Potential of Sentinel-1 C-Band Time Series to Derive Structural Parameters of Temperate Deciduous Forests

Abstract

With the increasing occurrence of forest fires in the mid-latitudes and the alpine region, fire risk assessments become important in these regions. Fuel assessments involve the collection of information on forest structure as, e.g., the stand height or the stand density. The potential of airborne laser scanning (ALS) to provide accurate forest structure information has been demonstrated in several studies. Yet, flight acquisitions at the state level are carried out in intervals of typically five to ten years in Central Europe, which often makes the information outdated. The Sentinel-1 (S-1) synthetic aperture radar mission provides freely accessible earth observation (EO) data with short revisit times of 6 days. Forest structure information derived from this data source could, therefore, be used to update the respective ALS descriptors. In our study, we investigated the potential of S-1 time series to derive stand height and fractional cover, which is a measure of the stand density, over a temperate deciduous forest in Austria. A random forest (RF) model was used for this task, which was trained using ALS-derived forest structure parameters from 2018. The comparison of the estimated mean stand height from S-1 time series with the ALS derived stand height shows a root mean square error (RMSE) of 4.76 m and a bias of 0.09 m on a 100 m cell size, while fractional cover can be retrieved with an RMSE of 0.08 and a bias of 0.0. However, the predictions reveal a tendency to underestimate stand height and fractional cover for high-growing stands and dense areas, respectively. The stratified selection of the training set, which we investigated in order to achieve a more homogeneous distribution of the metrics for training, mitigates the underestimation tendency to some degree, yet, cannot fully eliminate it. We subsequently applied the trained model to S-1 time series of 2017 and 2019, respectively. The computed difference between the predictions suggests that large decreases in the forest height structure in this two-year interval become apparent from our RF-model, while inter-annual forest growth cannot be measured. The spatial patterns of the predicted forest height, however, are similar for both years (Pearson’s R = 0.89). Therefore, we consider that S-1 time series in combination with machine learning techniques can be applied for the derivation of forest structure information in an operational way.