1Shkaruba A., PhD in geography, 2Kireyeu V., PhD in environmental sciences

1Central European University, 2Erda RTE

During the past decade, high resolution climate, elevation, land use, and vegetation datasets become available, as well as affordable and functional GIS tools. This brings forward in the ecosystem evaluation studies the methods based on statistical analysis with no or minimal contribution of expert judgements. In this study, we demonstrate such a method for the analysis of forest ecosystems’ exposure to climate change in Belarus. The study included: (1) identification of relevant terrain and climate variables, (2) development of methods for their integration, (3) running the classification, and (4) analysis of geographical shifts. On the stage of selecting relevant variables, a comprehensive analysis of characteristics directly or indirectly conditioning spatial differentiation of biophysical environments is needed. These characteristics should be quantitative and spatially distributed, have the same resolution, be in the same units, come from a reliable source, and have a significant impact on plant growing conditions in the scale of the study area. All the selected variables were grouped into two types: (1) accounting for the distribution of biophysical factors within the area and (2) influencing local re-distribution of these factors. The latter group included: elevation, slope angle, exposition, levels of groundwater table, and grain-size composition of underlying rocks. Elevations have been extracted from USGS 3 arc second (equal to c.a. 90 m in Belarus) SRTM dataset. This database has also been used to calculate a topographic index of hydrological similarity (also topographic wetness index, which accounts (on the scale) for the spatial variability of groundwater levels [1]. The grain-size composition of underlying rockswas  derived  from the 1:500 000 Map of Quaternary Deposits [2].

The central component of this methodological block is a dynamic and spatially explicit classification of climatic variables. The dynamic nature  of these variables calls for a methodological framework that allows for instant updates of resulting maps, sensitivity analysis of integration results, and a straightforward evaluation of uncertainties. The composition of climate datasets should reflect specific needs of the classification. In this study, we are exploring the potential effects of climate change on forests, and thereforethe selected variables also accounted for plant-growing conditions: growing degree  days (base 5°C and 10°C), growing season lengths (base 5°C and 10°C), annual averages of mean daily temperature, annual sums of mean daily precipitation, growing season precipitation, and hydrothermal coefficient (after Seljaninov).

All the climate datasets were constructed from data series from meteorological stations in Belarus and adjacent areas of neighbouring countries, and using Co-Kriging method. Six sets of interpolation surfaces have been made to capture how the variables were changing in each cell from  the  baseline decade (1950s) to 2000s. These gridded surfaces have been re-sampled to the same resolution (1 km) and combined into one multi-dimensional matrix. To remove redundancy and to reduce dimensionality, principal component analysis (PCA) has been applied to this dataset. Three 1st principal components (PCs) have further been used to combine grid cells into 30 clusters by applying MATLAB kmeans function. Grid cells have then been classified based on their cluster affiliation. This way, each cell would receive its six class ids, one per each decade.

The values of the 2nd PC, accounting mostly for precipitations and precipitation-related indices, have been classified into 4 groups by equal intervals. The 1st PC values, accounting for temperature-related variability, have been classified by the same method into 19 groups. Intersection of these groups of 1st and the 2nd PCs gave us 76 combinations; 30 of them corresponded to actual bioclimatic classes that occurred in Belarus in one of the six decades.

In this methodology, we assume that the bioclimatic conditions of the baseline decade in a cell are close to the equilibrium conditions for the ecosystems located within this cell. Thus, biophysical vulnerability of a given forest cell can be calculated as a function of shifting bioclimatic classes. To capture shifts in both temperature and humidity, each class received two ranking scores according to its 1st and 2ndPC group. To measure the shift in bioclimatic conditions, we can use the coefficient of linear regression (βpc1), giving us an average number of ranked classes that shifted through the cell per decade.We then combine 1st and 2nd PC shifts into the index of Biophysical vulnerability (Vb) by using the following equation:

where βpc1 is a regression coefficient for the 1st PC ranking values, and βpc2 for the 2nd PCones. Biophysical vulnerability index therefore shows the length of the bioclimatic shift in ranking scores per decade.


  1. Kirkby M.J., Weyman D.R. Measurements of contributing area in very small drainage basins // Seminar Series B, Department of Geography, University of Bristol. – 1974. – No. 3
  2. Геологическая карта четверстичных отложений Белорусской ССР. – Л.: Аэрогеология, 1978.