Prediction of post-disturbance forest recovery in the Alps – new paper published!

By Lisa Mandl (Technical University of Munich, Professorship of Earth Observation for Ecosystem Management)

Understanding forest recovery dynamics after disturbances is essential for effective conservation and management, especially as we face a future with more climate extremes. This is particularly relevant in mountainous landscapes like the Alps, where steep terrain and frequent extreme weather can hinder natural tree regeneration. Many existing studies use spectral indices to measure recovery, interpreting decreases in these indices as disturbances and subsequent increases as recovery. However, these indices don’t reveal which type of vegetation is driving the post-disturbance recovery. To address this, we propose using tree cover as a tangible and ecologically informed unit for assessing post-disturbance recovery.

We investigated post-disturbance forest recovery in the Alps using satellite remote sensing from Landsat and Sentinel-2, covering the period from 1990 to 2021. and applying a state-of-the-art technique called temporally generalized synthetic spectral unmixing, which allows to disentangle the land cover types within each pixel, giving us accurate tree cover fractions. We computed recovery intervals for both stand-replacing and non-stand-replacing disturbances using two recovery indicators: baseline-normalized and absolute recovery. Absolute recovery relies on the FAO definition of closed forests, requiring a minimum tree cover of 40%. Baseline-normalized recovery sets the threshold at 80% of the pre-disturbance tree cover, aligning with previous studies. We also calculated recovery intervals using spectral indices like NDVI and NBR to compare our findings with existing research. Additionally, we used bare ground shares, disturbance severity, and pre-disturbance tree cover shortly after disturbance events to predict long-term recovery trajectories. This approach allows us to assess recovery success for recent disturbances without needing long time series.

An Earth observation data cube spanning the year 1990 – 2021 and the area of the eastern Alps & surroundings (~ 130,000 km²). Each pixel contains information about its fractional cover for the classes trees, shrubs/grassland and bare soil.

For tree cover-based absolute recovery, the mean interval was 5.5 ± 0.03 years (mean and standard error) for non-stand-replacing and 13.4 ± 0.16 years for stand-replacing disturbances. Baseline-normalized tree cover recovery took 6.9 ± 0.04 years and 10.2 ± 0.64 years, respectively. Spectral index-based indicators showed faster recovery: NBR intervals were 3.5 ± 0.01 years for non-stand-replacing and 7.3 ± 0.36 years for stand-replacing disturbances, while NDVI intervals were 1.6 ± 0.006 years and 5.9 ± 0.32 years, respectively. Over the entire observation period of 32 years, nearly all disturbances have recovered, independent from the recovery indicator used. Within 10 years post-disturbance, 61% and 70% recovered by baseline-normalized and absolute recovery, respectively, while 83% recovered by NBR-based and 93% by NDVI-based recovery. Recovery based on spectral indices was thus far more rapid and overestimated the recovery success compared to recovery measured in terms of canopy cover.

Recovery intervals stratified by disturbance type (a) and percentage of recovered disturbances (b) for different indicators of recovered disturbances.

We finally predicted the long-term recovery success fitting logistic regression models based on pre- and early post-disturbance characteristics, that is pre-disturbance tree cover, disturbance severity and the bare ground share 3-years post-disturbance. For the absolute recovery indicator, the overall model accuracy was 83%; for baseline-normalized recovery we found an accuracy of 76%. We thus provide evidence that long-term recovery trajectories can be projected based on the signal from just a few years post-disturbance.

Mapped predictors used as input for the logistic regression model, which are translated into probabilities of recovery success (right). The column in the middle shows observed recovery success derived from the fractional cover maps. The site shows a disturbance patch in Crnivec (Slovenia), which was affected by a storm in 2008.

For more information check the full article published in Remote Sensing of Environment:

Mandl, L.; Viana-Soto, A.; Seidl, R.; Stritih, A.; Senf, C. (2024) Unmixing-based forest recovery indicators for predicting long-term recovery success.

Leave a Reply

Your email address will not be published. Required fields are marked *