Tree fall onto railway lines puts passengers at risk and causes large economic losses due to disruption of train services and damage to infrastructure. Railway lines in Germany are vulnerable to tree fall because of the large number of trackside trees that exist in that country with approximately 70% of all railway lines being tree-lined. In this paper we first tested whether a hybrid-mechanistic tree wind damage model, ForestGALES, could identify the sections of the railway network affected by tree fall in two federal states of Germany, Northrhine-Westphalia (NRW) and Thuringia (TH). We secondly tested whether the model, in combination with meteorological forecast models, could predict where tree fall occurred during a damaging windstorm. We used information on tree characteristics derived from LiDAR and aerial photography along the railway line network in NRW and TH to calculate the critical wind speed (CWS) at which damage is expected to happen for every individual tree as a function of its size and species, and the underlying soil. The railway network was then divided into 500 m sections and the statistics of the CWS, tree height, and species composition (broadleaf/conifer mix) within each section were calculated. Analysis of past tree fall events recorded by Deutsche Bahn AG (DB) showed that there was a significantly lower minimum CWS and significantly greater maximum tree height in sections that had recorded damage. In a second step we compared the calculated CWS values for all trees against downscaled wind speed assessments across the two federal states during Storm Friederike (named Storm David internationally) on 18 January 2018 and tested the ability of the model to discriminate sections with recorded damage during the storm. Excellent model discrimination was found with an AUC value of 0.82 and an overall model accuracy of 74.2%. The first test showed that the ForestGALES model with precise individual tree information can identify the sections of a railway network most vulnerable to tree fall. The second analysis showed, for the one storm tested, that the ForestGALES model when combined with predicted storm wind speeds can identify the most probable sections of the railway network to experience tree fall during an approaching damaging storm. Such information could be of value in firstly planning remedial work along railway lines, and secondly preparing the railway network ahead of a major storm.

https://www.dzsf.bund.de/SharedDocs/Downloads/DZSF/Veroeffentlichungen/Forschungsberichte/2023/ForBe_41_2023_SensVeg.pdf?__blob=publicationFile&v=8

Concerted use of CNN + UAV can close the gap in reference data scarcity Multitemporal, large-scale maps of standing deadwood with high spatial resolution (10 m) Best model with all S1 + S2 bands R² = 0.38 after 5-fold CV Model slope: y = 0.45x + 0.26 Error equally distributed across range Optimization of co-registration improved results Robust across heterogeneous landscape and years % standing deadwood 2 bidirectional LSTM layer 100 LSTM units each Sampled to equal distribution Long short-term memory (LSTM) network UAV-derived reference data Sentinel-2 Level-2A • BOA (sen2cor) • Cloud filter < 70% • 11 bands + kNDVI Sentinel-1 Level-2 • Terrain corrected backscatter (CARD-BS) • Interferometric coherence (CARD-COH6) 2-years time series Linear interpolation of missing values Aggregation to 7-day intervals Satellite time series Step 2: Mapping fractional cover of standing deadwood from Sentinel time series (landscape level) delineation CNN prediction U-net CNN architecture Independent, pixel-based evaluation: F1-Score = 0.89

update umweltmonitoring Potsdam von 1992-2016 Aktuelle Methoden und Erfassungen Neue Daten 2016 Trends zu Nutzung, Biotopwerten, Grünvolumen, Versiegelung und Hinweis zu Klimaanpassung http://10dfns.ioer.info/fileadmin/user_upload/10dfns/files/16_5_2018/21_3_Frick.pdf