Retrieving plant functional traits through time series analysis of satellite observations using machine learning methods

Plant functional traits (e.g. leaf pigment and water contents, specific leaf area) serve as important indicators of plant condition, both within a given vegetation season and between years. Remote sensing-based methods allow for non-destructive and repeatable monitoring of the Earth's surface and thus offer an efficient way to map and monitor these traits. In our study, we used a large database of ground survey data sampled at several contrasting phenological phases of vegetation to develop and compare different machine learning models trained to estimate selected plant functional traits at two different sites: mixed floodplain forest at Lanzhot and beech forest at Stitna, both in the Czech Republic. Empirical models were trained as predictors using 1) Sentinel-2 satellite data (a data set with higher spatial and spectral resolution), and 2) Harmonized Landsat Sentinel-2 (HLS) product (a data set with higher temporal resolution). The most successfully retrieved traits were chlorophyll and carotenoid content (R-2 = 0.78 and 0.65, respectively). Although models trained with Sentinel-2 predictors proved to be slightly better in terms of validation statistics compared to HLS predictors, the HLS product may be preferable for some applications requiring analysis at a high frequency. The best-performing machine learning algorithm, canonical correlation forest, was then applied per pixel to all cloud-free images from the HLS product at both study sites for the years 2019-2021. This allowed us to create a time series of plant functional traits useful for observing differences between the two sites, as well as between growing seasons, and also to observe patterns of spatial behaviour using map outputs.