Implications of JPEG2000 lossy compression on multiple regression modelling

Multiple regression is a common technique used when performing digital analysis on images to obtain continuous, quantitative, variables (as temperature, biomass, etc). In this scenario pixels are treated as samples from which several independent variables are known; when remote sensing images are available, the different spectral bands offered by a given sensor are often used as independent variables. The dependent variable is also a quantitative variable, such as a forest inventory variable or a climate variable (e.g., temperature). This paper presents an evaluation of the implications of JPEG2000 lossy compression when applied to these regression processes. Annual minimum and annual mean air temperature are modelled using two methods according to the independent variables used: only geographical, and geographical and remote sensing images as independent variables. Raster matrix representing independent variables were compressed using compression ratios from 50% up to 0.01% of the original file size. Results have revealed that, even at high compression ratios, practically the same accuracy, measured with independent reference climatic stations, is obtained, so JPEG2000 seems an interesting technique not heavily affecting these common modelling approaches.