TL;DR
PLANTEUR is a new open-access benchmark for crop-type segmentation specifically designed for tropical regions, filling a major geographic gap in remote sensing datasets. By providing aligned multi-sensor satellite time series across six French overseas territories, the authors demonstrate that current state-of-the-art models struggle significantly with the unique agronomic and topographic complexities of the tropics compared to temperate regions.
What Problem Does This Paper Address?
While machine learning has revolutionized remote sensing, the benchmarks we use to track progress are heavily biased toward temperate climates, particularly North America and Europe. The "gold standard" for crop-type mapping, the PASTIS dataset, is limited to metropolitan France.
This is a problem because agricultural systems in tropical regions—characterized by smallholder farms, complex topography, and different crop cycles—are often where mapping is most critically needed for food security and environmental monitoring. Models trained or validated on European landscapes often fail to generalize to these diverse tropical environments. PLANTEUR aims to bridge this "tropical gap" by providing a high-fidelity dataset that challenges models to handle geographic and agronomic diversity.
Methods
The authors designed PLANTEUR to be a direct extension of the PASTIS framework to ensure ease of adoption. Key methodological highlights include:
- Multi-Sensor Alignment: The dataset provides georeferenced and temporally aligned sequences from three different satellite constellations: Sentinel-2 (optical), Sentinel-1 (radar), and Landsat-8 (multispectral).
- High-Fidelity Labels: Labels are derived from France’s Land Parcel Identification System (LPIS). Because these are based on farmer declarations tied to financial subsidies and audits, the label quality is exceptionally high (claimed error rate < 0.3%).
- Evaluation Protocols: Beyond standard train/test splits, the paper introduces a Leave-One-Island-Out (LOIO) protocol. This forces models to train on a subset of territories and test on an entirely unseen region, directly measuring cross-territory generalization.
- Model Benchmarking: The authors evaluated three families of models: supervised models trained from scratch (like U-TAE and TSViT), linear probes on frozen foundation model embeddings (e.g., Google’s AEF, Tessera), and fully fine-tuned foundation models (OlmoEarth).
Datasets
The primary contribution is the PLANTEUR benchmark:
- Scope: 2,750 multi-spectral sequences covering 4,506 km² across Réunion, Guadeloupe, Martinique, Guyane, Mayotte, and Saint-Martin.
- Classes: An 8-class label space including Meadow, Fruit/Veg/Flower, Orchard, Sugarcane, Banana, Pineapple, Tropical tuber, and Background.
- Comparison: It mirrors the structure of the PASTIS dataset but focuses on tropical crop regimes absent in metropolitan France.
Key Insights and Ideas
- The "Tropical Difficulty" is Real: All evaluated models performed substantially worse on PLANTEUR than on the temperate PASTIS benchmark. This confirms that tropical regions present a unique set of challenges that current models are not yet equipped to solve.
- Foundation Models Lead the Way: Fine-tuned foundation models (specifically OlmoEarth v1.2) outperformed traditional supervised models trained from scratch. This suggests that large-scale pre-training on diverse satellite data captures fundamental features that are useful even in niche tropical environments.
- Generalization is the Bottleneck: The massive drop in performance during Leave-One-Island-Out (LOIO) testing shows that "tropical" is not a monolith. A model that learns to identify sugarcane in Réunion doesn't automatically succeed in Guadeloupe, highlighting the need for more robust domain adaptation techniques.
- Temporal Context Matters: Performance increases significantly as the temporal sequence length grows from one to twelve monthly mosaics, emphasizing that crop-type mapping is inherently a time-series problem.
Key Results
- Top Performer: The fine-tuned OlmoEarth v1.2 achieved the highest in-domain performance with 40.82 mIoU.
- Supervised Baselines: Established models like TSViT (36.33 mIoU) and U-TAE (35.82 mIoU) lagged behind the fine-tuned foundation models.
- The Generalization Gap: In the LOIO setup, the best mean mIoU dropped to 25.28, a roughly 40% decrease compared to in-domain results.
- Few-Shot Transfer: The authors found that adding PASTIS data to the training set improved performance in few-shot scenarios, suggesting that features learned from temperate crops can partially transfer to tropical ones.
Limitations and Future Work
The authors acknowledge several limitations inherent in the data source:
- Completeness: The LPIS system only tracks farmers participating in the EU Common Agricultural Policy. Consequently, many small-scale or subsistence farms are missing from the labels.
- Class Granularity: Some diverse tropical crops are lumped into generic categories (e.g., mango and lychee are both under "Orchards").
- Seasonal Sync: For regions in the Southern Hemisphere (like Réunion), growing seasons are out of sync with Northern Hemisphere imagery, which may confuse models not accounting for seasonality.
Future work will focus on sourcing additional high-quality labels to expand the distribution to even more small-scale and diverse tropical crops.
Summary updated