Machine Learning Using Hyperspectral Data Inaccurately Predicts Plant Traits Under Spatial Dependency

Abstract

Spectral, temporal and spatial dimensions are difficult to model together when predicting in situ plant traits from remote sensing data. Therefore, machine learning algorithms solely based on spectral dimensions are often used as predictors, even when there is a strong effect of spatial or temporal autocorrelation in the data. A significant reduction in prediction accuracy is expected when algorithms are trained using a sequence in space or time that is unlikely to be observed again. The ensuing inability to generalise creates a necessity for ground-truth data for every new area or period, provoking the propagation of “single-use” models. This study assesses the impact of spatial autocorrelation on the generalisation of plant trait models predicted with hyperspectral data. Leaf Area Index (LAI) data generated at increasing levels of spatial dependency are used to simulate hyperspectral data using Radiative Transfer Models. Machine learning regressions to predict LAI at different levels of spatial dependency are then tuned (determining the optimum model complexity) using cross-validation as well as the NOIS method. The results show that cross-validated prediction accuracy tends to be overestimated when spatial structures present in the training data are fitted (or learned) by the model.

Imaging spectroscopy reveals the effects of topography and logging on the leaf chemistry of tropical forest canopy trees

“…As anticipated, predictive models based on airborne hyperspectral measurements are less accurate, because tree-level spectral signatures combine light reflected by overlapping and adjacent crowns, in addition to canopy structural properties, that further modulate the signal (Asner et al, 2011(Asner et al, , 2017Doughty et al, 2017). Assessments of predictive performance are likely to be inflated when field samples are collected in a clustered manner, which is not accounted for during analysis (Rocha, Groen, Skidmore, Darvishzadeh, & Willemen, 2018), particularly if both trait and spectral variation is small within clusters relative to between clusters (e.g. Chadwick & Asner, 2016).…”

Tuning a statistical trade-off between spectral and spatial domains to predict plant traits with hyperspectral remote sensing

“…The other layers of plant traits were utilised only to simulate the hyperspectral data (see Rocha et al 2018 for full details on the simulation). The simulated spectra were used as explanatory variables to fit spatial and non-spatial regression models to predict LAI.…”

“…Machine learning regression can deal with the serially correlated predictors derived from the spectral domain if properly tuned (Dormann et al, 2007). However, these algorithms may pose challenges when dealing with serially correlated observations from spatially dependent plant traits and remote sensing data (Rocha et al, 2018).…”

“…These algorithms consider the entire range of the wavelengths as covariates to fit a model (Feilhauer et al, 2015;Moisen and Frescino, 2002;Ramoelo et al, 2013). Wavelengths captured by hyperspectral sensors have very strong multicollinearity because bands are serially correlated (Dormann et al, 2013;Rocha et al, 2018). Also, considerable noise can be captured in specific regions of the spectrum, depending on the capacity of a sensor's platform to control variations in illumination and view geometry (Combal et al, 2002).…”

“…Machine learning algorithms when tuned to restrict model complexity are capable of dealing with serially correlated hyperspectral wavelengths (Dormann et al, 2007;Rocha et al, 2017). However, non-spatial regression approaches face challenges when modelling with spatially autocorrelated observations derived from plant traits or remote sensing data (Rocha et al, 2018). Spatial autocorrelation is often neglected even when the in situ plant trait measurements are exceptionally close to each other, yet in analyses are considered as randomly distributed observations (Hoeting, 2009;Legendre et al, 2004).…”