Detection of linear trends in multisensor time series in the presence of autocorrelated noise: Application to the chlorophyll‐a SeaWiFS and MERIS data sets and extrapolation to the incoming Sentinel 3‐OLCI mission

Abstract

[1] The detection of long-term trends in geophysical time series is a key issue in climate change studies. This detection is affected by many factors: the size of the trend to be detected, the length of the available data sets, and the noise properties. Although the noise autocorrelation observed in geophysical time series does not bias the trend estimate, it affects the estimation of its uncertainty and consequently the ability to detect, or not, a significant trend. Ignoring the noise autocorrelation level typically leads to an overdetection of significant trends. Due to satellite lifetime, usually between 5 and 10 years, sea surface time series do not cover the same period and are acquired by different sensors with different characteristics. These differences lead to unknown level shifts (biases) between the data sets, which affect the trend detection. In this work, we develop a generic framework to detect and evaluate linear trends and level shifts in multisensor time series of satellite chlorophyll-a concentrations, as provided by the Medium Resolution Imaging Spectrometer instrument (MERIS) and sea-viewing wide field-of-view sensor (SeaWiFS) ocean-color missions. We also discuss the optimization of the observation networks, in terms of needed time overlap between successive time series to reduce the uncertainty on the detection of long-term trends. For the incoming Sentinel 3-Ocean and Land Color Instrument (3-OLCI) mission that should be launched at the end of 2014, we propose a global map of the number of months of observations to enhance the trend detection performed with the joint SeaWiFS-MERIS analysis.

Variability and 20-Year Trends in Satellite-Derived Surface Chlorophyll Concentrations in Large Marine Ecosystems around South and Western Central America

“…Two types of time series were generated at each pixel, each using data from a different sensor for the overlapping period The analysis of trends was done at all pixels jointly for the SWF and AQ time series following the methodology developed in Saulquin et al (2013). This method considers the noise autocorrelation in the time series, which affects the estimation of the uncertainty in the trend estimate and consequently the ability to detect a significant trend.…”

“…Only trends satisfying the 95% detection threshold are considered in the analyses. More details on the numerical resolution of the equations can be found in Tiao et al (1990) and Saulquin et al (2013).…”

“…Trend analyses of global sea surface temperature (SST) indicate mean increases of 0.71 • C century −1 since 1900 (Wu et al, 2012) and between 0.09 • ± 0.03 and 0.18 • ± 0.04 • C decade −1 since the 1980s (Lawrence et al, 2004;Good et al, 2007). Longterm trends in chlorophyll concentrations are variable, with areas of increasing and decreasing chlorophyll around the globe (Gregg et al, 2005;Demarcq, 2009;Saulquin et al, 2013;Boyce et al, 2014;Gregg and Rousseaux, 2014;O'Brien et al, 2017). The factors driving these changes in environmental conditions are not fully understood, and although climate driven changes can have a global effect (i.e., increased SST and more intense stratification), other factors have variable effects on regional scales.…”

Impact of inter-mission differences and drifts on chlorophyll-atrend estimates

“…In any case, the objective of the study was not to derive accurate estimates of actual chl-a trends in the ocean, for which more advanced statistical methods are desirable (e.g. Beaugrand, Ibañez, and Lindley 2003;Henson and Thomas 2007;Saulquin et al 2013), but to illustrate how significant bias or drift can be for trend detection. This being said, the trend map (Figure 1(b)) obtained for the period 1998-2012 with the reference merged series (combining SeaWiFS and bias-corrected MODIS data) is very coherent with the results of Gregg and Roussseaux (2014) who also used a bias correction approach and data assimilation in a biogeochemical model (see their Figure (6)).…”

“…Behrenfeld et al 2006;Martinez et al 2009;Mélin 2009, 2011) but the inclusion of other subsequent missions in the overall data record is now necessary to extend the temporal basis of such analyses. Several investigations for local or global applications actually used the SeaWiFS record with data from other missions (McClain, Signorini, and Christian 2004;Djavidnia 2009, Mélin et al 2011;Kahru et al 2012;Bélanger, Babin, and Tremblay 2013;Coppini et al 2013;Saulquin et al 2013;Gregg and Casey 2010;Gregg and Roussseaux 2014;Park et al 2015;Signorini, Franz, and McClain 2015), and various merging techniques were proposed to combine data sets from multiple missions (Kwiatkowska and Fargion 2003;Maritorena and Siegel 2005;Pottier et al 2006;Mélin et al 2011;IOCCG 2007).…”