data: Random forests versus support vector regression

Cited by 68 publications

References 89 publications

A support vector regression model to predict nitrate-nitrogen isotopic composition using hydro-chemical variables

A support vector regression model to predict nitrate-nitrogen isotopic composition using hydro-chemical variables

Cereal Yield Forecasting with Satellite Drought-Based Indices, Weather Data and Regional Climate Indices Using Machine Learning in Morocco

Mapping Regional Soil Organic Matter Based on Sentinel-2A and MODIS Imagery Using Machine Learning Algorithms and Google Earth Engine

Contact Info

Product

Resources

About