SITS-DECO: A Generative Decoder Is All You Need For Multitask Satellite Image Time Series Modelling

TL;DR

This work introduces SITS-DECO, a decoder-only, GPT-style model that treats diverse Earth Observation tasks as unified sequences of continuous and symbolic tokens, enabling multi-task learning and promptable inference without task-specific heads or encoders. By focusing on dense multi-temporal satellite time series, the authors demonstrate that a lightweight, data-centric approach can outperform several larger EO foundation models on crop-type classification, highlighting the importance of temporal modelling over architectural complexity. The framework supports self-supervised objectives and regression tasks within the same token-sequence paradigm, enabling flexible domain adaptation and minimal preprocessing. Despite not matching full spatio-temporal state-of-the-art, SITS-DECO provides a practical, extensible path toward generative EO foundation models and emphasizes data diversity as a primary driver of capability. The work also delineates clear future directions for scaling data, integrating language, and incorporating spatial context to realize true, zero-adaptation EO foundation modelling.

Abstract

Earth Observation (EO) Foundation Modelling (FM) holds great promise for simplifying and improving the use of EO data for diverse real-world tasks. However, most existing models require additional adaptation before they can be used and are structured rigidly around particular data sources or training approaches. To address this, we take inspiration from large language models, where diverse tasks, both pre-training and downstream, are implicitly captured through next-token prediction over unified token sequences, leveraging the structure and diversity of the training data. We introduce SITS-DECO (Satellite Image Time Series-DECoder Only), a proof-of-concept generative model that applies this unified-sequence framing to EO data. Using a simple GPT-style decoder-only architecture, and demonstrate its ability to perform useful EO tasks (pixel-wise, multi-temporal, multi-modal crop-type classification) in a purely generative framework. Through symbolic prompting, we show that the model can perform multiple supervised and self-supervised tasks within a single unified architecture, without task- or modality-specific adaptation. Despite its simplicity and lack of spatial context, SITS-DECO outperforms much larger EO foundation models on crop-type classification (PASTIS-R) demonstrating that dense temporal sequence modelling is a critical missing ingredient in the current paradigm. This work exemplifies a data-centric modelling paradigm in which capability arises from the diversity and structure of the training data rather than from architectural complexity. SITS-DECO provides a lightweight, practical route to multi-modal, multi-task EO modelling, and a conceptual bridge toward future generative EO foundation models.

Figures (7)

• Figure 1: Symbolic illustrations of the token sequence structures used for the different experiments. Note, experiment 2 uses the same sequence as experiment 1.
• Figure 2: True (A, C) and predicted (B, D) normalised confusion matrices for Sentinel 2 only (A, B) and multi-modal Sentinel 1 + 2 (C, D) PASTIS-R crop classification on fold 5 (trained on folds 1, 2 and 3).
• Figure 3: Inference examples (true & predicted) for Sentinel 2 or Sentinel 1 & 2 crop classification 2 patches in fold 5.
• Figure 4: Illustration of the gap between spatial first EO foundation models on PASTIS-R compared with this work and comparable SotA models. References in Table \ref{['tab:tab2']}.
• Figure 5: True and predicted normalised confusion matrices for Sentinel 1 + 2 + tile + lat-lon.