Resumen de Multispectral analysis of geospatial data through deep learning techniques.: Application on dense forested areas and transportation systems
The management of forest and road health is a relevant topic for both environmental conservation and infrastructure maintenance, particularly considering the proximity of grey and green infrastructures in different scenarios, i.e., natural elements and man-made structures. The proximity of forests and roads serves as an example of this situation by increasing the susceptibility to natural disasters and human-induced disturbances such as pollution, traffic accidents and wildfires. By considering both types of infrastructure, we gain a better understanding of potential vulnerabilities and the spread of risks, which allows us to detect and monitorize them towards effective mitigation of potential hazards. For instance, forests can influence fire behaviour near roadways, affecting infrastructure integrity and safety, while road networks can serve as firebreaks or conduits for spreading fire into adjacent green spaces.
Therefore, integrated assessments that examine the interaction between green and grey infrastructure provide valuable insights for enhancing resilience and reducing risks in such complex environments.
Remote sensing techniques, especially Light Detection and Ranging (LiDAR), offer a powerful solution for assessing the interface between forests and roads. Different type of LiDAR-based platforms, like terrestrial or aerial laser scanning (TLS and ALS, respectively), provide detailed information into forest structure and road conditions, enabling the identification of potential hazards such as unstable slopes and vegetation encroachment. By leveraging LiDAR-derived metrics, scientists and workers from different specializations can make informed decisions to proactively address risks and safeguard both natural environments and transportation facilities.
This integration of LiDAR technology into forest and road management practices holds a potential for enhancing resilience and sustainability in the face of evolving environmental challenges, and this has not been fully exploited yet.
Leveraging advances in Artificial Intelligence (AI) and Deep Learning (DL), this study introduces a comprehensive approach to perform different types of analyses compressing geospatial 3D data with spectral kernels at multiple bands, i.e., multispectral (MS) information.
Through a dissertation compressing different types of novel methodologies, the main objective of this doctoral thesis relies in the integration of novel methodologies based on DL and MS LiDAR for quantifying risk parameters related to forest wildfires close to transportation networks. To achieve this goal, this work is structured in four main chapters in which specific objectives are addressed. These specific objectives envision the integration of Machine Learning (ML) and DL based algorithms to segment LiDAR point clouds in places where roads and forests coexist, the development of synthetic data generators to feed these classifiers and wildfire risk modelling in these scenarios through the combined use of the previous algorithms to retrieve forest fuel models.
Firstly, as an initial step for the PhD candidate into LiDAR data acquisition and processing, a heuristic algorithm is designed to facilitate instance segmentation of forest 3D point clouds. This approach, which differs from most of region-growing-based algorithms from the state-of-the art of the moment, allows scientists and forest workers to conduct different types of biomass analyses, delving into individual tree (IT) characteristics through the intermediate application of ML-based techniques. Among the results of this algorithm, it is worth mentioning the high tree detection rate and accuracies in the classifications of TLS point clouds, which are over 90% and 93% respectively. By discerning geometrical patterns within forest LiDAR point clouds, this methodology lays the groundwork for informed decision-making in forestry management.
Secondly, a 3D simulation software called ROADSENSE is developed, which has the capability to automatically synthesize realistic point clouds with semantic labels and MS information.
ROADSENSE transcends traditional limitations like the intermediate requirements of simulating full-wave pulses or a 3D trajectory, offering the possibility of generating different types of scenarios within forest and transportation environments while limiting the dependence on laser scanner parameters. From dense forested areas to intricate transportation networks like highways, this tool helps researchers with a versatile platform for experimentation and analysis.
In a paradigm-shifting hybrid methodology, the ROADSENSE simulator serves as the backbone for training state-of-the-art DL-based classifiers. By using these synthetic data generated within ROADSENSE, these classifiers can achieve mean accuracies over 90% in semantic classifications within real-world point clouds. This symbiotic relationship between simulation and DL propels the boundaries of geospatial analysis, unlocking new avenues for understanding and harnessing the intricacies of our natural and built environments.
Finally, a hybrid approach is proposed, combining all the aforementioned methodologies to retrieve information regarding potential wildfire hazards in forests, like fuel models, which can predict the behaviour and characteristics of vegetation that can contribute to wildfire ignition, spread, and intensity. Because traditional approaches do not consider species identity and spectral features, relying only in geometrical distributions, one novel aspect of the proposed method consists in a previous classification of forest species into different semantic classes regarding their fire response.
From segmenting LiDAR point clouds in different green-grey close scenarios to discerning fuel models in forested areas, this composite framework serves as an example of the current potential of AI and DL techniques in environmental research, as demonstrated through the dissemination and publication of peer-reviewed scientific publications and an international conference. Through ongoing innovation and collaboration, this study offers a significant step towards enhancing environmental management by introducing novel methodologies whose purpose relies in mitigating the catastrophic effects of wildfires on both natural and built infrastructures.
