AeroRIT

我们研究了应用卷积神经网络 (CNN) 体系结构来促进空中高光谱场景的理解，并提出了一个新的高光谱数据集-aerorit，该数据集足够大，可以进行CNN训练。迄今为止，大多数高光谱机载设备都局限于植被和道路的各个子类别，该场景引入了两个新类别: 建筑物和汽车。据我们所知，这是第一个具有近700万个像素注释的综合大规模高光谱场景，用于识别汽车，道路和建筑物。我们通过密集语义分割的任务来比较三种流行的体系结构 (SegNet，U-Net和Res-U-Net) 的性能，以实现场景理解和对象识别，从而为场景建立基准。为了进一步加强网络，我们添加了挤压和激励块以实现更好的通道交互，并使用自我监督学习来实现更好的编码器初始化。航空高光谱图像分析仅限于具有有限的训练/测试拆分功能的小型数据集，我们相信AeroRIT将通过更复杂的对象分布来帮助推动该领域的研究。

We investigate the application of Convolutional Neural Network (CNN) architectures to facilitate aerial hyperspectral scene understanding, and propose a novel hyperspectral dataset, AeroRIT, which is sufficiently large for CNN training. To date, most airborne hyperspectral datasets have only covered various subcategories of vegetation and roads, while this dataset introduces two new categories: buildings and cars. To the best of our knowledge, this is the first comprehensive large-scale hyperspectral dataset with nearly 7 million pixel-level annotations for the recognition of cars, roads, and buildings. We compare the performance of three popular architectures (SegNet, U-Net, and Res-U-Net) via the dense semantic segmentation task for scene understanding and object recognition, thereby establishing a benchmark for this research field. To further enhance network performance, we add squeeze-and-excitation blocks to enable better channel interaction, and employ self-supervised learning to achieve superior encoder initialization. Aerial hyperspectral image analysis has been restricted to small datasets with limited train/test split capabilities, and we firmly believe that AeroRIT will help drive forward research in this domain by presenting more complex object distributions.