Predicting plant diversity in revegetated grasslands with Sentinel-2: comparing performance of spatio-temporal features with input time series

Abstract. Mining companies are continuously looking for cost efficient methods to monitor the success of their rehabilitation efforts. Although open access satellite imagery is available at regular temporal intervals, its usefulness for grassland biodiversity monitoring has been questioned due to its coarse spatial resolution with respect to the species size. To compensate for the low spatial resolution, previous studies have successfully explored the benefits of using a multitemporal set of Sentinel-2 (S2) images. However, unless the temporal patterns are studied as a whole, some of the phenological information such as growth rates are lost, and delayed snow cover may spread events like growth onset over multiple dates between plots. This study aims to explore the added value of temporal fitting of Sentinel-2 time series (ts) over existing baseline models applied using the full time series as such. Our set of temporal features included functional components, harmonic decomposition, frequency decomposition, and phenological metrics. Out of the compared models, the Random Forest regression model using a set of fitted temporal features achieved the highest holdout prediction accuracy (R2 = 0.36, RMSE = 3.87, relative RMSE = 0.20) and cross-validation accuracy similar to the baseline models. However, all the compared regression models underestimated extreme plant diversity to some extent. Future studies should account for varying vegetation cover and terrain features by incorporating auxiliary data.