AI helps scientists see what giant pandas see

Zhu combined very high-resolution WorldView images with decades of satellite data and deep-learning models. By combining images taken under different light conditions, she reduced the shadows and revealed forest structures that were hard to detect before. The models also learned to recognise evergreen bamboo, even where the treetops block the view.

The result was a detailed map of bamboo across the Qinling Mountains, one of China’s most important panda regions. The west held large, connected patches of bamboo. The east was far more broken up. These differences affect how pandas move and how well the habitat can support them.

Zhu did not stop at mapping forests. She also asked a central question in panda conservation: what really decides where pandas live? Conservationists have long assumed that the amount of bamboo is the main factor.

Zhu found something different. Using her forest maps and ecological models, she showed that human activity matters even more than food supply. The distance to cropland and roads predicted panda numbers better than bamboo alone. In other words, enough bamboo is not enough. A forest can still be unsuitable if there is too much human pressure nearby.

Although Zhu developed her methods for pandas, they can be used far more widely. Similar approaches could help researchers study tropical rainforests and other places where shadows and dense vegetation make satellite mapping difficult. “This is not only about pandas,” says Tiejun Wang, Zhu’s co-promotor. “It is about understanding ecosystems in ways that were not possible before.”

Xiao Zhu carried out her PhD research at ITC, the Faculty of Geo-Information Science and Earth Observation of the University of Twente. Her promotor was Prof Dr Andrew K. Skidmore, and her co-promotor was Prof Dr Tiejun Wang of Sun Yat-sen University, China. Her thesis is titled Multi-resolution deep learning for forest structure mapping to support giant panda habitat and distribution modelling.