Reproduce “SPEAR NeXT Causal Latent Forecasting Across Multiple Horizons for Spectral Temporal Earth Representation Learning” on OpenResearch

About this paper

Rajiv Ranjan, Udaiveer Singh, Shashank Tamaskar, Dharmendra Saraswat

SPEAR-NeXT extends spectral Earth-observation representations with a causal Transformer that directly forecasts multiple future latent states from past pixel observations. Using frozen 32-dimensional SPEAR embeddings, it improves several land-cover, agricultural, and crop-yield benchmarks, including a reported mean crop-wise yield R² of 0.721 and land-cover F1 scores of 94.79% in India and 90.66% in CONUS, although detailed forecasting and ablation results remain incomplete.

Key insights

• Forecasting future latent representations, rather than reconstructing raw observations, encourages embeddings to retain information about how Earth-surface states evolve over time.
• Combining relative temporal order with month and year information helps distinguish elapsed time from seasonal phase, while inverse-horizon weighting balances near-term and longer-range prediction.
• The strongest reported gains occur in temporally structured regression, especially crop phenology and yield prediction; classification improvements are more uneven across datasets and adaptation settings.

Summary by alphaXiv.