Crop types are best recognized from how fields change over time (phenology), not from a single image. Our work focuses on parcel-level crop mapping in Europe using Sentinel-2 time series, with models that (1) learn directly from raw observations, (2) scale to large areas, and (3) can make early predictions while the season is still unfolding.

Key ideas

  • Time series modeling: use the full seasonal trajectory, not just a snapshot.
  • Scalability & benchmarking: compare methods fairly on large, public datasets.
  • In-season decision making: predict as early as possible, while staying accurate.

Selected works and main takeaways

AI time-series models (Transformers / Self-Attention)

Self-attention for raw optical satellite time series classification (ISPRS JPRS, 2020)

  • Introduces a self-attention / transformer-style model for crop-type mapping from raw Sentinel-2 time series.
  • Shows that self-attention and recurrent models are particularly strong on raw sequences, and provides analyses of which observations matter most for classification.
  • Establishes a practical recipe for end-to-end learning that reduces reliance on hand-crafted preprocessing.

Recurrent Neural Networks (RNNs / LSTMs)

Multi-Temporal Land Cover Classification with Sequential Recurrent Encoders (ISPRS IJGI, 2018)

  • Proposes sequential encoders using convolutional recurrent layers (e.g., LSTM/GRU variants) to summarize an entire Sentinel-2 sequence into a single representation.
  • Treats clouds/atmospheric effects as temporal noise and learns robustness by integrating evidence over time.
  • Demonstrates how RNN-style models naturally fit crop phenology and multi-temporal land-cover mapping.

Large-scale benchmark dataset for Europe

BreizhCrops: A Time Series Dataset for Crop Type Mapping (arXiv / ISPRS Archives, 2019–2020)

  • Introduces BreizhCrops, a public parcel-based Sentinel-2 time series benchmark for crop-type mapping (Brittany, France).
  • Provides both TOA and BOA time series and establishes strong baselines across several deep architectures plus Random Forest.
  • Enables reproducible method comparison and supports the community with dataset + model implementations.

Early time-series classification (predict early, not only accurately)

End-to-end learned early classification of time series for in-season crop type mapping (ISPRS JPRS, 2023)

  • Introduces ELECTS, a generic mechanism that augments any time-series classifier with a learned “stop / enough information” probability.
  • Optimizes a balanced objective of accuracy + earliness, enabling reliable crop maps earlier in the season.
  • Reduces the amount of data that must be downloaded/processed by stopping once the model is confident—important for operational, large-scale monitoring.