Statistical Inference for Spatiotemporal Trends in Remote-Sensing Data

Abstract:

Existing and future remote-sensing datasets provide unprecedented ability to detect broad-scale changes in the world through time. However, rigorously testing for patterns in these datasets requires a solid statistical foundation. To make predictions about future changes, it is necessary to test hypotheses about the causes of past changes and to separate predictive patterns in the data from non-predictive stochastic patterns. The non-predictive patterns often have temporal and spatial autocorrelation. Proper accounting for temporal and spatial autocorrelation has two benefits: it guards against false positives by correctly attributing patterns in the past to predictive variables, and it can decrease false negatives by increasing the statistical power of predictions. Our goal is to first explain the value of accounting for temporal and spatial autocorrelation when testing hypotheses with large, spatially rich datasets, and then to present a new statistical approach designed for remotely sensed data. For illustration, we analyze trends from 1981 to 2013 in annual cumulated values of NDVI at 8-km resolution for six continents. NDVI is a synoptic measure that integrates growing season length and peak productivity, and therefore we expected it to be influenced by factors that do not predict future changes in NDVI. Our statistical analyses of global NDVI data show the two benefits of accounting for temporal and spatial autocorrelation. First, there is little statistical support for latitudinal increases in greening trends in NDVI at continental scales (guarding against false positives). Second, there is strong statistical support that different land-cover classes show different latitudinal greening trends (guarding against false negatives). These results illustrate the value of formulating hypotheses for time-trend data that can – and have to be – statistically tested at broad scales.