MT4AFE - A Deep Learning Benchmark Dataset for Agricultural Field Extraction

MT4AFE (Multi-Task Learning for Agricultural Field Extraction) MT4AFE is a dataset composed of high-resolution Sentinel-2 Level-3A satellite imagery and detailed reference data on agricultural fields derived from the French graphical parcel registry (RPG). The dataset includes 132,222 labeled image patches, each measuring 256 × 256 pixels, extracted from multiple Sentinel-2 tiles across France. It covers a wide diversity of landscapes, environmental settings, and agricultural conditions. MT4AFE has been curated to support researchers in the development and evaluation of semantic segmentation deep learning methods for agricultural field delineation. It provides not only pixel-level segmentation labels, but also field boundary annotations and distance maps, making it particularly well-suited for multi-task learning approaches. Dataset Structure Description The dataset is organized into three main folders corresponding to the years [2021, 2022, and 2023]. Each folder contains two subfolders: S2_FE: Contains the Sentinel-2 Level-3A patches. LABELS_FE: Contains the corresponding label patches. 1. Training and Validation Data [2021, 2022] The folders for 2021 and 2022 are intended for model training and validation.Each of their subfolders (S2_FE and LABELS_FE) is further divided into three subsets: training (~ 75% of the data per site) validation (~ 15%) test (~10%) Both the 2021 and 2022 datasets cover the same geographic sites, but each year uses different Sentinel-2 acquisitions and the corresponding labels. 2. Generalization Data [2023] The 2023 folder is designed for spatio-temporal generalization testing of the trained models. It contains data from different geographic sites not included in the training, validation or test phases. This folder is not split into training, validation, or test subsets. 3. Data Format All data are provided in GeoTIFF format, ensuring geospatial compatibility and easy integration with remote sensing workflows. Each Sentinel-2 image patch contains five channels in the following order:[Red (R), Green (G), Blue (B), Near-Infrared (IR), and NDVI] The corresponding label patches include three masks in the following order:[Field extent masks, Field boundaries, Distance maps] 4. File Naming Convention Each Sentinel-2 image patch and its corresponding label patch follow a consistent naming format : First part: Unique identifier (ID). Second part: Data source (S2 for Sentinel-2 images or Labels for labels masks). Middle parts: Tile ID and acquisition date. Last part: Dataset split (_train.TIF, _valid.TIF, _test.TIF). Example: 1_S2_T30TYP_20210315_train.TIF — Sentinel-2 image patch 1_Labels_T30TYP_20210315_train.TIF — Corresponding label patch 5. Statistics Files The dataset includes four statistical summary files (.txt format), providing key information about the Sentinel-2 data before image patching: one file for each year (2021, 2022, 2023) and one for the combined 2021–2022 dataset. Each file contains the mean and standard deviation values for five bands/features [Red (R), Green (G), Blue (B), Near-Infrared (IR), NDVI] These statistics are calculated prior to patch extraction, based exclusively on the training subset, ensuring no data leakage between the training and validation datasets. This allows for consistent standardization of patches during model training. In addition, each file includes information calculated prior to patch extraction, derived from the full label extent masks (i.e., training, validation, and test masks), including: Total number of pixels Count and percentage of class 0 (non-field) Count and percentage of class 1 (field) 6. Dataset Summary Data Partition Number of Patches Data Partition Number of Patches Training 93,198 Validation 18,252 Test 12,128 Generalization 8,644 Total 132,222 7. Notes Note 1: For more details about the data preparation and differences between years, please refer to the original article associated with this dataset. Note 2: NDVI was computed from surface reflectances following the standard formula, with a small constant (0.01) added to the red band to reduce sensitivity to atmospheric correction noise. This approach is based on the recommendation by Hagolle (CESBIO), detailed here. 8. Citation Please cite the following letter when publishing work derived from this dataset. IEEE style : G. Amin, O. Hagolle, and V. Demarez, “MT4AFE: A Deep Learning Benchmark Dataset for Agricultural Field Extraction,” IEEE Geosci. Remote Sens. Lett., vol. 22, pp. 1–5, 2025, doi: 10.1109/LGRS.2025.3613688. Elsevier/Harvard style : Amin, G., Hagolle, O., Demarez, V., 2025. MT4AFE: A Deep Learning Benchmark Dataset for Agricultural Field Extraction. IEEE Geosci. Remote Sens. Lett. 22, 1–5. https://doi.org/10.1109/LGRS.2025.3613688MT4AFE（Multi-Task Learning for Agricultural Field Extraction，多任务学习农业地块提取）是由高分辨率Sentinel-2 Level-3A卫星影像，以及源自法国图形地块登记册（Registre Parcellaire Graphique，RPG）的详细农业地块参考数据构建的数据集。该数据集包含132,222个标注图像块，每个图像块尺寸为256×256像素，提取自法国境内多幅Sentinel-2影像瓦片，涵盖多样的景观类型、环境场景与农业生产条件。本数据集专为支持研究人员开发、评估用于农业地块边界勾画的语义分割深度学习方法而整理，不仅提供像素级分割标签，还包含地块边界标注与距离图，尤其适配多任务学习方法。 ## 数据集结构说明 本数据集按年份划分为三个主文件夹，分别对应2021、2022与2023年。每个文件夹包含两个子文件夹： - S2_FE：存放Sentinel-2 Level-3A影像块。 - LABELS_FE：存放对应的标签块。 1. 训练与验证数据（2021、2022年） 2021与2022年的文件夹用于模型训练与验证。其下的S2_FE与LABELS_FE子文件夹进一步划分为三个子集：训练集（约占各站点数据的75%）、验证集（约15%）、测试集（约10%）。2021与2022数据集覆盖相同的地理站点，但每年使用不同的Sentinel-2影像采集数据与对应标签。 2. 泛化测试数据（2023年） 2023年文件夹用于对训练完成的模型进行时空泛化测试，其数据来自训练、验证与测试阶段未涉及的不同地理站点，且未划分为训练、验证或测试子集。 3. 数据格式 所有数据均采用GeoTIFF格式存储，保障地理空间兼容性，可轻松集成至遥感工作流中。每个Sentinel-2影像块包含5个波段，顺序为：红（R）、绿（G）、蓝（B）、近红外（IR）与归一化植被指数（Normalized Difference Vegetation Index，NDVI）。对应的标签块包含3类掩码，顺序为：地块范围掩码、地块边界掩码、距离图掩码。 4. 文件命名规范 每个Sentinel-2影像块及其对应标签块遵循统一命名格式： - 第一部分：唯一标识符（ID）。 - 第二部分：数据来源（S2代表Sentinel-2影像，Labels代表标签掩码）。 - 中间部分：影像瓦片ID与采集日期。 - 最后部分：数据集拆分标识（_train.TIF、_valid.TIF、_test.TIF）。 示例： 1_S2_T30TYP_20210315_train.TIF — Sentinel-2影像块 1_Labels_T30TYP_20210315_train.TIF — 对应标签块 5. 统计文件 本数据集包含4个统计汇总文件（.txt格式），提供影像分块前的Sentinel-2数据关键信息：分别对应2021、2022、2023年各年份，以及2021-2022年合并数据集。每个文件包含5个波段/特征（红、绿、蓝、近红外、NDVI）的均值与标准差。这些统计量在影像分块前仅基于训练子集计算，确保训练与验证集之间无数据泄露，可在模型训练阶段实现一致的影像块标准化。 此外，每个文件还包含影像分块前基于完整标签范围掩码（即训练、验证与测试掩码）计算的信息，包括：总像素数、类别0（非地块）的计数与占比、类别1（地块）的计数与占比。 6. 数据集汇总 数据分区 图像块数量 训练 93,198 验证 18,252 测试 12,128 泛化 8,644 总计 132,222 7. 注意事项 注意1：如需了解数据准备细节与各年份间的差异，请参考本数据集关联的原创学术论文。 注意2：归一化植被指数（NDVI）通过地表反射率按标准公式计算得出，为降低大气校正噪声的影响，在红光波段中加入了微小常数（0.01）。该方法基于Hagolle（CESBIO）的建议，详见此处。 8. 引用规范 使用本数据集衍生的研究成果发表时，请引用以下文献： IEEE格式：G. Amin, O. Hagolle, and V. Demarez, "MT4AFE: A Deep Learning Benchmark Dataset for Agricultural Field Extraction," IEEE Geosci. Remote Sens. Lett., vol. 22, pp. 1–5, 2025, doi: 10.1109/LGRS.2025.3613688. Elsevier/哈佛格式：Amin, G., Hagolle, O., Demarez, V., 2025. MT4AFE: A Deep Learning Benchmark Dataset for Agricultural Field Extraction. IEEE Geosci. Remote Sens. Lett. 22, 1–5. https://doi.org/10.1109/LGRS.2025.3613688