跳到正文
Ai2 / Allen Institute for AI·· 2026-10-22AI 評分50

PLANTEUR 將作物製圖基準擴展至六個熱帶地區

A crop-mapping benchmark for the tropics

AI 導讀

Ai2 的 PLANTEUR 將 PASTIS 作物製圖基準擴展至六個熱帶地區。數據集包含 2,750 組對齊的 Sentinel-2、Sentinel-1 和 Landsat-8 時序,以及覆蓋 4,506 km²、涵蓋七類作物的地塊標註。所有受測方法在 PLANTEUR 上的得分均明顯低於 PASTIS;模型在未參與訓練的島嶼測試時,準確率約減半。

正文

October 22, 2026

Ai2


Tech ReportDataCode

Foundation models and open data are driving rapid progress in machine learning for remote sensing, and benchmarks play a key role in understanding model improvement. PASTIS is one of the community's most trusted benchmarks, thanks to its high-quality labels and geospatial metadata. It also represents an important problem. Crop-type mapping is the starting point for food security planning, water allocation, and tracking how farmland changes. However, its coverage is limited to metropolitan France. We present PLANTEUR, an extension of PASTIS to six tropical regions, in hopes that teams building geospatial foundation models can check whether their model performs in agricultural systems where mapping could be most valuable: smallholder, tropical, topographically complex.

What’s in PLANTEUR

PLANTEUR was built from the same reliable data source as PASTIS, France’s Land Parcel Identification System, and spans Réunion, Guadeloupe, Martinique, Guyane, Mayotte, and Saint-Martin. To make PLANTEUR easy to adopt into existing evaluation pipelines, its format closely resembles that of PASTIS: 2,750 aligned Sentinel-2, Sentinel-1, and Landsat-8 time series, and 4,506 km² of parcel-derived labels across seven crop classes, including sugarcane, banana, pineapple, and tropical tubers (root crops such as cassava, yam, and taro).

We include baseline numbers for several linear probes and supervised methods on a standard split, a leave-one-island-out split, a few-shot setup, and even one that includes PASTIS samples in the training data. These baselines introduce new challenges, in new places, like performing well in the tropics, generalizing to unseen regions, learning from scarce labels, and transferring knowledge across agricultural systems.

What it opens up, and how far there is to go

We built PLANTEUR to increase the geographic diversity of remote sensing benchmarks. By building on the success of the PASTIS dataset, we hope to provide builders of remote sensing models a way to evaluate how good their models are at crop type mapping in many different geographies.

Every method we evaluated, ours included, scores substantially lower on PLANTEUR than on PASTIS, and every model loses roughly half its accuracy when tested on an island it never saw in training. Just as PASTIS motivated the community to develop strong methods for crop mapping in metropolitan France, we hope PLANTEUR motivates the community to develop methods which perform well in tropical contexts as well.

PLANTEUR is available on Hugging Face and GitHub.

來源:Ai2 / Allen Institute for AI · allenai.org