High-Resolution Remote Sensing Benchmark Dataset ：AsiaMountain-Road

1. Dataset Overview AsiaMountain Road is the first large-scale high-resolution remote sensing image road extraction benchmark dataset specifically designed for complex mountainous environments in Asia. This dataset aims to solve the problem of insufficient adaptability of existing road datasets in mountainous scenes with large terrain undulations, strong background interference, and complex road shapes, and provide data support for the research of high-precision mountain road extraction algorithms. 2. Data sources and coverage Image source: The data is sourced from high-resolution satellite imagery from Google Earth, with a spatial resolution of approximately 1 meter. Coverage area: The dataset selects 29 representative mountain areas in many Asian countries, including the Qinghai Tibet Plateau (Xizang, Qinghai), the Yunnan-Guizhou Plateau (Yunnan) in China, and typical mountain areas in Nepal, Bhutan and other countries. Coverage area: The total coverage area of the image reaches 1836 km ², covering various landform types from high mountains and canyons to plateau hills. 3. Data features and challenges: AsiaMountain Road contains complex scene features unique to mountainous roads, which are highly challenging Geometric distortion and occlusion: including a large number of road breakage phenomena caused by mountain shadows and tree occlusion. Complex topology structure: covering the unique "zigzag" shaped highways, hairpin bends, and narrow rural dirt roads in mountainous areas. Background interference: The background includes features such as terraced fields, dry riverbeds, and gravel slopes that are similar to the spectral characteristics of roads (which can easily lead to false positives). 4. Data Preprocessing and Organizational Structure (2048&1024 Slices) In order to meet the input scale requirements of different deep learning models and fully preserve the contextual information of long-distance roads, the original large image underwent strict geometric correction and manual visual interpretation annotation, and two sizes of data slices were generated through sliding window cropping: 2048 × 2048 specification: preserves more macro level geographic contextual information, suitable for large receptive field models that need to capture long-distance road dependencies. 1024 × 1024 specification: As a standard input size, it adapts to the memory limitations and training requirements of most mainstream semantic segmentation networks. All cuts have been filtered to remove pure black/invalid areas without road background. The dataset is divided into training set, validation set, and testing set (with a ratio of approximately 7:2:1), and corresponding pixel level binary labels (Ground Truth) are provided, where white (pixel value 255) represents roads and black (pixel value 0) represents background.