Predicting fractional cover of standing deadwood at landscape level based on long short-term memory networks and Sentinel time series

Increasing climate extremes lead to the global phenomenon of increased tree mortality. Remote sensing provides great opportunities to track such dynamics. Effective and scalable approaches usually rely on multispectral data with moderate spatial resolution and high temporal resolution that are available at regional and global level (e.g., Landsat, Sentinel-2). There exist various remote sensing approaches that aim to track tree mortality by assessing vegetation status and changes, e.g., vegetation indices, classification, and time series analysis. Such products can either only inform on whether a stand is degraded (e.g., due to stress induced loss of foliage or vegetation health status) or the information on deadwood is only a single observation at a particular point in time. Until now, there is no detailed data product available that can explicitly inform on tree mortality over larger areas and multiple years. However, temporal and spatial information on tree mortality is of the uttermost importance, primarily to understand the extent and dynamics of this phenomenon and secondarily to understand the underlying mechanisms and environmental drivers. ... mehrThe lack of such products can be mostly attributed to reference data scarcity to build models that can extrapolate across large spatial scales and time. In this study, we attempt to close this gap using excessive amounts of Unoccupied Aerial Vehicle (UAV) imagery and a Convolutional Neural Network (CNN) for an automated way of reference data extraction. These automatically extracted reference data will then be used to predict the fractional cover of standing deadwood at Sentinel-2 (S2) pixel level using time series data and Long Short-Term Memory (LSTM) Neural Networks.
Our UAV data is spread over Central Europe and include heterogeneous temperate forests in Germany (Southern Black Forest, Black Forest National Park, Hainich National Park, Dresdner Heide, Hardtwald Karlsruhe, Bretten) and Helsinki, Finland. We used a U-net CNN for the segmentation of standing deadwood in the UAV-based RGB-imagery over approximately 470 ha of forest and achieved an F1 Score of 0.89 (from independent validation). Given the very high spatial resolution (smaller 2 cm) of the UAV-imagery we were able to create high resolution prediction maps of standing deadwood at UAV-scale. These high-resolution products then enabled us to precisely quantify the standing dead tree cover per S2-pixel. We accessed S2-data via the Data and Information Access Services (DIAS) platform and extracted time series of the four 10m-resolution spectral bands and the kernel NDVI (kNDVI). The time series cover a two-year period before the respective UAV acquisitions (2017-06-11 until 2021-10-17). Preprocessing included removal of pixels with kNDVI smaller than 0.075, missing-value interpolation (e.g., due to clouds), averaging to 7 day-intervals, and Savitzky-Golay filtering. The LSTM consisted of two bidirectional LSTM layers, each with 100 hidden units, followed by a fully connected layer with sigmoid activation that calculates the fractional cover of standing deadwood. The final model achieved R² = 0.65, RMSE = 0.169, and MAE = 0.123 in an independent validation.
Our results show that automatically extracted reference data from very-high resolution UAV-based RGB-imagery can be an effective source of training data for large-scale, multitemporal and satellite-based mapping applications. The results also show that with a high amount of detailed training data, accurately predicting standing deadwood cover at S2-pixel level is possible. With the LSTM Neural Network based on time series, the presented approach was not only able to extrapolate the results spatially but also temporally, thus enabling large-scale assessment of tree mortality dynamics over years. Such data products may not only be of great value to track forest loss, but also to understand its spatiotemporal patterns as a function of environmental drivers. Moreover, our study demonstrated that increasing open availability of high-resolution UAV-data will not only facilitate our capabilities for forest assessments on local scales, but also enhance our possibilities to exploit satellite missions for forest monitoring. Our findings therefore emphasize the importance of data sharing initiatives.