High-resolution remote sensing data and machine learning approaches for mapping and monitoring habitats
Citation:
Cruz, Charmaine, High-resolution remote sensing data and machine learning approaches for mapping and monitoring habitats, Trinity College Dublin, School of Natural Sciences, Geography, 2024Download Item:
Abstract:
The increasing decline in the status of habitats, mainly due to anthropogenic stressors, has spurred the development and implementation of many conservation-related legislation. This legislation involves mapping, a critical task that aims to gather important information about the habitats, including their locations, spatial extents and changes over time. In Ireland, mapping habitats is often based on field surveys. However, the methods of these surveys can be time-consuming and resource-intensive, especially if they need to be done frequently over large or remote areas. Advancements in mapping technologies, such as remote sensing and machine learning, can be integrated with field surveys to improve the efficiency and effectiveness of mapping habitats. However, they have not been well-examined in the context of mapping habitats in Ireland protected under the European Union?s Habitats Directive. As a result, the potential of these mapping technologies for assessing Irish habitats remains underexplored. Motivated by this, the research has explored the use of remote sensing data and machine learning techniques in developing robust and repeatable methodologies for assessing Irish habitats. While the research was focused on mapping coastal and upland habitats at selected protected sites in Ireland, the developed methodologies could broadly apply to other habitat types or target plant species globally.
The initial task was the assessment of the effectiveness of remote sensing data acquired by Unoccupied Aerial Vehicles (UAVs) and machine learning as tools for accurate and detailed mapping of highly dynamic and fine-scale mosaics of habitats in a coastal dune environment. UAV images are characterised by high spatial resolution (centimetre-level), providing detailed habitat information. These images can also be flexibly acquired at specific periods of the year, aligning to key phenological stages of vegetation communities when spectral discrimination is best, for example, during the flowering or senescence period. In this study, UAV imagery and field data were acquired in a dune site in Co. Kerry during the growing season in 2020: 26 May (early), 28 July (mid), and 15 October (late). Topographic data representing the terrain of the study site were also generated during the photogrammetric processing of UAV images. These datasets were then processed and analysed using the Random Forest machine learning technique to classify dune habitats at this site. The results showed that using multiple UAV datasets acquired throughout the vegetation growing season achieved higher classification accuracy compared to using just a single dataset (92.37% vs. 84.09%, respectively). Also, including topographic data consistently improved the accuracy, regardless of the number of datasets. Comparing the three UAV-acquired datasets, the analysis suggested that the dataset acquired in the middle period of the growing season, i.e., the flowering period, was better than those acquired in the early or late periods for dune habitat mapping. The high-resolution maps produced from this study provide valuable information about the detailed distribution of dune habitats, as well as highlight areas that are vulnerable to the impacts of human activities, such as frequently used tracks and beach access. These maps can then be used to guide site managers in developing timely and appropriate mitigation strategies.
A critical aspect of habitat conservation is tracking the location and expansion of invasive species, which is considered a major threat to and pressures on Irish habitats. This study explored the potential of utilising UAV-acquired imagery and deep learning (DL) techniques for detailed mapping of invasive species, including those in the early stages of invasion, i.e., occurring in small patches. Creating a robust DL model is challenging due to the requirement for large and diverse labelled training data. In remote sensing, the available labelled data can be limited as it is time- and resource-intensive to generate. This limitation can restrict DL models to obtain better predictive performance when applied to new, unseen sets of data. The study implemented a DL semantic segmentation on UAV imagery, testing different model parameters?architectures, encoder backbones and input image patch sizes?to determine an effective network structure. Moreover, the potential of applying data augmentation and pseudo-labelling to increase the amount and diversity of labelled data was investigated. Results showed that the U-Net model architecture with Inception v3 as the backbone and trained on 128 ? 128 image patches was the best model network structure based on its predictive performance: highest mean Intersection-Over-Union (mIOU) score of 0.832. Furthermore, the model trained on the augmented and pseudo-labelled data achieved an mIOU score of 0.712 on an independent dataset, while there was a decrease of 0.158 in model performance when only the original labelled data were used. This result suggests the potential of using data augmentation and pseudo-labelling techniques in creating more robust models. Automating invasive species mapping through deep learning, reducing the time and effort required, is a significant step towards a more efficient assessment of habitat conditions.
To date, a single UAV flight can only capture images over a small area?usually a few square kilometres?due to factors such as the limited battery life, the requirement of operating a UAV within the visual line-of-sight of the operator, and the specified maximum flight altitudes (120 metres in Ireland). These factors can make UAV surveys impractical for mapping habitats in an extensive, often rugged terrain, such as uplands. Therefore, high-resolution (2m) satellite imagery was used for mapping upland habitats. The study specifically investigated the impact of varying spatial resolutions on classification accuracy. Moreover, a fuzzy classification technique based on Random Forest was implemented as upland habitats are typically heterogeneous and occur in complex mosaics. Results from the analysis indicated that using higher spatial resolution images generally led to a more accurate classification compared to using images of lower resolutions (2m-resolution image: 80.34% vs 10m-resolution image: 73.55%). The fuzzy classified maps provided information on the probability of each habitat being present in every pixel of the image, giving more insights into the habitat composition. The generated habitat probabilities were further used to estimate the spatial confidence of classification over an area. In this way, more targeted assessments of areas of high uncertainty can be recommended.
Overall, the combined use of high-resolution remote sensing data and machine learning techniques offers massive potential for repeatable and systematic approaches to the fine-scale characterisation of habitats. The resulting detailed maps from these approaches can provide critical information to guide and inform habitat conservation efforts. Moreover, these maps can support and contribute to the evidence-based implementation of Goal 15 of the Sustainable Development Goals, which focuses on ?protecting, restoring and promoting sustainable use of terrestrial ecosystems and preventing biodiversity loss?.
Sponsor
Grant Number
Environmental Protection Agency (EPA)
E3 scholarship
Description:
APPROVED
Author: Cruz, Charmaine
Advisor:
Connolly, JohnPublisher:
Trinity College Dublin. School of Natural Sciences. Discipline of GeographyType of material:
ThesisAvailability:
Full text availableMetadata
Show full item recordLicences: