JingtaoLi/E2Forecast

E²Forecast checkpoints

Checkpoints for Forecasting Earth Surface Change in Embedding Space. Code, training and evaluation: https://github.com/ucam-eo/e2forecast.

file	experiment	embedding	context	selection	date offset
s1_embed.pt	S1_Embed forecasting	OLMoEarth-nano, 128-D	7 observations	last epoch (600)	order
s2_embed.pt	S2_Embed forecasting	Prithvi-EO-2.0-300M, 1024-D	37 observations	last epoch (72)	order
gsa.pt	GSA anomaly change detection	OLMoEarth-nano, 128-D	3 observations	last epoch (300)	order
inr.pt	INR, location → embedding (S2-100K)	SatCLIP, 256-D	none	best validation epoch (108)	–
inr_date.pt	INR with query date (SSL4EO-S12)	SatCLIP, 256-D	none	best validation epoch (288)	–

import torch
from e2forecast import load_pretrained, load_model

model, checkpoint = load_pretrained("s1_embed", torch.device("cuda"))         # download (cached) and load
model, checkpoint = load_model("s1_embed.pt", torch.device("cuda"))           # or a local file

Each file stores its model configuration (checkpoint["model_config"]), so the network is rebuilt from the file alone. checkpoint["date_offset"] selects the date offset of the history: year = year(τ*) − year(τ), order = −C, …, −1. Data: https://huggingface.co/datasets/JingtaoLi/E2Forecast-data.