linlin xu

Supported by MITACS and BMO over the next three years, this project aims to predict wildfire risk across Canada using AI and heterogeneous datasets, which is critical for environmental management and disaster mitigation. By leveraging diverse data sources such as climate records, weather patterns, remote sensing data, and historical wildfire occurrences, AI algorithms can discern subtle spatial-temporal patterns and correlations, enabling more accurate and timely forecasts of wildfire outbreaks. This research aims to improve the current state-of-the-art predictive capability, such that it empowers wildfire management to minimize the devastating impact of wildfires on ecosystems, human settlements, and infrastructure.

LiDAR has been widely used to measure the structures, position and geometry of ground targets in many environmental applications. The processing and analysis of LiDAR 3D point cloud data, as a fundamental task in 3D computer vision, is critical to support these LiDAR applications. Deep neural network (DNN) models, due to their flexibility and strong feature learning capability, have been widely used for improving key 3D point cloud processing tasks. I and my HQPs have been exploring cutting-edge DNN models and methods for improving 3D point cloud processing and analysis in different LiDAR applications.

I and my students develop novel hierarchical 3D Transformer Networks (3DCTN) to improve point cloud classification and segmentation. This work has led to many journal paper published on top journals and conferences.

The above classification and segmentation models have enabled us to build a deep Transformer-based point cloud representation framework, to improve other key 3D point cloud processing tasks, i.e., object detection, 3D reconstruction, super-resolution and compression, towards building a complete toolbox for improving LiDAR 3D point cloud processing and analysis. This work has led to a highly-cited paper published on arXiv (also submitted to IEEE PAMI), and another paper submitted in Nov. 2023 to Remote Sensing of Environment for tree species mapping with airborne multispectral LiDAR data.

We will continue publishing on high-rank journals and conferences, and developing intelligent LiDAR analytics to improve environmental monitoring applications. We have been and will continue fusing LiDAR with multispectral and hyperspectral data to improve accuracy. Particularly, we will use the intelligent analytics to improve forest mapping and monitoring, to better support environmental impact evaluation, land management, and fire prevention planning in Canada.

Hyperspectral imaging, due to its high spectral resolution, can provide rich information for identifying and distinguishing spectrally-similar materials, and has become an essential tool to support various environmental monitoring and resource exploration applications. Nevertheless, due to the large data volume, the innate high­-dimensionality, spatial­-spectral heterogeneities and the noise effect of hyperspectral image (HSI), there are significant challenges for users to efficiently and accurately transform HSI into value­-added information.

To address these challenges, I and my students have been developing novel ML, DL and spectral analysis models and methods for achieving computer­-aided processing and analysis of the large­-volume complex HSI that can significantly improve the quantification of valuable bio/geo-variables, and thereby improve environmental monitoring and resource exploration.

We develop novel DL/ML and spectral analysis models to improve many different HSI processing tasks, e.g., HSI classification, spectral unmixing, HSI denoising, feature extraction and visualization. We implement these algorithms into advanced HSI software tools to facilitate the development of data processing workflows, pipelines and analytic solutions to key HSI applications.

We use UAV/drone hyperspectral systems with Cubert UHD185 and Resonon Pika XC2 hyperspectral cameras to collect HSI on vegetated and industrial areas for vegetation monitoring and pollution investigation. We use the PRISMA satellite hyperspectral images for mineral and vegetation mapping. We develop data processing pipelines to improve the processing and analysis of UAV and satellite hyperspectral imaging data to support our industry collaborators, e.g., SkyWatch, in many key applications, e.g., crop mapping, mineral exploration and methane emission monitoring applications.

We develop low-cost DIY UAV hyperspectral systems based on the openHSI project https://openhsi.github.io/.

Satelliet sensors (e.g., SAR, Passive Microwave, LiDAR, Thermal) have been widely used in the monitoring of the Canadian water and Northern ice environment to support many critical applications, e.g., e.g., climate change studies, coastal environmental management, indigenous communities, and Arctic shipping route planning and navigation. Nevertheless, efficient extraction of sea/lake ice, oil spill and whale information from these images is difficult due to many challenges, such as large data volume, noise effects, the complex marine environment, and weak signatures of targets.

To address these challenges, I and my students have been using advanced ML and deep learning (DL) techniques to develop advanced image processing models and algorithms to better support sea/lake ice mapping, oil spills monitoring and whale detection.

For sea/lake ice mapping, we collaborate with Canadian Ice Service (CIS) in developing advanced DL and image analysis algorithms for improving SAR image denoising, SAR image segmentation and object-based SAR image classification that enable a complete data processing pipeline for large-scale high-resolution daily Pan-Arctic SAR ice mapping, which enabled us to win the 1st place in the AI4EO AutoICE competition (https://platform.ai4eo.eu/auto-ice ).

For SAR oil spill monitoring, we collaborate with CIS in developing advanced ML and computer vision algorithms that accurately identify the boundaries of the potential oil spills from ­noisy SAR imagery. We compare and optimize different classifiers to improve the classification of true oil spills from the "look­-alikes". We develop active learning and semi-supervised learning approaches to improve model training using limited training samples.

For whale detection and counting, we collaborate with Fisheries and Oceans Canada (DFO) in designing advanced DL-based small object detection algorithms that improve the identification of beluga whales from high-resolution airborne optical images.