Increasing Accuracy in Analysis Ndvi-precipitation Relationship through Scaling down from Regional to Local Model

Spatial relationship between vegetation and rainfall in Central Kazakhstan has been modelled using Normalized Difference Vegetation Index (NDVI) and rainfall data from weather stations. The modelling based on application of two statistical approaches: conventional ordinary least squares (OLS) regression, and geographically weighted regression (GWR). The results support the assumption that the average impression provided by the OLS model may not accurately represent conditions locally. The OLS model applied to the whole study area was strong (R2 = 0.63), however it gave no local description of the relationship. Application of the OLS at the scale of individual land cover classes resulted in a better prediction power of the model (R2 = 0.75) and revealed that the response of vegetation to rainfall varies between the land cover classes. The GWR approach, dealing with spatial non-stationarity, significantly increases the model’s accuracy and prediction power (R2 = 0.97), as well as highlights local conditions within every land cover class. In order to compare the OLS and the GWR models in terms of prediction uncertainty, we calculated Moran’s I for their residuals. Our results demonstrated that the GWR provides a better solution to the problem of spatially autocorrelated errors in spatial modelling compared to the OLS modelling.