Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates

Researchers have proposed two new frameworks for improving time series forecasting, one for predicting vegetation dynamics and another for autonomously generating forecasting algorithm code.

A team of researchers has developed a probabilistic forecasting framework for field-level Normalized Difference Vegetation Index (NDVI) prediction under sparse, irregular clear-sky acquisitions^[1]. The framework separates the encoding of historical NDVI and meteorological observations from future exogenous covariates, and uses a temporal-distance weighted quantile loss to address irregular revisit patterns and horizon-dependent uncertainty. Experiments on European satellite data showed that the proposed approach outperformed statistical, deep learning, and time-series baselines on both pointwise and probabilistic evaluation metrics. Meanwhile, another research team introduced a framework called SEA-TS, which autonomously generates, validates, and optimizes forecasting algorithm code for time series forecasting^[2]. SEA-TS uses a self-evolution loop to generate, validate, and optimize algorithm code, combining three mechanisms: MA-MCTS, code review with running prompt refinement, and global steerable reasoning. The framework outperformed strong baselines TimeMixer, Timer, and SEMixer in seven out of eight comparisons.