Landsat-8

数据集以JSON字符串格式存储在Google Cloud Storage中。有100个文件，每个文件都有Landsat-8的文件名、场景中凝结尾迹的多边形边界，以及每个标记场景的不标识的平流飞行航路点。我们通过使用matplotlib集合 [26] 从多边形边界构造2个地面真相掩码来手动预处理数据。然后，我们将图像和遮罩的尺寸调整为512x512尺寸进行训练; 任何小于512的尺寸都会导致结果恶化。由于凝结尾迹和背景之间的班级不平衡，我们仅使用具有至少一个轨迹的图像进行训练，并帮助模型更好地推广。数据集被进一步划分为用于训练的80% (1737) 和用于测试的20% (434)。我们选择使用假彩色图像，因为它在视觉上比RGB图像更容易识别轨迹。在实验上，这些模型在假彩色图像上的表现也更好。

The dataset is stored as JSON strings in Google Cloud Storage. There are 100 files, each containing the Landsat-8 filename, the polygonal boundaries of contrails in the scene, and unlabeled stratospheric flight waypoints for each annotated scene. We manually preprocessed the data by constructing two ground-truth masks from the polygonal boundaries using the matplotlib collection [26]. Then, we resized both the images and masks to a 512×512 dimension for training; any size smaller than 512 would result in deteriorated performance. Due to the class imbalance between contrails and the background, we only used images with at least one contrail for training to help the model achieve better generalization. The dataset was further split into 80% (1737 samples) for training and 20% (434 samples) for testing. We chose to use false-color images, as they make contrails easier to visually identify compared to RGB images. Experimentally, models also perform better on false-color images.

搜集汇总

背景与挑战

背景概述

Landsat-8数据集包含100个JSON格式的文件，记录了Landsat-8的图像、凝结尾迹边界和飞行航路点。数据集经过预处理，调整为512x512尺寸，并划分为80%训练和20%测试集，特别使用假彩色图像以提高模型表现。该数据集由Manipal Institute of Technology于2022年发布。