Training-Free Monitoring of Kenya with Pre-computed Satellite Embeddings: Lake Expansion, Construction Dating and Crop Retrieval with TESSERA

Abdul Majeed Mohamed Fathir

Pre-computed embeddings from geospatial foundation models turn the most expensive step of satellite analysis into shared data. We test how much analysis this enables for one country, using TESSERA, which publishes a 128-dimensional embedding for every 10 m pixel and year. An audit of Kenya shows that all 4,766 tiles centred in the country carry all nine years (2017–2025), but that the landmask removes the sea and not inland water. We then map Lake Baringo, a Rift Valley lake that has flooded its shores, by assigning each pixel to the nearer of two reference embeddings (open water and urban land), with no training. The lake grew from 181.3 km² in 2017 to 223.0 km² in 2025 (+23%), 73% of the newly flooded land went under water in 2020–2021, and our 2020 area lies 5% below an independent field-checked Landsat estimate. Classified independently each year, 67% of the changing area switches label exactly once. On the same embeddings, a single weighted change point per pixel dates 92% of changed pixels along the Nairobi–Naivasha railway to its 2018–2019 construction, while a control line opened in 2017 shows 2 changed pixels out of about 9,000; a similarity search seeded with one tea field recovers 59% of the pixels of 51 held-out tea fields with almost no detections in other landscapes; and a linear probe agrees with ESA WorldCover on 68.6% of pixels in held-out regions. Code and notebooks are released under the MIT licence.