UAV-based reference data for the prediction of fractional cover of standing deadwood from Sentinel time series

Abstract

Increasing tree mortality due to climate change has been observed globally. Remote sensing is a suitable means for detecting tree mortality and has been proven effective for the assessment of abrupt and large-scale stand-replacing disturbances, such as those caused by windthrow, clear-cut harvesting, or wildfire. Non-stand replacing tree mortality events (e.g., due to drought) are more difficult to detect with satellite data – especially across regions and forest types. A common limitation for this is the availability of spatially explicit reference data. To address this issue, we propose an automated generation of reference data using uncrewed aerial vehicles (UAV) and deep learning-based pattern recognition. In this study, we used convolutional neural networks (CNN) to semantically segment crowns of standing dead trees from 176 UAV-based very high-resolution (<4 cm) RGB-orthomosaics that we acquired over six regions in Germany and Finland between 2017 and 2021. The local-level CNN-predictions were then extrapolated to landscape-level using Sentinel-1 (i.e., backscatter and interferometric coherence), Sentinel-2 time series, and long short term memory networks (LSTM) to predict the cover fraction of standing deadwood per Sentinel-pixel. The CNN-based segmentation of standing deadwood from UAV imagery was accurate (F1-score = 0.85) and consistent across the different study sites and years. Best results for the LSTM-based extrapolation of fractional cover of standing deadwood using Sentinel-1 and -2 time series were achieved using all available Sentinel-1 and --2 bands, kernel normalized difference vegetation index (kNDVI), and normalized difference water index (NDWI) (Pearson’s r = 0.66, total least squares regression slope = 1.58). The landscape-level predictions showed high spatial detail and were transferable across regions and years. Our results highlight the effectiveness of deep learning-based algorithms for an automated and rapid generation of reference data for large areas using UAV imagery. Potential for improving the presented upscaling approach was found particularly in ensuring the spatial and temporal consistency of the two data sources (e.g., co-registration of very high-resolution UAV data and medium resolution satellite data). The increasing availability of publicly available UAV imagery on sharing platforms combined with automated and transferable deep learning-based mapping algorithms will further increase the potential of such multi-scale approaches.