Reducing uncertainty in analysis of relationship between vegetation patterns and precipitation

Pavel Propastin 1, Nadiya Muratova 2 and Martin Kappas 1
1 Department of Geography, Georg-August-University
Goldschmidtstr. 3, 37077 Goettingen, Germany
Tel.: + 049 0551 389 487; Fax: + 049 0551
ppavel@gmx.de; mkappas@uni-goettingen.de 
2 Institute for Space Research and Earth Observation
Shevchenko Str. 15, 480100, Almaty, Kazakhstan
Tel.: + 007 3272 494 274; Fax: + 007 3272 918 077
nmuratova@mail.com

Abstract
The spatial relationship between vegetation patterns and rainfall and its trend over the period 1985-2001 in desert, semi-desert and steppe  grassland of the Middle Kazakhstan was investigated with Normalized Difference Vegetation Index (NDVI) images (1985-2001) derived from the Advanced Very High Resolution  Radiometer (AVHRR), and rainfall data from weather stations. The growing season relationship was examined using conventional, global, ordinary least squares (OLS) regression technique, and a local regression technique known as geographically weighted regression (GWR). Regression models between NDVI and precipitation for every  analysis year (1985-2001) were calculated using separately the both statistic approaches, the OLS and the GWR. The study found a presence of high spatial and temporal non-stationarity in the strength of relationships and regression parameters. The ordinary least squares regression model had been applied to the whole study area was superficially strong ( 2 R = 0.63), however it gave no local description of the relationship. Applying the OLS at the scale of the separate land cover classes revealed a different response of various vegetation types to rainfall within  the study area. The strength of the relationship between NDVI and rainfall increased in order from desert ( 2 R = 0.36), to semi-desert ( 2 R = 0.52), and to steppe grassland ( 2 R = 0.67). The approach of  geographically weighted regression provided considerably stronger relationships from the same data sets (mean value of 2 R = 0.88), as well as highlighted local variations within the land cover classes. Relationships between vegetation patterns and rainfall amounts are generally assumed to be spatially and temporally stationary. This assumption was not satisfied in this study. The study found that the relationship varied significantly  in space and time. In such circumstances the results provided by a global  regression model were uncertain and incorrect presented the relationship between the both variables locally. The application of local regression techniques such as GWR, may reveal local patterns of  relationship and significantly reduces the uncertainties of calculations. 

Keywords: NDVI, precipitation, regression modeling

In: Caetano, M. and Painho, M. (eds). Proceedings of the 7th International Symposium on Spatial Accuracy Assessment in Natural Resources and Environmental Sciences, 5 – 7 July 2006, Lisboa, Instituto Geográfico Português

AttachmentSize
Propastin2006accuracy.pdf860.23 KB