amazon-from-space
收藏资源简介：
该数据集来源于Kaggle 2017年竞赛“Planet: Understanding the Amazon from Space”，提供JPG格式的卫星图像。图像为256×256像素的裁剪图，来自Planet公司的PlanetScope卫星对亚马逊流域的观测，地面采样距离约3米。每个图像被分配一个大气条件标签（clear、partly_cloudy、cloudy、haze），以及零个或多个土地覆盖/土地利用标签（共13类，包括常见类别如primary、water、agriculture、road、habitation、cultivation、bare_ground，以及稀有类别如slash_burn、selective_logging、blooming、conventional_mine、artisinal_mine、blow_down）。标注为cloudy的图像没有其他标签。数据集包含40,479张带标签的训练图像和61,191张未标记的测试图像。作者将训练图像采用多标签迭代分层方法划分为10折用于交叉验证。提供两种配置：default（训练、验证、测试分割）和folds（所有训练图像作为训练集，测试集独立）。特征字段包括：image（JPEG图像，注意原始文件为4通道CMYK，需转换为RGB）、labels（正标签的ClassLabel索引序列）、label_names（正标签名称列表）、image_id（原始文件名前缀）、fold（折索引，测试图像为null）。该数据集适用于多标签图像分类任务，竞赛评价指标为样本平均F2分数。
This dataset is from the 2017 Kaggle competition Planet: Understanding the Amazon from Space, providing satellite images in JPG (visual RGB) format. The images are 256×256 pixel crops from the PlanetScope satellite observations of the Amazon basin, with a ground sampling distance of about 3 meters. Each image is assigned an atmospheric condition label (clear, partly_cloudy, cloudy, haze) and zero or more land cover/land use labels (13 classes in total, including common ones like primary, water, agriculture, road, habitation, cultivation, bare_ground, and rare ones like slash_burn, selective_logging, blooming, conventional_mine, artisinal_mine, blow_down). Images labeled cloudy have no other labels. The dataset contains 40,479 labeled training images and 61,191 unlabeled test images. The authors split the training images into 10 folds using a multi-label iterative stratification method for cross-validation. Two configurations are provided: default (train, validation, test splits) and folds (all training images as train set, test set independent). Feature fields include: image (JPEG image; note that original files are 4-channel CMYK, need to be converted to RGB), labels (sequence of ClassLabel indices for positive labels), label_names (list of positive label names), image_id (original file name prefix), fold (fold index, null for test images). This dataset is suitable for multi-label image classification tasks, and the competition evaluation metric is sample-averaged F2 score.
