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

Mountain forest ecosystems are characterised by complex topography, pronounced environmental gradients, and strong structural heterogeneity, which together play a critical role in shaping biodiversity patterns and wildlife habitat quality. However, accurately mapping forest structure in mountainous regions remains challenging due to terrain-induced shadow effect, limitations of conventional machine learning approaches in fractional cover estimation, and the scarcity of spatially continuous reference data for understory vegetation mapping. These challenges are particularly evident in giant panda habitats in China, where forest structure, evergreen conifers, understory bamboo, and human disturbance jointly influence panda distribution and conservation outcomes. This thesis develops and applies a multi-resolution deep learning framework to map forest structure and assess its ecological relevance in complex mountain landscapes, with a particular focus on giant panda habitats in the Qinling Mountains. By integrating very high-resolution (VHR) satellite imagery with medium-resolution satellite time-series data and advanced deep learning models, this research addresses key methodological limitations related to terrain shadow mitigation, forest fractional cover estimation, understory vegetation mapping, and habitat modelling. To mitigate terrain shadows in VHR imagery acquired under low solar elevations, this thesis proposes a bi-temporal image fusion approach that combines shadow-affected WorldView-2 multispectral images with geometrically registered WorldView-1 panchromatic images acquired at higher solar elevations. An enhanced shadow detection method was developed to identify terrain shadows, followed by a dedicated discrimination method to distinguish between light and dark terrain shadows in mountainous forest environments. Across four mountainous study sites, the proposed approach substantially reduced terrain shadows and significantly improved the detection accuracy of evergreen conifers, particularly in heavily shaded areas. Building on the shadow-mitigated VHR imagery, a deep learning framework was developed to map the fractional cover of evergreen conifers at a spatial resolution of 30 m across the entire giant panda habitat.
A 3D U-Net model was used to generate high-quality reference fractional cover maps from VHR imagery acquired during leaf-off conditions, which were subsequently integrated with Landsat time-series data using a long short-term memory (LSTM) network for regional-scale estimation. The proposed framework demonstrated strong spatial and temporal transferability and consistently outperformed traditional machine learning approaches, highlighting the value of combining multi-source and multiresolution satellite imagery with advanced deep learning techniques for large-area forest structure mapping. This thesis further addresses the long-standing challenge of mapping understory vegetation in dense forest ecosystems by developing a multiscale remote sensing framework for evergreen understory bamboo mapping. Fine-scale bamboo reference maps derived from visually interpreted WorldView-2 imagery were upscaled using Landsat timeseries data and deep learning models. Comparisons between LSTM and Transformer-based approaches showed that attention-based architectures provided more accurate and robust predictions across a range of bamboo fractional cover thresholds. Finally, forest structural indicators derived from remote sensing were integrated with human disturbance variables to investigate the factors shaping giant panda spatial distribution in the Qinling Mountains. Habitat modelling results revealed that human disturbance—particularly proximity to cropland—outweighs food availability in shaping panda distribution patterns. Nevertheless, bamboo availability remains a critical component of habitat quality, influencing panda habitat use. These findings underscore the importance of jointly considering forest structure, food resources, and human pressures in conservation planning. Overall, this thesis demonstrates that integrating multi-resolution satellite imagery with deep learning provides a scalable and effective approach for mapping forest structure and linking structural attributes to wildlife habitat modelling in mountainous landscapes. The framework developed here advances methodological capabilities in terrain shadow mitigation, forest fractional cover estimation, and understory vegetation mapping, while also offering actionable insights for biodiversity conservation and habitat management in complex mountain ecosystems.