RSI-CB: A Large-Scale Remote Sensing Image Classification Benchmark Using Crowdsourced Data

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep convolutional neural networks (DCNNs) excel at natural image recognition.
• The remote sensing field lacks a large-scale benchmark comparable to ImageNet.
• Existing benchmarks may not be sufficient for the demands of the big data era.

Purpose of the Study:

• To propose and construct a novel, large-scale benchmark for remote sensing image classification.
• To leverage crowdsourced data for effective annotation of remote sensing imagery.
• To provide a robust dataset for training and evaluating DCNNs in remote sensing.

Main Methods:

• Utilized crowdsourced data, including Open Street Map (OSM), for annotating remote sensing images.
• Developed a worldwide, large-scale benchmark (RSI-CB) with significant geographical distribution.
• Established a classification system with six categories and 35 sub-classes, inspired by ImageNet's hierarchy.

Main Results:

• Constructed a benchmark containing over 24,000 images (256x256 pixels).
• Demonstrated RSI-CB's suitability as a benchmark through comparative experiments.
• RSI-CB proved more effective than SAT-4, SAT-6, and UC-Merced datasets for current remote sensing tasks.

Conclusions:

• RSI-CB is a valuable resource for advancing remote sensing image classification.
• The benchmark's scale and diversity make it suitable for the big data era.
• RSI-CB has broad potential applications in various remote sensing domains.