Country-wide, high-resolution monitoring of forest browning with Sentinel-2

Samantha Biegel; David Brüggemann; Francesco Grossi; Michele Volpi; Konrad Schindler; Benjamin D. Stocker

stat.APApr 2, 2026

Country-wide, high-resolution monitoring of forest browning with Sentinel-2

Authors: Samantha Biegel, David Brüggemann, Francesco Grossi, Michele Volpi, Konrad Schindler, Benjamin D. Stocker

Organizations: Photogrammetry & Remote Sensing, ETH Zurich and ETH AI Center · Swiss Data Science Center, ETH Zürich and EPFL · Forest and Soil Ecology, Swiss Federal Institute for Forest, Snow and Landscape Research WSL, 8903 Birmensdorf, Switzerland · Institute of Geography, University of Bern, Hallerstrasse 12, 3012 Bern, Switzerland ·

Abstract

Natural and anthropogenic disturbances are impacting the health of forests worldwide. Monitoring forest disturbances at scale is important to inform conservation efforts. Here, we present a scalable approach for country-wide mapping of forest greenness anomalies at the 10 m resolution of Sentinel-2. Using relevant ecological and topographical context and an established representation of the vegetation cycle, we learn a predictive quantile model of the normalised difference vegetation index (NDVI) derived from Sentinel-2 data. The resulting expected seasonal cycles are used to detect NDVI anomalies across Switzerland between April 2017 and August 2025. Goodness-of-fit evaluations show that the conditional model explains 65% of the observed variations in the median seasonal cycle. The model consistently benefits from the local context information, particularly during the green-up period. The approach produces coherent spatial anomaly patterns and enables country-wide quantification of forest browning. Case studies with independent reference data from known events illustrate that the model reliably detects different types of disturbances.

Figures & tables

Variable	Source
NDVI	Swisstopo / Copernicus
Median Forest Height	Ginzler (2021)
Forest Mix Rate	Waser et al. (2025)
Tree Species	Koch et al. (2024)
Habitat Type	Price et al. (2024)
Elevation	SwissAlti3D DEM

Table 1: Variables used in the NDVI seasonal cycle model.

Figure 1: Exemplary predicted NDVI ranges and anomalies for 4 forest pixels. Green curves denote the upper quartiles ( $f_{0.75}(t)$ ), red curves denote the lower quartiles ( $f_{0.25}(t)$ ). Interquartile ranges are shaded in red. Dashed lines indicate the anomaly thresholds derived from those curves. NDVI observations that fall within the anomaly thresholds are coloured black, observations labelled as anomalies are coloured by their anomaly score, with lighter colour denoting more negative scores (larger deviations from the normal range).

Figure 2: Example anomaly map for forest pixels based on NDVI observations from 22/08/2023 and 24/08/2023 combined. A pixel is flagged if it is detected as anomalous on either of the days. The three panels on the left show a magnified view of the anomalies at the indicated regions on the map. Normal pixels are coloured in black and anomalous pixels are shaded in red according to their anomaly score.

Model	Pinball loss			D 2			Quantile coverage			MAE	RMSE	Bias	R 2
Model	q = 0.25	q = 0.5	q = 0.75	q = 0.25	q = 0.5	q = 0.75	q = 0.25	q = 0.5	q = 0.75	MAE	RMSE	Bias	R 2
Global baseline	0.081	0.080	0.048	0.0	0.0	0.0	0.248	0.500	0.752	0.16	0.23	0.08	-0.15
Climatology baseline	0.048	0.054	0.039	0.41	0.32	0.19	0.250	0.500	0.750	0.11	0.16	0.02	0.46
Conditional model	0.037	0.042	0.031	0.55	0.47	0.36	0.245	0.497	0.748	0.08	0.13	0.01	0.65

Table 2: Quantile model goodness of fit. The global baseline uses a single, country-wide empirical quantile. The climatology baseline uses a country-wide quantile per day of the year. MAE, RMSE, bias and R 2 are for per-day median values. “Bias” is the difference between the mean of model predictions and the mean of observations.

Figure 3: D ${}^{2}_{\text{pinball}}$ per day of the year for each of the predicted quartiles. The daily values are shown alongside the 7-day rolling mean. For each day of the year, the value of D ${}^{2}_{\text{pinball}}$ expresses the fraction of pinball loss explained when compared to a model that uses the empirical quantile at that day of the year (the climatology baseline ). A value of 0 corresponds to a model that predicts the same value at all locations per day of the year.

Figure 4: Daily fraction of forest pixels affected by negative NDVI anomalies. The negative anomaly fractions are computed per day of the year and smoothed with a 7-day rolling average. The dashed horizontal line indicates the overall fraction of anomalies detected at forest pixels. Fractions are relative to the number of valid observations at a pixel. The vertical lines indicate the season starts on March 1 (spring), June 1 (summer), September 1 (autumn) and December 1 (winter).

Figure 5: Fraction of observations per pixel affected by negative NDVI anomalies (“browning”) over the entire observation period. Fractions are relative to the number of valid observations at a pixel. The histogram on the right shows the distribution of anomaly fraction values across all pixels. The three panels on the left show a magnified view of the anomaly fractions at the indicated regions on the map.

Figure 6: Aggregated anomaly scores (top row) and fraction of anomalous pixels (bottom row) for each of the five cases (one case per column). Coloured curves denote the cases and dashed grey lines denote the reference areas. The vertical dashed lines indicate the date of the event for the fire and storm, the start and end of drought impact determined with the anomaly data for the drought event, and the start of the year for the clearcut harvest and beetle outbreak. In addition to the median, the 25th and 75th percentiles of the anomaly scores are shown by the shaded area.

Figure 7: Visualisation of forest embeddings, projected to two dimensions with $t$ -SNE. Points are labelled with their TypoCH habitat names and associated hierarchical codes, and coloured according to the highest hierarchy level.

Remote Sensing Change Detection

Remote sensing change detection finds and interprets differences across Earth-observation images acquired at different times.

See all 89 papers

ECHOSAT: Estimating Canopy Height Over Space And Time
Feb 24, 2026Jan Pauls, Karsten Schrödter, Sven Ligensa +7Remote Sensing Remote Sensing Change Detection
Department of Information Systems, University of Münster, Germany · Laboratoire des Sciences du Climat et de l’Environnement (LSCE), France · Zuse Institute Berlin (ZIB), Germany +2
A Two-Stage Cascade for Near-Real-Time Forest Anomaly Detection from Sentinel-1 SAR Time Series
Oct 2, 2026Pann Thinzar Seint, Subas Chhatkuli, Bryan AtwoodRemote Sensing Change Detection Time-Series Anomaly Detection
DAI Labs, K.K.
Reconstructing Multi-Decadal Forest Disturbances: A Spatio-Temporal Transformer Approach
Jun 5, 2026Linus Scheibenreif, Anton Raichuk, Maxim NeumannRemote Sensing Change Detection Remote Sensing
ETH Zurich Zurich, Switzerland · Google DeepMind Zurich, Switzerland

Feb 24, 2026J/K move · Enter open · S save

ECHOSAT: Estimating Canopy Height Over Space And Time

Jan Pauls, Karsten Schrödter, Sven Ligensa, Martin Schwartz+6

Department of Information Systems, University of Münster, Germany · Laboratoire des Sciences du Climat et de l’Environnement (LSCE), France · Zuse Institute Berlin (ZIB), Germany+2

Remote Sensing Remote Sensing Change Detection

Open paper AskPDF arXiv

Forest monitoring is critical for climate change mitigation. However, existing global tree height maps provide only static snapshots and do not capture temporal forest dynamics, which are essential for accurate carbon accounting. We introduce ECHOSAT, a global and temporally consistent tree height map at 10 m resolution spanning multiple years. To this end, we resort to multi-sensor satellite data to train a specialized vision transformer model, which performs pixel-level temporal regression. A self-supervised growth loss regularizes the predictions to follow growth curves that are in line with natural tree development, including gradual height increases over time, but also abrupt declines due to forest loss events such as fires. Our experimental evaluation shows that our model improves state-of-the-art accuracies in the context of single-year predictions. We also provide the first global-scale height map that accurately quantifies tree growth and disturbances over time. We expect ECHOSAT to advance global efforts in carbon monitoring and disturbance assessment. The maps can be accessed at https://github.com/ai4forest/echosat.

Abstract

Remote Sensing Change Detection

Explore similar work

ECHOSAT: Estimating Canopy Height Over Space And Time

A Two-Stage Cascade for Near-Real-Time Forest Anomaly Detection from Sentinel-1 SAR Time Series

Reconstructing Multi-Decadal Forest Disturbances: A Spatio-Temporal Transformer Approach