Ocean color remote sensing data can provide a temporal and spatial view of the phytoplankton biomass distribution for long data series. However, there is also a need to develop quantitative methods for identifying and interpreting the fluctuations in the biomass distribution over the duration of the time series [45
]. The present study at Sagres has developed the stl.fit() as a tool to provide an analysis of ocean color time series data from this region, to identify the historical changes over time and to provide a richer diagnosis and clearer interpretation of the temporal variability and its potential causes. The study has focused on the inter-annual variation of the changes of the phytoplankton to better understand the dynamics of the coastal waters off Sagres.
4.1. The stl.fit() Approach
The characterization the changes that occur in the Sagres region is based on the STL decomposition method proposed by Cleveland et al.
], but instead of considering the standard stl(), where the seasonal smoothing parameter is assumed to be periodic, stl.fit() has been developed to obtain “automatically” the seasonal and trend smoothing parameters by minimizing an error measure that enables optimal modeling of the data. This approach is an iterative computational estimate of the seasonal and trend components, where the trend component is obtained using the non-parametric loess regression and the seasonality from the trend adjusted series.
Another feature that is considered by STL is its robustness to extreme and other aberrant observations. There are studies (e.g., Vantrepotte and Mélin [3
], Vantrepotte and Mélin [4
], Vantrepotte and Mélin (2009) [46
], and Volpe et al.
]) that remove these observations from the data, whereas STL solve this situation without reducing the dataset. Moreover, an analysis has been performed that detects the dominant components in the time series, using a statistical dispersion measure that is not influenced by outliers. Thus, the empirical results based on an error measure indicates that the best model is achieved using the stl.fit(), instead of the standard approach. This is important because the initial objective of the study was to obtain a statistical model that can better describe the stochastic behavior of a time series, enabling a statistical analysis that will improve the diagnosis and interpretation of the changes in the ocean color products retrieved by the satellite data. In summary, stl.fit() is capable of capturing the variations in the seasonal dynamics and takes advantage of the useful features of the STL time series decomposition method. Nevertheless, stl.fit() also inherits the disadvantages of STL, i.e.
, that it does not handle calendar variations and only considers an additive decomposition [12
4.2. MERIS Time Series
Using around 10 years of MERIS ρw
(λ) and API 1 products, this study has evaluated the performance of stl.fit() for the decomposition of time series for ocean color satellite data. The decomposition (Figure 6
and Figure 7
) shows that the data can be split into seasonal, trend and irregular components. The decomposition has detected the strength of the time series components, which were deduced from the percentage of data represented within the IQR. Station A shows a seasonal dominance for all the reflectance wavebands, which is a similar finding to that observed by Vantrepotte and Mélin [3
] for the normalized water leaving radiances from SeaWiFS for the Iberian area. For the other two stations, the model is not capable of modeling the trend and the seasonality, probably due to the high variability in the data.
When the seasonal components of the ρw
(at 443, 490, 510 and 560 nm) and the API 1 product are compared, a relationship is observed between them, as might be expected as the visible wavelengths of ρw
(λ) are part of the API 1 algorithm. However, in Figure 6
d, the seasonal effect is less variable (for 510 and 560 nm) and seems to decrease over time (for 560 nm) compared to 443 and 490 nm. The lower variability in the green bands is probably because the reflectance is not strongly influenced by the absorption of phytoplankton, which is the dominant optical constituent within this region [24
]. However, the higher seasonal variability at the start of the time series is also seen in the MERIS API 1 product, which could indicate a change in the dominant phytoplankton functional types, i.e.
, higher values in the green bands indicate higher scattering that could be coming from the phytoplankton. The decline in the concentration of phytoplankton (TChla
, retrieved by API 1), together with the simultaneous decrease in seasonal amplitude, is consistent with the findings of Vantrepotte and Mélin [4
] for the Northeast Atlantic area.
The MERIS API 1 product is dominated mainly by the seasonal component, with an increase in dominance as the distance increases offshore from Station A to Station C. As in Vantrepotte and Mélin [4
], this site shows an inter-annual variability that is dominated by the phytoplankton seasonal variability that is linked to the physical oceanography influencing the study area. The main source of nutrients for the phytoplankton depends on the seasonal pattern of upwelling events [25
]; other sources of nutrients are limited as there are no permanent rivers [29
Although the difference between the seasonal and the irregular components is not significant, the dominance of the irregular component is slightly higher for the coastal Station (A). This can probably be explained by an increased failure of the atmospheric correction due to the adjacency effects from land, which are discussed in more detail in Cristina et al.
]. Therefore, these results suggest that future studies should try processing the time series with the Improved Contrast between Ocean and Land (ICOL) processor; to reduce errors in the processing chain for sites close to land by correcting for the adjacency effects.
An example of how stl.fit() can contribute to the analysis of the MERIS time series at the Sagres site is provided be comparing plots of the inter-annual variability for the seasonal component of the MERIS API 1 at the Sagres Stations (Figure 8
). Essentially, these seasonal plots show the months with the highest variability at each Station. These data provide important information about periods when management action might be needed, for example, for aquaculture activities, or when in situ
sampling should be implemented for monitoring purposes, e.g., to conform to EU Directives. In the case of Sagres the data shows that each Station should be sampled at least once per season, to show changes in the ecosystem. In particular Station A and B should be sampled in spring (from March and April), in summer (during June), in autumn (from September and November) and in the winter (during December). Alternatively, Station C should be sampled in spring (from March to May), in summer (from June to August), in autumn (from September to November) and in the winter (from December to February).
With regard to the two research questions presented in the Introduction Section:
Will stl.fit() capture the dynamics of the time series of the study area better?
How can stl.fit() be used to describe and explain the variability of the study area?
The final section of both the Results and the Discussion provide an example showing the inter-annual variability of the seasonal component of MERIS API 1 (see Figure 8
). Many other examples could be provided.