Sensitivity of EVI-based harmonic regression to temporal resolution in the lower Okavango Delta

Abstract

In this study, we examined how satellite time-series-based characterization of ecological cycles and trends is sensitive to the temporal depth and spacing of the time series and whether the observed sensitivities were cover and/or cycle specific. We fitted a harmonic regression (with annual, semi-annual and quasi-decadal cycles) to an 85-image Landsat time series (1989–2002) covering the lower Okavango Delta, varying the temporal depth and spacing of time-series vectors following two different but comparable approaches (i.e. systematic vs random variation). The results show that as the temporal depth decreases, the sensitivity to both short- and long-term ecological cycles was lost in the seasonally dynamic environment. The degree to which characterization of ecological cycles and trends was influenced by temporal depth and spacing was dependent on the functional type of vegetation and dynamics of the disturbance regime(s). Ecological cycles were comparatively better characterized, even at reduced temporal depths, for woodland-dominated areas, riparian vegetation and more frequently flooded areas (i.e. less dynamic systems) compared with less routinely and non-flooded grasslands, mixed savannas and more frequently burned areas. Detection sensitivity was also found to be cycle specific, as, at reduced temporal depths, the semi-annual cycle was less easily detected than the annual cycle for the majority of the study area. Temporal spacing of time-series vectors was also found to be important: the chances of misinterpreting system variability as a change were much higher when vectors were randomly chosen. Interestingly, at a reduced temporal depth, the harmonic model unexpectedly produced a higher coefficient of determination due to model under-specification for both vector selection approaches. However, randomly selected vectors produced much larger artefacts. These results indicate the need for caution in selecting a time series not only for its temporal depth but also for its temporal spacing, particularly as related to the inherent environmental periodicity of the study area.