Counting living carbon, from tree crowns to sediment archives

Trees that models cannot yet find everywhere

Counting trees from the air sounds solved until the model leaves its home forest. A preprint introduces MillionTrees, a global benchmark for airborne tree detection with over one million individual tree annotations and six million weakly supervised annotations from diverse sources. Current open-source tree-detection models do not generalize across the benchmark's diverse datasets, and overfitting to test data is common, with hyperparameter tuning identified as a critical challenge for zero-shot performance. Training on the MillionTrees dataset improves both within-domain and out-of-domain performance, which suggests the benchmark's value is less in ranking today's models than in exposing how much their accuracy depends on the forest, sensor, and annotation style they were tuned for. For carbon accounting, that exposure matters: a tree count that is accurate in one region and unreliable in another will bias any total built from it. m:…9739

Lidar fills where optical estimates saturate

A second preprint shows how combining sensors changes a continental carbon estimate. Integrating GEDI lidar observations into a data-assimilation system reduces woody biomass uncertainty by 63.7 percent relative to unconstrained models, and it resolves saturation biases in high-biomass regions where optical estimates of aboveground biomass fail. Data quality, cloud coverage, and elevation remain the primary drivers of spatial error, which means the improvement is real but uneven: the system is most uncertain exactly where observations are hardest to get. The work is limited to North America, and scaling to other biomes remains to be validated, so its continental number should not be quoted as a global one. The method, though, is the transferable part: harmonizing repeated optical and lidar observations iteratively, rather than trusting either sensor alone, gives a carbon estimate whose uncertainty can be stated and, over time, shrunk. m:…7925

Hidden diversity below the canopy

Not all ecosystem counting happens from above. A preprint from Maludam National Park in Borneo reports that tropical peatlands harbor Arcellinida communities, a group of testate amoebae, whose composition differs fundamentally from extratropical ecosystems, with species richness significantly higher in peatland zones where understory vegetation is developed. The correlation with vegetation-hydrological zones suggests these microscopic communities track the structure of their habitat closely enough to serve as indicators of peatland condition. The study is limited to a single national park and lacks functional metabolic data for the identified taxa, so it cannot yet say what these communities do, only that they differ and where they are richest. Even with that limit, it widens the counting problem: an ecosystem assessment that counts trees and carbon but ignores microbial diversity will miss changes that only the smallest residents record first. m:…0822

Sediments separate two centuries of pressure

A final preprint reaches back two centuries to separate pressures that present-day surveys see only mixed together. Using ancient environmental DNA metabarcoding of sediment records, the authors distinguish the ecological impacts of climate change and industrial whaling in the Southern Ocean. Industrial whaling in the twentieth century is negatively associated with the abundance of metazoan families that feed from the water column, indicating a loss of pelagic productivity, and biodiversity change in the Western Antarctic Peninsula has accelerated in recent decades compared with the pre-industrial baseline. The study rests on five sediment records, so broader spatial validation is still needed. Its contribution is methodological as much as historical: archives that record both human extraction and climate change can, with the right analysis, attribute change to each, rather than leaving every decline to be argued over. Across all four preprints, the pattern is the same. Better ecosystem numbers come not from one perfect sensor, but from combining sensors and archives whose blind spots differ, and from stating those blind spots plainly. m:…7210