Climate warming is increasing forest disturbances, with hotter summers and longer droughts causing widespread tree mortality. Yet the timing of these deaths remains unclear. While standing dead trees can be delineated from high‑resolution aerial imagery, their infrequent updates limit temporal analysis. Although satellite images have been used to map dead trees, few studies have estimated when trees died using time series data. We present a method to estimate mortality timing within known mortality areas using PlanetScope (3 m) time series in a boreal forest (Finland) and a temperate forest (Luxembourg). We used summer imagery (June–August) from 2021 to 2022 for Helsinki and 2020–2022 for Luxembourg. The reference dataset contained 468 standing dead trees in Helsinki and 3070 in Luxembourg, mapped from very high-resolution aerial (5–10 cm) and satellite images (50 cm). The Helsinki study area was characterized by scattered tree mortality with isolated and small clusters of dead trees, whereas the Luxembourg study area had more clustered tree mortality and larger clusters of dead trees. We evaluated four vegetation indices: kNDVI, GNDVI, SR 800/550, and PSSRc2, and used the Kernelized Change Point Detection (PELT) algorithm to identify sustained declines in pixel values. kNDVI performed best, detecting spectral change for 80% (2832 of 3538) of the dead trees across both areas, with detection increasing with cluster size. In Helsinki, kNDVI detected 204 dead trees (43.6%), while in Luxembourg, 2628 dead trees (86%). The differences in performance between the study areas were mainly attributed to the spatial distribution of tree mortality (scattered vs. clustered). Detection was most reliable for clusters of ≥ 3 trees, while isolated trees were rarely detected, likely due to mixed pixels (only 19% detected). For timing estimation, the overall RMSE across both areas was 245 days with a mean bias of + 6 days (i.e., six days later than visual confirmation). In Helsinki RMSE = 211, bias = − 103; Luxembourg RMSE = 254, bias = +40. The relatively high RMSE alongside a small overall bias indicates variability among individual timing estimates due to reference‑date lags and mixed pixels. With further improvements to change‑point detection and spectral inputs, PlanetScope time series show initial promise for estimating mortality timing in clusters (≥3 trees) of standing deadwood.
Nowak et al. (Mon,) studied this question.