← Papers

Unverified paper record

Improving plant diversity prediction in revegetated grasslands using compact Sentinel-2 time series descriptors

Copernicus GmbH · 10 Mar 2026 · 10.5194/wbf2026-342

Abstract

Previous studies have shown that multitemporal data can strengthen the relationship between grassland diversity and spectral reflectance. However, most studies used interpolated Sentinel-2 time series as such. This study investigated whether species richness models can be improved using fitted Sentinel-2 temporal features describing the overall plant growth patterns.We measured plant species richness in 77 quadrats (1m x 1m) from revegetated alpine grasslands located around an open pit mine in Hochfilzen, Austria, characterized by a high range of species richness. Several multi-year Sentinel-2 vegetation index time series were used as inputs for fitting temporal features such as phenology descriptors, harmonic decomposition, frequency decomposition and functional principal components. Those features were compared to interpolated time series of Sentinel-2 bands used in state of the art baseline models (Fauvel et al., 2020; Muro et al., 2022).The feature sets were inserted into a Random Forest regression model pipeline, first selecting the best performing features in a nested cross-validation, then applying the final model over the grassland areas to produce species richness maps. Lastly, SHAP feature analysis was performed to improve model interpretability.Our best model, using fitted CIRE time series features, achieved coefficient of determination R2 = 0.19 in cross-validation and R2 = 0.36 on holdout set (16 quadrats). Features describing events around peak growing season were found especially important. Our model clearly outperformed all baseline models on holdout set across all metrics: R2 (+0.15 to +0.33), absolute Root Mean Squar Error RMSE (-0.35 to -0.91) and relative RMSE (-0.02 to -0.05).All models highlighted similar areas of high or low richness. Differences were observed for most extreme species richness, or less densely vegetated pixels. Results suggest our features are better suited to small datasets. The comparison of models will also be carried out on larger field inventories.Future steps will include extension to species abundance metrics, and a comparison to spectral variation features extracted from multi-scale hyperspectral imagery from Hyspex Mjölnir VS-620 drone and EnMAP satellite.This research is part of the MultiMiner project funded by the European Union’s Horizon Europe research and innovations actions programme under Grant Agreement No. 101091374.

Plant phenotyping relevance

Sentinel-2時系列から植物種多様性(種数)を推定する特徴抽出・機械学習パイプラインを開発・比較し、交差検証とホールドアウトで評価しているため、植物形質推定法が中心です。

abstractThis study investigated whether species richness models can be improved using fitted Sentinel-2 temporal features describing the overall plant growth patterns.
abstractThose features were compared to interpolated time series of Sentinel-2 bands used in state of the art baseline models
abstractThe feature sets were inserted into a Random Forest regression model pipeline, first selecting the best performing features in a nested cross-validation

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.

This is an automatically classified, unverified record. Curator approval is required before any resource enters the Catalog.