Assessing the detectability of standing dead trees in PlanetScope imagery

Standing dead trees (SDTs) play a critical role in forest biodiversity assessment, disturbance monitoring, and carbon cycling. High-resolution optical satellite imagery offers an opportunity for large-scale SDT detection by providing finer spatial detail than moderate-resolution sensors. Despite this advantage, isolated SDTs may occupy only part of a PlanetScope pixel, allowing their signals to mix with those of live canopy, non-woody vegetation, and shadows. As a result, it is necessary to quantify signal detectability and detection stability across these conditions. To address this issue, this study uses PlanetScope imagery at 3 m resolution and Uncrewed Aerial Vehicle (UAV)-derived reference annotations to develop a multi-level, response-based SDT framework for analyzing SDT detectability from the aspect of the segmentation models, SDT density, and SDT patch sizes. UAV annotations are first aggregated to the PlanetScope pixel scale to generate hard labels and soft labels. The deep and machine learning models are then trained to select a stable reference semantic model and assess the consistency of SDT detection responses across architectures. All deep learning segmentation models produced consistent detectability responses, with HRNet achieving the best performance, with a binary F1 of 0.86. A single-pixel SDT signal became evident mainly when SDT cover proportion reached the 70 %-80 %, whereas neighborhood support lowered this change to a local SDT density of 20 %–30 %. Patch-level analysis indicated more stable detection for SDT patches larger than 90–180 m 2 . Furthermore, reliable PlanetScope detection is associated with SDT dominance, spatial aggregation, and high local SDT density . The proposed framework provides a detectability analysis approach for evaluating the stability of SDT detection in 3 m satellite imagery and the consistency of these responses across semantic segmentation models, linking high-resolution satellite imagery probability mapping with ultra-high-resolution UAV-based observations.