Late July: foundation models, phenology manuscripts, and clearing the decks

A short research note rather than a full week diary. The main news this month was the Science of Remote Sensing paper; this is where the rest of the programme sits.

Foundation models → ecology

The Trentino work was the first ecological application I led with TESSERA embeddings. The useful lesson was not just accuracy: with a few percent of the training parcels you are already near the performance ceiling, so the scarce resource is good reference data, not more spectral feature engineering.

I want the next steps to stay ecological rather than purely methodological — transferring these embeddings to habitat and community maps, and asking how crown- and plot-level traits can become training signal at landscape scale.

Tropical phenology and species mapping

Two first-author phenology manuscripts from Paracou are close: one on canopy green-up preceding the light season, and one linking leaf-cycle frequency (not deciduousness alone) to growth. Alongside those, the hyperspectral / multi-temporal species-mapping strand continues — including a new preprint on phenological regularity and spectral separability, and an updated version of the hyperdiverse species-mapping preprint.

The through-line is the same as detectree2: make individual-tree and community processes observable beyond the plot, then use that to train models that scale.

Software and open tools

detectree2 remains the main open package I maintain. On the foundation-model side, the Cambridge EO group’s GeoTessera tooling is the practical entry point for embeddings. TESSERA itself is heading to CVPR 2026.

Site housekeeping

I finally added this Notes section to the research site and caught the homepage up with the SRS paper, the phenology preprint, and TESSERA’s CVPR framing. Expect occasional notes when something ships or a research strand shifts — not a weekly cadence for its own sake.