PhD Defence Xiao Zhu | Multi-resolution deep learning for forest structure mapping to support giant panda habitat and distribution modelling

Multi-resolution deep learning for forest structure mapping to support giant panda habitat and distribution modelling

Xiao Zhu is a PhD student in the Department of LIFE. (Co)Promotors are prof.dr. A.K. Skidmore from the Faculty of Geo-Information Science and Earth Observation and prof.dr. T. Wang from the Faculty of Geo-Information Science and Earth Observation and School of Geospatial Engineering and Science, Sun Yat-sen University.

Mountain forests are typically heterogeneous ecosystems, where steep topography, mixed vegetation, and vertical forest stratification create highly variable habitat conditions. Giant panda habitats represent distinctive mountain forest landscapes in which forest structure and understory bamboo are key components of habitat composition. However, mapping these habitat components remains difficult due to terrain shadows, mixed pixels, canopy occlusion, and limited field observations. This thesis explores how multi-resolution deep learning can improve forest structure mapping and support more ecologically informed giant panda habitat and distribution modelling.

The research first addresses terrain shadow effects in very high-resolution WorldView imagery. A bi-temporal image fusion approach is developed to combine shadow-affected multispectral imagery with higher-solar-elevation panchromatic imagery, improving evergreen conifer detection in rugged mountain forests. Building on this, the thesis develops a deep learning framework that transfers fine-scale information from WorldView imagery to Landsat time series, enabling 30 m mapping of evergreen conifer fractional cover across giant panda habitat.

The thesis then extends this multi-scale framework to evergreen understory bamboo, the key food resource for giant pandas. Fine-scale bamboo information derived from WorldView imagery is upscaled using Landsat time series and deep learning models, including LSTM and Transformer networks, to produce bamboo maps. Finally, these remotely sensed forest structure and bamboo indicators are integrated with panda occurrence and human disturbance variables to examine drivers of panda distribution in the Qinling Mountains.

Overall, this thesis demonstrates how deep learning, remote sensing, and ecological knowledge can be integrated to learn, transfer, and interpret fine-scale forest structural information across spatial scales. By linking forest structure mapping with understory resource assessment and wildlife distribution modelling, the thesis provides a transferable framework for AI-enabled biodiversity monitoring and habitat assessment in complex mountain ecosystems.