ForTy
收藏资源简介:
ForTy是一个全球性的、多模式的、多时相的基准数据集,旨在推动森林类型制图的研究。该数据集包含了20万个图像块的时间序列,每个图像块都包含Sentinel-2、Sentinel-1、气候和海拔数据。每个时间序列以月度或季节性频率捕捉变化。数据集的每个像素都进行了标注,包括森林类型和其他土地利用类别,支持图像分割任务。ForTy通过利用多个公共数据源,实现了全球覆盖。数据集被用于评估几个基线模型,包括卷积神经网络和基于Transformer的模型。此外,我们提出了一种新型的基于Transformer的模型,专门设计用于处理多模态、多时相卫星数据,用于森林类型制图。我们的实验结果表明,该模型在性能上优于基线模型。
ForTy is a global, multimodal, and multitemporal benchmark dataset developed to advance research on forest type mapping. This dataset contains time series of 200,000 image patches, each containing Sentinel-2, Sentinel-1, climatic, and elevation data. Each time series captures changes at monthly or seasonal frequencies. Every pixel in the dataset is annotated with forest types and other land use categories, supporting image segmentation tasks. The dataset achieves global coverage by leveraging multiple public data sources. It has been used to evaluate several baseline models, including convolutional neural networks and Transformer-based models. Furthermore, we propose a novel Transformer-based model specifically designed to handle multimodal, multitemporal satellite data for forest type mapping. Our experimental results demonstrate that this model outperforms the baseline models.

- 1Not Every Tree Is a Forest: Benchmarking Forest Types from Satellite Remote SensingGoogle DeepMind · 2025年



