A new GIS landscape classification method for rain/snow temperature thresholds in surface based models

Precipitation falling as rain or snow can affect many different aspects of life such as transportation safety on roads, strength or thickness of sea ice (Lundberg & Feiccabrino 2009), runoff in rivers, animal migration patterns (Stenseth et al. 2002), or water storage in reservoirs for electricity production, drinking, or recreation. Surface based models for these purposes use a precipitation phase determination scheme (PPDS) to assign precipitation to a liquid (rain) or solid (snow) phase. Many PPDS apply a surface air temperature threshold (T_A) (e.g. USACE 1956), surface dew point temperature threshold (T_D) (e.g. Marks et al. 2013), or a surface wet bulb temperature threshold (T_W) (e.g. Matsuo et al. 1981).

A PPDS scheme often uses either: (1) a single temperature threshold (T_RS) where all precipitation events occurring in temperatures equal to or colder than a T_RS are classified as snow, while all events occurring in warmer temperatures are classified as rain; or (2) a dual temperature threshold scheme using a rain (T_R) and snow (T_S) temperature threshold with all precipitation occurring in temperatures equal to or colder than T_S classified as snow, all events occurring in temperatures equal to and warmer than T_R classified as rain, and for precipitation in temperatures between T_S and T_R a decreasing snow fraction (SF) from T_S = 100% to T_R = 0% is used to assign the precipitation phase. The advantage of a dual threshold scheme is that it can account for sleet occurring in temperatures near a T_RS, however in some areas sleet is rarely reported. For PPDS purposes, the T_RS or T_S and T_R are set to the temperature value/s resulting in the least amount of misclassified precipitation. These threshold values are often applied over large areas irrespective of terrain, ocean, or seasonal influences on different landscapes. These and other methods, along with meteorological reasoning for their use, are reviewed in Feiccabrino et al. (2015).

Surface based PPDS methods do not account for microphysical processes, e.g. evaporation, sublimation, freezing and melting (Thériault & Stewart 2010), which are very calculation intensive, and important for assigning a precipitation phase in atmospheric models. The calculations of microphysical processes require detailed lower atmospheric measurements not often brought into surface (hydrological) models (Harder & Pomeroy 2013). Instead, surface models rely on the assumption of constant atmospheric conditions, such as a set decrease in air temperature with height (e.g. the CHRM model using −7.5 °C/km lapse rate) (Fang et al. 2013).

Assuming constant atmospheric conditions over a large area with varying landscapes is not ideal when, for example, a snowpack (open water) conductively cools (warms) near surface air. The atmospheric depth to which land or ocean temperatures have an effect is generally restricted to a couple hundred meters, and this depth can increase or decrease depending on the amount of time an air-mass is over the cooling (heating) source, wind speeds, or other factors. Terrain effects, for example orographically enhanced precipitation, will increase the depth of the atmosphere required to melt falling snow (the melting layer) toward the ground due to increased heat energy required to melt the additional snow. This results in a lowering of the 0 °C isotherm and snow elevations compared to upwind locations (Minder et al. 2011). Therefore, increasing the precipitation rate would cause changes from the surface model assumption of constant atmospheric conditions by lowering cloud layers, snow elevations, and the 0 °C isotherm. The degree to which orographic enhancement occurs is dependent on the geometry of the terrain, and melting layer properties (Minder et al. 2011), for example, relative humidity (RH). On the leeside of mountains, air descends drier, and warms at or near the dry adiabatic lapse rate (9.8 °C/km) causing another change from normal atmospheric conditions. These terrain effects are known to occur on the individual to hill-slope scale.

Dai (2008), in a global study, found a warm bias of +0.7 °C for T_A over ocean regions (1.9 °C) compared to land regions (1.2 °C). The warm bias over oceans is most likely caused by open water transferring heat to the near surface atmosphere. The added surface warmth increases atmospheric instability (Ólafsson & Haraldsdóttir 2003), which by definition increases the environmental lapse rate, or actual decrease in air temperature with height, compared to the model assumed constant lapse rate. This decreases the depth of an atmospheric warm layer from the expected depth, thus increasing the chance of snow not fully melting before reaching the ground at a given surface air temperature. Stewart (1992) observed that precipitation phase transition often occurred near the coast, supporting the idea of ocean heat affecting lower atmospheric air temperatures which affects the precipitation phase in near coastal locations.

Topography and warm ocean temperatures are therefore known to affect atmospheric conditions above individual locations which should result in different T_A values. However, it is not known if terrain effects can be projected to larger geographical landscapes such as the leeward (LW) side of a mountain range rather than being restricted to the individual hill-slope scale. If station T_A values within a geographical landscape are found to be similar, and differ from values in other landscapes, then geographical regions (countries) could be broken into sub-regions based on landscape categorization to reduce misclassified precipitation error.

This study has two main goals: (1) present a landscape classification technique using geographic information system (GIS) to (a) automate geographic landscape classification, (b) decrease the subjectivity from manual classification found in the earlier conference paper by Feiccabrino (2015), and (c) determine if changes in T_A values for landscape (sub-regional) can be used to decrease error without downscaling to the individual hill-slope scale; and (2) using the same weather stations as the earlier paper, decrease the temperature resolution from 0.5 to 0.1 °C for a comparison of both (a) the misclassified precipitation percentage resulting from PPDS schemes using T_A, T_D, and T_W and (b) the misclassified precipitation as a result of manual and automated landscape classification.

The weather data are free and publicly available from the Swedish Meteorological and Hydrological Institute (SMHI 2016) and the Norwegian Meteorological Institute (NMI 2016) websites. A description of the data and further station information are found in Feiccabrino (2015). In Feiccabrino (2015), the weather stations were manually assigned to the geographic landscapes; WW ocean, WW coast, WW fjords, WW hills, and WW mountain (valleys) in Norway, and LW mountain (valleys), LW hills, LW rolling, and LW Coast in Sweden. The 2015 study found that using different T_A values for geographic landscapes is a viable option for decreasing misclassified precipitation error in regional studies. It also found T_W to be 0.5 °C in all landscapes (at a 0.5 °C scale), which resulted in less misclassified precipitation than T_D and T_A for eight of nine landscapes. However, dew point, wet-bulb, and RH measurements are much rarer than surface air temperature at hydrological and meteorological sites (e.g. Kane & Stuefer 2015). Therefore it is important to find techniques to decrease misclassified precipitation using surface air temperature, a commonly reported value.

An initial test applying slight modifications to Dai's (2008) method with different landscape T_A for WW ocean, WW coast, WW fjords, inland, and LW coast stations was found to decrease error from country values (Table 1 and Figure 2). However, better landscape definitions could be used to further decrease misclassified precipitation, especially for inland areas.

Geographic coordinates and associated tabular data for each meteorological station were imported into ArcGIS (version 10.2.1) and initially displayed using the WGS 1984 geographic coordinate system. In order to perform spatial analysis, the data were reprojected in a UTM projected coordinate system (ETRS 1989 – TM32). Elevation/hillshade and slope base maps were created using a European digital elevation model (GTOP30) with a 30-arc second (ca. 1 km) resolution. This layer was produced by the USGS EROS Data Center (USGS 1996) and was obtained through ArcGIS online. ArcGIS was used to classify the stations based on landscape features and also to produce a series of maps that highlight the spatial distribution of the results.

The classification of meteorological stations using ArcGIS involved three sets of landscape-based parameters. The first parameter in the classification was used to approximate the water (ocean) or land influence on a weather station. Ice-free water in the winter would normally be a surface based heat source that affects temperature thresholds. Buffer zones of 15 km were created surrounding each station and then intersected with the land outline for Norway and Sweden. The area of land contained within each of the intersected buffer zones was quantified using the calculate geometry function. The following categories were established based on these results: <10% land = ocean; 10–60% land = coast; 60–90% land = fjord (inlet); >90% land = land (Table 2). These numbers were chosen for their ability to separate the near coastal stations into distinct landscapes with different relative influences of land and ocean. Some smaller islands off the main shore were categorized as ocean stations which could be separated out if another landscape category was added, or the 10% land requirement for coast was decreased.

The land sites were further categorized based on the range of elevation that fell within the 15 km buffer zones. This relief value should indicate the amount of orographic enhancement (Minder et al. 2011) or down-sloping (LW) influence on lower atmospheric conditions over an area, therefore providing a reasonable reflection of terrain affects at a higher spatial resolution than the regional scale without having to consider wind datum or more detailed topographic data. The process of calculating the relief value involved extracting the digital elevation model (DEM) that fell within the buffer zones and then converting the extracted DEM into a vector (polygon) feature class. The polygon data were then summarized by station number under the fields of maximum and minimum values. A new table was created during the summarization and the range (max–min elevations) was then calculated for each site buffer using the field calculator. Finally, the new table with range was joined to the original buffer layer. Now each site could be categorized by the range of elevation within each buffer zone. The following classifications were made based on these data: range of 1,500–1,000 m = mountain; range of 1,000–500 m = hill; range of 500–0 m = rolling (Table 2). The values for these classifications were chosen to be both objective with a range of 500 m each, and representative of the amount of relief in the area. The values can be expanded or retracted depending on the topography of a study area, or have extra categories added.

The third and final parameter used in the classification was WW versus LW based on the prevailing direction of westerly flow and the location of the Scandinavian Mountains.

Datasets for precipitation events with surface air temperatures of −3 to 5 °C were built for each country, landscape, and individual weather station to determine T_A, T_D, and T_W values.

The results in Figure 3 and Table 3 indicate that the proposed GIS method is able to replicate and/or improve the reduction in misclassified precipitation from manual landscape classification. This landscape classification was done to embrace terrain (Minder et al. 2011) and ocean effects (Dai 2008) in the lower atmosphere which are ignored when using a global or regional T_A value, while not downscaling to the individual/hill-slope scale.

The landscape T_A values agreed well with Dai (2008). The warmest values were over the ocean and (WW) Norwegian coast (1.8 and 1.7 °C respectively). However, one surprise was the LW hill stations which have a T_A of 1.6 °C. This could be due to down-sloping effects leading to a deeper dry air layer between the cloud and ground. In a dry layer the air could be assumed to warm at between 9.8 and 7.5 °C/km (Fang et al. 2013). Depending on the atmospheric conditions, snow could fall through a couple hundred meters of dry air without completely melting (Venne et al. 1997). The coolest T_A values were in the lowlands near the shores indicating that the near surface heating of the ocean does not affect surface or lower atmospheric temperatures far inland (Table 3). Interestingly, T_A values for inland and mountains are warmer (1.2 and 1.3 °C with LW hills at 1.6 °C) than WW fjord and LW coastal stations (1.0 and 1.1 °C respectively). T_RS value fluctuations for individual stations within hill and mountain landscapes suggest much of the differences occur on the hill-slope scale as local terrain and winds differ greatly compared to lowland flat terrain. In the winter, oceans cause relatively greater lower atmospheric instability than land due to conductive heating at the ground atmosphere interface leading to warmer T_RS values over open water (Ólafsson & Haraldsdóttir 2003). Applying the same logic, one would expect snow covered ground to increase lower atmospheric stability due to cooling at the ground atmosphere interface leading to cooler T_RS values over snow covered ground. However, Feiccabrino (2015) did not find any notable variation in T_A values for either land with and without snow cover, or ocean stations with varying sea surface temperatures.

The GIS landscape classification method was both able to add objectivity to a previous, more subjective, manual classification with similar decreases in misclassified precipitation, indicating that it is a viable method and should be tested in a surface hydrological model.

The GIS landscape T_A results indicate that decreasing the geographical scale from regional/country to sub-regional/landscape reduces misclassified precipitation by 2.29% of the possible 7.60% when using individual station thresholds. The landscapes show ocean stations to have the warmest T_A values while low relief land stations have the coolest. Generally warmer T_A values were found in mountain and hill landscapes than for the inland rolling hill or LW coast stations.

Over land, high variability in individual station T_A values were found in mountain and hill landscapes. This indicates that while sub-regional T_A values based on a 15 km GIS landscape identification scheme improved precipitation phase determination, there may be further benefit from a smaller scale GIS analysis modified to identify upslope, downslope, or valley areas within high relief terrain.

Ocean and land effects on T_A were able to advect with prevailing atmospheric winds over coastal stations, most notably for the WW coast and WW ocean landscapes with T_A = 1.7 and 1.8 °C respectively and LW rolling, LW coast landscapes with T_A = 1.3 and 1.1 °C respectively. The ocean effect did not push further than 15 km inland as fjord/inlet stations had a T_A value of 1.0 °C which is more representative of low relief land stations.

GIS landscape T_W (8.26%) reduced misclassified precipitation from regional T_A by more than three times as much as landscape T_A values (2.29%). Wet-bulb thresholds are superior to T_A for point locations, however many stations do not report the required information to calculate wet-bulb. Therefore, the use of T_W is highly recommended if the information exists, or a surface wet-bulb layer is verified to project well through a model layer.