tacofoundation/SEN2NAIPv2

SEN2NAIPv2数据集是SEN2NAIP数据集的扩展版本，包含62,242对低分辨率（LR）和高分辨率（HR）图像对，比第一版本增加了约76%的图像。数据集文件名为`sen2naipv2-unet-000{1..3}.part.taco`。数据集包括合成的RGBN NAIP波段，分辨率为2.5米和10米，降级为相应的Sentinel-2图像，并可能包含x4的放大因子。降级模型包括高斯模糊和双线性下采样、反射率协调和添加噪声三个步骤。反射率协调是最关键的步骤。在版本1中，协调模型使用U-Net架构将高斯模糊的NAIP图像转换为反射率校正的Sentinel-2类似图像。在版本2中，U-Net模型被重新训练，时间阈值从一天扩展到两天，并包括美国可用的完整Sentinel-2存档，增加了跨传感器数据集的大小到34,640张图像。除了合成数据集（`sen2naipv2-unet`），SEN2NAIPv2还引入了三个新变体：`sen2naipv2-histmatch`、`sen2naipv2-crosssensor`和`sen2naipv2-temporal`。

The SEN2NAIPv2 dataset is an extension of the SEN2NAIP dataset, containing 62,242 low-resolution (LR) and high-resolution (HR) image pairs, about 76% more images than the first version. The dataset includes synthetic RGBN NAIP bands at 2.5 and 10 meters, degraded to corresponding Sentinel-2 images and a potential x4 factor. The degradation model to generate the LR pair comprises three sequential steps: (1) Gaussian blurring and bilinear downsampling, (2) reflectance harmonization, and (3) adding noise. Reflectance harmonization is the most critical of these steps. In version 2, the U-Net model was retrained. The temporal threshold was expanded from one day to a 2-day range, and the search included the full Sentinel-2 archive available for the USA, increasing the cross-sensor dataset size to 34,640 images. Additionally, three new variants are introduced in SEN2NAIPv2: `sen2naipv2-histmatch`, `sen2naipv2-crosssensor`, and `sen2naipv2-temporal`. The dataset is designed for the task of super-resolution in Sentinel-2 imagery.