Carbon, Measured From Orbit, Ocean, and Forest

From orbit: tracking change at power plants

In a preprint, a ConvGRU deep learning model classifies power-plant NOx emission trajectories from TEMPO satellite imagery with a macro-F1 score of 0.56, outperforming a multilayer perceptron baseline score of 0.45. A macro-F1 of 0.56 is not a claim of near-perfect detection; it is a measured improvement over the baseline on a hard classification task, and the gap between the two scores is the result worth noting. The baseline matters because a score without a comparator says little; here the comparison is explicit, and the ConvGRU clears it. m:…0546

The preprint's framing choice is arguably more important than the score. Change detection is a more tractable framework for modeling power-plant activity from TEMPO observations than estimating absolute emissions. Satellites see plants well enough to say that something changed, and when, more reliably than they can say exactly how much was emitted in absolute terms. Building monitoring systems around change — which plants ramped up, which went quiet — plays to what the observations actually support. m:…0546

From the ocean: floats that accidentally trap carbon

A second preprint finds a carbon measurement in an unlikely instrument. Argo floats, deployed to profile the ocean, act as serendipitous particle traps, capturing sinking carbon particles during deep-sea drift. The floats were not designed as sediment traps, but their drift through the deep sea leaves them collecting the very particles whose sinking carries carbon out of the surface ocean. m:…8514

From that accidental archive, the preprint estimates the global sinking carbon flux at 1,000 meters depth at 790 teragrams of carbon per year, with a 95 percent confidence interval of 470 to 1,120 teragrams per year. The interval is wide — the upper bound is more than double the lower — and that width is part of the finding: the deep-ocean carbon sink is large, but pinning it down to a narrow number from existing observations is not yet possible. The estimate's value lies in giving the flux a measured range grounded in a global, repurposed dataset rather than in a single model run. m:…8514

From the forest: two decades of loss, accelerating

The third measurement comes from land. Ghana lost 1.75 million hectares of tree cover between 2001 and 2024, representing 25.1 percent of its 2000 canopy cover, according to a preprint mapping two decades of regional loss. A quarter of a country's canopy in under a quarter-century is a structural change, not an annual fluctuation, and it reframes Ghana's forest from a standing stock to a shrinking one. m:…2235

The loss is also accelerating and concentrated. Tree cover loss in Ghana increased by 9 percent between the 2015-2019 and 2020-2024 periods, and the Western and Central regions experienced the highest percentage of their own tree cover loss, at 32.0 and 30.3 percent respectively. Regional concentration matters for any response: a national average hides where the loss actually happens, and the preprint's hotspot framing points protection and restoration toward the regions losing the largest share of their own forests. Across orbit, ocean, and forest, the three preprints share a discipline — report what the observations support, state the uncertainty, and let change, not just totals, carry the story. The period comparison is the acceleration signal: the later five-year window lost more than the earlier one, rather than holding steady. m:…2235