Sequence-based mapping approach to spatio-temporal snow patterns from MODIS time-series applied to Scotland

Highlights

• National scale spatio-temporal snow patterns mapping.

• Improved methods for high cloud cover and irregular spatio-temporal snow presence.

• Sequence-based approach to describe time series patterns of snow cover.

• Good agreement with validation data of cloud filling method.

• Good caption of temporal patterns by indices and good relationship with landscape.

Abstract

Snow cover and its monitoring are important because of the impact on important environmental variables, hydrological circulation and ecosystem services. For regional snow cover mapping and monitoring, the MODIS satellite sensors are particularly appealing. However cloud presence is an important limiting factor. This study addressed the problem of cloud cover for time-series in a boreal-Atlantic region where melting and re-covering of snow often do not follow the usual alpine-like patterns. A key requirement in this context was to apply improved methods to deal with the high cloud cover and the irregular spatio-temporal snow occurrence, through exploitation of space-time correlation of pixel values. The information contained in snow presence sequences was then used to derive summary indices to describe the time series patterns. Finally it was tested whether the derived indices can be considered an accurate summary of the snow presence data by establishing and evaluating their statistical relations with morphology and the landscape. The proposed cloud filling method had a good agreement (between 80 and 99%) with validation data even with a large number of pixels missing. The sequence analysis algorithm proposed takes into account the position of the states to fully consider the temporal dimension, i.e. the order in which a certain state appears in an image sequence compared to its neighbourhoods. The indices that were derived from the sequence of snow presence proved useful for describing the general spatio-temporal patterns of snow in Scotland as they were well related (more than 60% of explained deviance) with environmental information such as morphology supporting their use as a summary of snow patterns over time. The use of the derived indices is an advantage because of data reduction, easier interpretability and capture of sequence position-wise information (e.g. importance of short term fall/melt cycles). The derived seven clusters took into account the temporal patterns of the snow presence and they were well separated both spatially and according to the snow patterns and the environmental information. In conclusion, the use of sequences proved useful for analysing different spatio-temporal patterns of snow that could be related to other environmental information to characterize snow regimes regions in Scotland and to be integrated with ground measures for further hydrological and climatological analysis as baseline data for climate change models.

Introduction

Snow cover is one of the most important climatic factors on Earth affecting surface albedo, energy balance and hydrological circulation. Monitoring snow cover is therefore essential because both the duration and depth of snow cover affect important environmental variables and ecosystem services with implications for human activities and biodiversity conservation over large areas of the planet. The characterization of snow and ice cover patterns is critical for understanding the Earth's water and energy cycles. Snow cover has the greatest seasonal variation in spatial extent of any component of the cryosphere. Therefore, accumulation and rapid melt are two of the most important seasonal environmental changes of any kind on the Earths surface (Frei et al., 2012, Rittger et al., 2013).

Water stored in the snow pack represents an important component of the hydrological balance in many regions of the world, especially in mountain and boreal regions.The pattern of snowfall and melt affects the timing of river flooding which can impact on e.g. habitat suitability of fresh water organisms or the generation of hydro-electric power. On land, plant and animal communities are adapted to different patterns of snow-lie, and several economic sectors rely on the predictability of snow cover, among these the ski industry, the farming and forestry sectors, and the transport sector. Climate-change adds extra complexity, because existing patterns of snow-lie, together with all of the above-mentioned, are likely to be affected by changes in temperature and precipitation.

While monitoring of snow cover is needed, given the vast areas in question, the point data collected at meteorological stations are relatively sparse and provide poor representation of different types of landscapes and topography. Snow cover mapping has been utilized in operational snow-melt, runoff forecasting, data assimilation and the calibration or validation of various hydrological models (Andreadis and Lettenmaier, 2006, Bloeschl et al., 1991, Grayson et al., 2002, Parajka and Bloschl, 2008, Parajka and Bloschl, 2008b). Maps based on ground data can be expected to have high uncertainty when produced for large spatial areas such as whole river catchments.

For regional snow cover mapping, the MODIS satellite sensors are particularly appealing due to their daily temporal resolution and medium spatial resolution of about 500 m. Potential applications include operational snow-melt runoff forecasting, data assimilation and the calibration or validation of hydrologic models (Parajka and Bloschl, 2008b).

Snow extent (i.e. presence or absence of snow, regardless of snow amount) is relatively straightforward to observe using visible observations because of the high albedo of snow compared to most land surfaces. However there are some limitations: (1) visible imagery is limited to the portion of the surface illuminated by sunlight; (2) cloud presence; (3) vegetation can obstruct visible and infrared information about snow from reaching the satellite sensor; e.g. forest canopies protrude above the snow pack, lowering the surface albedo and partially or completely masking the underlying surface; (4) surface heterogeneity can play an important role in the interpretation of visible and infrared imagery, affecting the quality of the resulting products.

Many studies (e.g. Hall and Riggs, 2007, Liang et al., 2008, etc.) have demonstrated that MODIS daily snow cover products in clear-sky conditions give quite good agreement with ground based observations or other satellite-based snow cover products. However high cloud cover has been found to be a real problem in using MODIS snow cover products for various applications (Ault et al., 2006, Bitner et al., 2002, Klein and Barnett, 2003, Lopez et al., 2008, Parajka and Bloschl, 2008, Simic et al., 2004, Tekeli et al., 2005, Zhou et al., 2005, Wang and Xie, 2009).

Maps of the snow extent can provide useful information to support modelling. Assimilation of MODIS snow cover data into a hydrological forecasting model was found to slightly improve the estimates of both spatial extent and temporal evolution of snow cover (Parajka and Bloschl, 2008b). However, the spatial and temporal information contained in a time series of snow extent maps may be further exploited. A possible approach is to consider the time series as a sequence of states. The primary objective of sequence methods is to extract simplified workable information from sequential data sets, i.e. to efficiently summarize and render these sets and to categorize the sequential patterns into a limited number of groups. A common approach for categorizing patterns consists of computing pairwise distances between them by means of sequence alignment algorithms (such as optimal matching) or other suitable metrics and using this information for clustering the sequences (Abbott and Tsay, 2000). A recent complementary approach considered in the literature (Elzinga and Liefbroer, 2007, Widmer and Ritschard, 2009) is to focus on sequence indices measuring for instance the longitudinal diversity and complexity of the sequences and to analyse them by means of conventional statistical tools for real-value variables. The sequence approach is often applied in genomics, social science (e.g. Studer et al., 2011) and survival analysis (e.g. Mills, 2011). However the applications to temporal series of environmental spatial data are so far limited. The use of frequence approach provides summary indices that can be more easily interpreted and used in further modelling, as the index is summarising in a single value information otherwise contained in long sequences, not immediately interpretable.

The aim of this paper was to characterise the spatio-temporal patterns of snow cover in an Atlantic-Boreal region (applied here to Scotland). A key requirement in this context was to apply improved methods to deal with the high cloud cover and the irregular spatio-temporal snow occurrence, through exploitation of space-time correlation of pixel values. The information contained in snow presence sequences was then used to derive summary indices to describe the time series patterns. Finally it was tested whether the derived indices can be considered an accurate summary of the snow presence data by establishing and evaluating their statistical relations with morphology and the landscape.