Análisis de métodos de validación cruzada para la obtención robusta de parámetros biofísicos

Abstract

Resumen: Los métodos de regresión no paramétricos son una gran herramienta estadística para obtener parámetros biofísicos a partir de medidas realizadas mediante teledetección. Pero los resultados obtenidos se pueden ver afectados por los datos utilizados en la fase de entrenamiento del modelo. Para asegurarse de que los modelos son robustos, se hace uso de varias técnicas de validación cruzada. Estas técnicas permiten evaluar el modelo con subconjuntos de la base de datos de campo. Aquí, se evalúan dos tipos de validación cruzada en el desarrollo de modelos de regresión no paramétricos: hold-out y k-fold. Los métodos de regresión lineal seleccionados fueron: Linear Regression (LR) y Partial Least Squares Regression (PLSR). Y los métodos no lineales: Kernel Ridge Regression (KRR) y Gaussian Process Regression (GPR). Los resultados de la validación cruzada mostraron que LR ofrece los resultados más inestables, mientras KRR y GPR llevan a resultados más robustos. Este trabajo recomienda utilizar algoritmos de regresión no lineales (como KRR o GPR) combinando con la validación cruzada k-fold con un valor de k igual a 10 para hacer la estimación de una manera robusta.Palabras clave: Hold-Out, k-fold, validación cruzada, aprendizaje automático, procesos gausianos, regresión de Kernel Ridge. Analysis of cross-validation methods for robust retrieval of biophysical parametersAbstract: Non-parametric regression methods are powerful statistical methods to retrieve biophysical parameters from remote sensing measurements. However, their performance can be affected by what has been presented during the training phase. To ensure robust retrievals, various cross-validation sub-sampling methods are often used, which allow to evaluate the model with subsets of the field dataset. Here, two types of cross-validation techniques were analyzed in the development of non-parametric regression models: hold-out and k-fold. Selected non-parametric linear regression methods were least squares Linear Regression (LR) and Partial Least Squares Regression (PLSR), and nonlinear methods were Kernel Ridge Regression (KRR) and Gaussian Process Regression (GPR). Cross-validation results showed that LR performed most unstable, while KRR and GPR led to more robust results. This work recommends using a nonlinear regression algorithm (e.g., KRR, GPR) in combination with a k-fold cross-validation technique with k=10 to realize robust retrievals.

Estimation of Total Nitrogen Content in Forage Maize (Zea mays L.) Using Spectral Indices: Analysis by Random Forest

“…In the present study, we used 70% of the data for adjustment and 30% for validation. These percentages of data division are similar to those reported by other authors [12,29,53,54,57].…”

Improving the remote estimation of soil organic carbon in complex ecosystems with Sentinel-2 and GIS using Gaussian processes regression

“…A Gaussian Process (GP) defines a distribution over functions. In other words, it generates a finite set of random variables, of which there is a joint Gaussian distribution; GPR is a non-linear regression method that uses non-parametric Bayesian modelling that considers the variance of the data set and a maximization of the probability margin in the training set using a scaled anisotropic Gaussian kernel function (see Table 3 ) (Pérez-Planells et al 2015 ). GPR allows identifying the important characteristics of the input variables (Rasmussen and Williams 2006 ) and evaluating the relative contribution highlighting the most relevant bands or parameters to the prediction model.…”

Satellite and historical data, and statistical modeling to predict potential fishing zones for dolphinfish, Coryphaena hippurus (Perciformes: Coryphaenidae) in Colombian Pacific

“…We used a cross-validation approach for the predictions of CPUE (Pérez-Planells et al, 2015). This approach was performed with a random subset containing 30 % of the data that was not involved in modeling.…”