All models of satellite-derived phenology are wrong, but some are useful: A case study from northern Australia

Satellite-derived phenology (or apparent phenology) is frequently used to illustrate changes in plant phenology (i.e. true phenology) and the effects of climate forcing. However, each study uses a different method to detect phenology. Plant phenology refers to the relationship between the life cycle of plants and weather and climate events. Phenology is often studied in the field, but recently studies have transitioned towards using satellite images to monitor phenology at the plot, country, and continental scales. The problem with this approach is that there is an ever-increasing variety of earth observation satellites collecting data with different spatial, spectral, and temporal characteristics. In this paper we ask if studies that detect phenology using different sensors over the same site produce comparable results. Mangrove forests are one example where different methods have been used to examine their apparent phenology. In general, plant phenology, including mangroves, is described using few individual plants, but continental-scale descriptions of phenological events are scarce or inexistent. Few attempts have been made to describe the phenology of mangroves using satellite imagery, and each study presents a different method. We hypothesize that apparent phenology changes with: 1) areal extent; 2) site location; 3) frequency of observation; 4) spatial resolution; 5) temporal coverage; and 6) the number of cloud contaminated observations. Intuitively, one would assume that these hypotheses hold true, yet few studies have investigated this. For example, one would expect that clouds change the observed phenology of vegetation, that the number of species captured at spatial resolution will impact the apparent phenology, or that mangroves in different places display different phenologies, but how are these changes represented in the apparent phenology? We use the Enhanced Vegetation Index (EVI) to examine the changes in the start of season and peak growing season dates, as well as the shape and amplitude of the apparent phenology in each hypothesis. We use Landsat and Sentinel 2 imagery over the mangrove forests in Darwin Harbour (Northern Territory, Australia) as a case study, and found that apparent phenology does change with the sensor, site, and cloud contamination. Importantly, the apparent phenology is comparable between Landsat and Sentinel 2 sensors, but it is not comparable to phenology derived from MODIS. This is due to differences in the spatial resolution of the sensors. Cloud contamination also significantly changes the apparent phenology of vegetation. In this paper we expose the complexity of modelling phenology with remote sensing and help guide future phenology investigations.

Automated summary

Satellite-derived phenology is commonly used to study changes in plant phenology and the impact of climate change, but the variety of earth observation satellites, each with different characteristics, can lead to discrepancies in apparent phenology results. Factors such as cloud contamination and spatial resolution can significantly affect the observed phenology of vegetation.