# A multi-label dataset for China’s agricultural and rural scenes classification from VHR satellite imagery

## TL;DR

This paper introduces China-MAS-50k, a new dataset of satellite images for classifying agricultural and rural scenes in China.

## Contribution

The novel contribution is the creation of the first VHR satellite dataset for multi-label classification of rural and agricultural scenes across China.

## Key findings

- The dataset contains 55,520 images with 135,289 labels across 18 categories.
- ResNeXt-101 achieved the best performance with an F1-score of 78.4%.
- Tail categories in the dataset remain challenging for current models.

## Abstract

This study releases China-MAS-50k, i.e., China Multi-label dataset for Agriculture & rural Scene 50k, the first very-high-resolution (VHR) remote sensing dataset for multi-label classification covering entire China’s agricultural and rural areas, filling the gap in finely annotated data for non-urban scene recognition. Based on a 50 km grid system, over 50,000 sample points were determined nationwide, where VHR Google Earth imagery were to be collected for subsequent multi-label annotation. A fine-grained label system comprising 18 categories (e.g., cropland, rural village, greenhouse and photovoltaic station, etc.) was established. Meanwhile, both a rigorously defined visual interpretation system and a labeling procedure including cross-check and error correction were proposed to maintain annotation quality. Finally, the proposed dataset has a total of 55,520 VHR images with 135,289 labels, which exhibits a long-tail distribution thus providing a challenging benchmark dataset. Furthermore, we evaluated the performance of mainstream multi-label classification models on the China-MAS-50k dataset, where ResNeXt-101 achieved the best performance with an F1-score of 78.4%, but exhibited limitations in recognizing tail categories.