Reconstructing Landsat reflectance time-series for assessing seasonal changes in Finnish lakes

Finnish lakes are spread across a large latitudinal gradient, and are extremely diverse in terms of trophic state, size and depth. Monitoring seasonal changes in the productivity of these lakes is therefore particularly interesting for assessing the impacts of climate change. Although remote sensing (RS) is a promising tool for monitoring lakes seasonality, the complex shapes, and often small size of Finnish lakes, require the use of high spatial resolution sensors. However, in RS, high spatial resolution often comes at the expense of coarser temporal resolutions. For instance, Landsat (LT) imagery have appropriate spatial resolution for monitoring Finnish lakes, but the low temporal resolution hinders the acquisition of images in cloudy seasons. As a result, data gaps in LT time-series makes it difficult to assess seasonal patterns of reflectance signals. The objective of this study was to apply time series analysis to reconstruct seasonal patterns in 35 years of LT data from a Finnish lake. Our study area was the Lake Köyliönjärvi, in SW Finland. We focused on assessing intra-annual changes between April and October (warm months), given that during the remaining months (cold months) the lake is often frozen and LT imagery have high frequency of cloud coverage. First, we extracted the lake's surface reflectance using every image from the LT 4, 5 and 7 archives. In total, 960 images from 1982 to 2016 were considered. Pixels contaminated by cloud, shadow or ice, were removed using standard masking algorithms (fmask). For this study, we analyzed the band ratio: B2(green)/B1(blue), which, according to previous studies, has a good relationship with Chlorophyll-a (Chl-a) concentration. Next, we fill the values from the cold months with a baseline value. Finally, we use a Kalman seasonal filter for filling the gaps in the warm months. Our approach could successfully retrieve the seasonal patterns during the warm months, showing a significant relationship (p<0.05) with field measurements of Chl-a. Further research will focus in optimizing the relationship between remote sensing data and phytoplankton dynamics, by testing other LT bands and spectral indices, as well as multi-variate models. Ultimately, the method will be applied for assessing long term changes in the phytoplankton dynamics of lakes across northern countries.