Reliability of resilience estimation based on multi-instrument time series

Many widely used observational data sets are comprised of several overlapping instrument records. While data inter-calibration techniques often yield continuous and reliable data for trend analysis, less attention is generally paid to maintaining higher-order statistics such as variance and autocorrelation. A growing body of work uses these metrics to quantify the stability or resilience of a system under study and potentially to anticipate an approaching critical transition in the system. Exploring the degree to which changes in resilience indicators such as the variance or autocorrelation can be attributed to non-stationary characteristics of the measurement process - rather than actual changes in the dynamical properties of the system - is important in this context. In this work we use both synthetic and empirical data to explore how changes in the noise structure of a data set are propagated into the commonly used resilience metrics lag-one autocorrelation and variance. We focus on examples from remotely sensed vegetation indicators such as vegetation optical depth and the normalized difference vegetation index from different satellite sources. We find that time series resulting from mixing signals from sensors with varied uncertainties and covering overlapping time spans can lead to biases in inferred resilience changes. These biases are typically more pronounced when resilience metrics are aggregated (for example, by land-cover type or region), whereas estimates for individual time series remain reliable at reasonable sensor signal-to-noise ratios. Our work provides guidelines for the treatment and aggregation of multi-instrument data in studies of critical transitions and resilience.