MODIS VCF should not be used to detect discontinuities in tree cover due to binning bias. A comment on Hanan et al. (2014) and Staver and Hansen (2015)

Abstract

Lloyd, Jon. 2017. MODIS VCF should not be used to detect discontinuities in tree cover due to binning bias. A comment on Hanan et al. (2014) and Staver and Hansen (2015). Global Ecology and Biogeography, 26 (7). 854-859. 10.1111/geb.12592 Contact CEH NORA team at noraceh@ceh.ac.ukThe NERC and CEH trademarks and logos ('the Trademarks') are registered trademarks of NERC in the UK and other countries, and may not be used without the prior written consent of the Trademark owner. ABSTRACTIn their recent paper, Staver and Hansen (Global Ecology and Biogeography, 2015, 24, 985-987) refute the case made by Hanan et al. (Global Ecology and Biogeography, 2014, 23, 259-263) that the use of classification and regression trees (CARTs) to predict tree cover from remotely sensed imagery (MODIS VCF) inherently introduces biases, thus making the resulting tree cover unsuitable for showing alternative stable states through tree cover frequency distribution analyses. We here provide a new and equally fundamental argument why the published frequency distributions should not be used for such purposes.We show that the practice of pre-average binning of tree cover values used to derive cover values to train the CART model will also introduce errors in the frequency distributions of the final product. We demonstrate that the frequency minima found at tree covers 8 % to 18 %; 33 % to 45 %; and 55 % to 75 % can be attributed to numerical biases introduced when training samples are derived from landscapes containing asymmetric tree cover distributions and/or a tree cover gradient. So it is highly likely that the CART, used to produce MODIS VCF, delivers tree cover frequency distributions that do not reflect the real world situation.2 Main body of text:The MODIS VCF tree cover product of Hansen et al. (2002b) provides worldwide estimates of percentage tree cover derived from MODIS data. Discontinuities in tree cover frequency distributions derived from this data have been used to support the hypothesis that the observed distribution of forest, savanna and grassland vegetation in the tropical and boreal regions of the world represent alternative stable states for equivalent environmental conditions (Hirota et al., 2011;Staver et al., 2011;Favier et al., 2012;Murphy & Bowman, 2012;Scheffer et al., 2012;Xu et al., 2016). But recently Hanan et al. (2014Hanan et al. ( , 2015 suggested that the adopted classification and regression trees (CART) approach, used to produce the MODIS VCF tree cover estimates, introduced a systematic bias which makes the MODIS VCF product inappropriate for the analysis of % tree cover frequency distributions. This point was counteredby Staver and Hansen (2015) arguing (i) that the approach taken by Hanan et al. (2014), using simulated EO data and pseudo satellite metrics to demonstrate that an artificial bias is generated by the CART approach, does not reflect the complexity and variability of landscapes and vegetation across the globe and (ii) that the CART model used by Hanan et al. (2014) was highly prun...

Alternative Vegetation States in Tropical Forests and Savannas: The Search for Consistent Signals in Diverse Remote Sensing Data

“…Metrics of spatial consistency in multimodality among data-streams based on Cohen's kappa (Table 2) indicate that VCF is distinct from LST and albedo (i.e., very little spatial consistency), with more agreement (but still considerable differences) in the locations of bimodality in the albedo and LST data streams. These results support concerns raised elsewhere [31][32][33] that the VCF product may not be appropriate for the diagnosis of alternative states (i.e., bifurcations) because of the calibration approach utilized and the discontinuities inherent in CART-based model-fitting approaches.…”

“…Further, we undertake a comprehensive analysis to identify evidence of multiple states across different data streams (VCF, NIR albedo, and LST) at the same geographic locations, noting that the independent (i.e., non-VCF) products are derived using fundamental physical principles with relatively little empirical model fitting, while VCF estimates rely on an empirical CART approach. We also note that while MODIS VCF is known to be unreliable for low tree cover values, distinctions of high and low tree cover may be more reliable [31][32][33]. Quantitative evidence for the presence or absence of unimodality (i.e., multimodality) can be inferred statistically using Hartigan's dip test [48] on a population of observations [49,50].…”

“…In these studies, potential spatial variability in ecological, edaphic, and anthropogenic factors within rainfall zones were mostly ignored. Further, the MODIS VCF retrieval was based on classification and regression tree (CART) algorithms and trained using pre-averaged binning techniques that have been linked to possible artifacts (artificial clumping of predictions) that could be wrongly interpreted as alternative states not present in reality [31][32][33]. Thus, the true extent and prevalence of bifurcating patch dynamics in the tropical savannas remains unknown.…”

Tropical climate–vegetation–fire relationships: multivariate evaluation of the land surface model JSBACH

“…globalfiredata.org/; last access: November 2013) provides globally gridded monthly burned area based on the MODIS sensor. We used the version 4 of the data set (Giglio et al, 2013).…”

Tropical tree cover in a heterogeneous environment: A reaction-diffusion model

“…In Africa, where there exist large areas of high tree cover savannas [27] and where fire occurrence is higher [11], bistability and cyclic behavior can be expected to play a larger role. A possibility other than the ones hitherto mentioned is that the observed bimodality is an artifact, resulting from data algorithms [28] or preprocessing methods [29]. Therefore, the multistability hypothesis should be tested on tree cover data produced with methods that are less likely to generate such artifacts.…”