Dataset for Land/Water Semantic Segmentation in Tonga and other Pacific regions

Dataset for training CNN models (e.g. U-Net) for land/water semantic segmentation of Sentinel-2 images.
Features 424 Sentinel-2 TOA images of size 256x256 pixels, with corresponding targets (binary segmentation maps).
The channels included are Sentinel-2 bands B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12, plus NDWI and NDVI (in this order).
Weight maps are included for training using a weighted loss function, in order to improve segmentation on important features, specifically small volcanic islands.

The weights of the pre-trained models used for our research are included as .keras files.
Code snippets include essential functionality to load the models, and pre-process Earth Engine images for input.

本数据集用于训练卷积神经网络（CNN）模型（如U-Net），以完成哨兵-2号（Sentinel-2）影像的水陆语义分割任务。数据集包含424张尺寸为256×256像素的哨兵-2号大气顶层反射率（TOA）影像，以及与之对应的二值分割标注图。
影像通道依次涵盖哨兵-2号的B2、B3、B4、B5、B6、B7、B8、B8A、B11、B12波段，外加归一化水体指数（NDWI）与归一化植被指数（NDVI）。
数据集附带权重图，可用于基于加权损失函数的训练流程，以优化小型火山岛等关键地物的分割效果。
本研究中使用的预训练模型权重以.keras文件格式提供。
附带的代码片段包含加载模型、将地球引擎（Earth Engine）影像预处理为模型输入所需的核心功能。