Liu, Shengjie Kris

This webpage is no longer updated. Please visit skrisliu.com for latest news.

I am a PhD student at University of Southern California. Prior to joining USC, I was a research assistant in light pollution at The University of Hong Kong.

My research interests include remote sensing, machine learning, and environment & sustainability. I develop advanced AI methods to process Earth Observation data (AI4EO) for sustainability research.

Email  /  CV  /  Github  /  Google Scholar

Map of local climate zone in Hong Kong

Map of Flickr photos density in Hong Kong

Estimated PM2.5 and PM10 simultaneously on the newly-launched Zhuhai-1 hyperspectral imagery to distinguish off-road and near-road air pollution

Analyzed the cloud amplification effect on NSB based on 30+ stations of the GaN-MN

Examined the relationship between night sky brightness (NSB) and two medium-resolution remote sensing data (Luojia-1 and ISS) in Hong Kong

Analyzed the long-term artificial light at night (ALAN) using VIIRS monthly product in Hong Kong between 2012 and 2019

Generated multi-label local climate zone map in Hong Kong using very high-resolution imagery

Identify unknown classes in hyperspectral land cover mapping in the open world using multitask deep learning. Empower the deep learning models with the ability to reject unseen categories of land cover.

Enhanced multiview active learning based on the disagreement of local minima of convolutional neural networks in a snapshot ensemble fashion. Improved the stability of neural networks for PolSAR image classification.

Using scene classification strategy with deep neural network to generate local climate zone maps in fifteen cities in China.

Enhanced hyperspectral image classification with spectral knowledge from other images using multitask deep learning. * The figure presented is from a working paper based on this method.

Proposed a novel method to empower CNNs to produce object-based classification maps, named “Object-Based Post-classification Refinement” (OBPR). Also analyzed the effect of fusing Sentinel optical and PolSAR data in urban land cover mapping.

Proposed a light-weight neural network named Wide Contextual Residual Network (WCRN), and combined it with active learning to reduce the need of training samples for remote sensing image classification.

2019-2023 Shengjie Liu
Modified from Jon Barron