Remote Sensing of Trees: Integrating deep learning and high-resolution images to map individual trees in East Africa
Publikation: Bog/antologi/afhandling/rapport › Ph.d.-afhandling › Forskning
Standard
Remote Sensing of Trees : Integrating deep learning and high-resolution images to map individual trees in East Africa. / Mugabowindekwe, Maurice.
Department of Geosciences and Natural Resource Management, Faculty of Science, University of Copenhagen, 2024. 119 s.Publikation: Bog/antologi/afhandling/rapport › Ph.d.-afhandling › Forskning
Harvard
APA
Vancouver
Author
Bibtex
}
RIS
TY - BOOK
T1 - Remote Sensing of Trees
T2 - Integrating deep learning and high-resolution images to map individual trees in East Africa
AU - Mugabowindekwe, Maurice
PY - 2024
Y1 - 2024
N2 - The world is halfway through the UN Decade on Ecosystem Restoration aiming to “Prevent, halt and reverse the degradation of ecosystems” (UN, 2019). Global efforts are ongoing to restore, protect, and conserve ecosystems such as “Forests, grasslands, croplands, wetlands, savannas, and other terrestrial to inland water ecosystems, marine and coastal ecosystems and urban environments”. The UN Decade on Ecosystem Restoration aligns and is supported by other preceding and ongoing global agreements on ecosystem restoration and sustainability. These include Sustainable Development Goals, the Paris Agreement, the Convention on Biological Diversity, the Aichi Biodiversity Target, the Kyoto Protocol, the Bonn Challenge, among others. These agreements and initiatives share a common interest to mitigate climate change and halt the biodiversity loss. Among other activities, the restoration of forests and other tree-dominated ecosystems has received a particular attention due trees’ ecological, social, economic, cultural, and aesthetic ecosystem services to support life on earth.Planting new trees has been one of the core parts of the landscape restoration efforts of many countries and organisations to meet their goals. Particularly since the ratification of Paris Agreement in 2015, there have been pledges to plant at least 1 trillion new trees across the world. Indeed, trees provide a range of fundamental ecosystem services such as storing carbon, providing habitats for a wide range of biodiversity, providing nutritional benefits, supporting increased income and diversified sources of likelihoods, among other benefits. However, most of the ongoing restoration initiatives do not necessarily have a fully functioning, comprehensive, and consistent monitoring system, which measures the progress and impact of the restoration efforts at the level of individual trees at large scale. Since “we can’t manage what we can’t measure” (Maus & Werner, 2023), there is a need to turn the “known unknowns” about trees into the known information regularly used to inform decisions and policies related to landscape restoration. For instance, questions related to location, count, size, biomass and carbon stock of trees within an ecosystem can be availed to inform the planning of tree restoration activities, to optimise the resources allocation and biophysical impact of the restoration efforts. Such a system can also provide information on the availability and location of areas to accommodate new trees, and potential impact of the planned new trees vis-a-vis the envisioned restoration targets.This thesis presents approaches that can be adopted towards developing a functional, comprehensive, and consistent monitoring system for regular updates on tree restoration initiatives at the level of individual trees at large scale. This is presented by four studies, which integrate deep learning methods and high resolution remote sensing images, to map and characterise individual trees both inside and outside of forests at national scales in East Africa. The studies cover all ecosystems in East Africa, including tropical rainforests, other natural forests, woodlands, shrublands, savannas, grasslands, farmlands, wetlands, mangroves, bare or sparse vegetation, and built-up and urban environments.The first study starts with a particular focus on Trees Outside Forests (TOF). It showcases how TOF can be mapped and monitored using the emerging new methods integrating deep learning with high resolution satellite images, to optimise their potential as part of natural climate solutions. The second study uses a similar approach but with a focus on producing comprehensive maps of location, size (i.e. crown area), biomass, and carbon stock of individual trees both inside and outside forests at the national scale of Rwanda. The study shows that the manually delineated forest maps have missed over 38% of individual trees scattered across open landscapes, which contain over 25% of the estimated total national carbon stock. The study also demonstrates the transferability of the approach in East Africa, beyond Rwanda. The third study demonstrates how the mapping and monitoring of individual trees can support the national climate change mitigation goals such as the “net zero by 2050”. It shows that for about one decade, smallholder farmers planted and maintained at least 3 trees per plot, which contributed to about 50 million new trees within farmlands nation-wide. Furthermore, under three different scenarios, the study presents options to “reach new tree densities for optimal restoration of both agroforestry areas and degraded forests”. It further provides information on ecological and socioeconomic benefits associated with each option, and illustrates how each option would support the national carbon neutrality goal in Rwanda.The fourth study, which is a preliminary draft of a manuscript under preparation, adapts deep patch-level regression principles from the field of computer vision, to develop an approach which estimates the count of individual overstory trees across all ecosystem types at national scale in Rwanda, Burundi, and Tanzania. The approach uses a composite of PlanetScope, Sentinel-1, and Sentinel-2 satellite images at 3 m resolution. This allows the model to take advantage of tree-specific visual cues as observed in PlanetScope, a broad array of multi-spectral features from Sentinel-2, and structure and roughness features from Sentinel-1, and relates them to the count of trees within an area. The approach includes young and small trees not necessarily visible at the resolution of input images, which lays a foundation for future work to provide information such as the survival rate of newly planted trees. Although this is one of the key performance indicators for tree restoration projects, conventional methods have been challenged to provide such information accurately, mainly due to the small size of the newly planted trees. The study estimated a total count of about 6.9 billion individual overstory trees across in Rwanda, Burundi, and Tanzania, and only 53.3% of the estimated count is found in “tree cover” class using WorldCover classification (Zanaga et al., 2022), while the rest is found in other ecosystems dominated by open landscapes.
AB - The world is halfway through the UN Decade on Ecosystem Restoration aiming to “Prevent, halt and reverse the degradation of ecosystems” (UN, 2019). Global efforts are ongoing to restore, protect, and conserve ecosystems such as “Forests, grasslands, croplands, wetlands, savannas, and other terrestrial to inland water ecosystems, marine and coastal ecosystems and urban environments”. The UN Decade on Ecosystem Restoration aligns and is supported by other preceding and ongoing global agreements on ecosystem restoration and sustainability. These include Sustainable Development Goals, the Paris Agreement, the Convention on Biological Diversity, the Aichi Biodiversity Target, the Kyoto Protocol, the Bonn Challenge, among others. These agreements and initiatives share a common interest to mitigate climate change and halt the biodiversity loss. Among other activities, the restoration of forests and other tree-dominated ecosystems has received a particular attention due trees’ ecological, social, economic, cultural, and aesthetic ecosystem services to support life on earth.Planting new trees has been one of the core parts of the landscape restoration efforts of many countries and organisations to meet their goals. Particularly since the ratification of Paris Agreement in 2015, there have been pledges to plant at least 1 trillion new trees across the world. Indeed, trees provide a range of fundamental ecosystem services such as storing carbon, providing habitats for a wide range of biodiversity, providing nutritional benefits, supporting increased income and diversified sources of likelihoods, among other benefits. However, most of the ongoing restoration initiatives do not necessarily have a fully functioning, comprehensive, and consistent monitoring system, which measures the progress and impact of the restoration efforts at the level of individual trees at large scale. Since “we can’t manage what we can’t measure” (Maus & Werner, 2023), there is a need to turn the “known unknowns” about trees into the known information regularly used to inform decisions and policies related to landscape restoration. For instance, questions related to location, count, size, biomass and carbon stock of trees within an ecosystem can be availed to inform the planning of tree restoration activities, to optimise the resources allocation and biophysical impact of the restoration efforts. Such a system can also provide information on the availability and location of areas to accommodate new trees, and potential impact of the planned new trees vis-a-vis the envisioned restoration targets.This thesis presents approaches that can be adopted towards developing a functional, comprehensive, and consistent monitoring system for regular updates on tree restoration initiatives at the level of individual trees at large scale. This is presented by four studies, which integrate deep learning methods and high resolution remote sensing images, to map and characterise individual trees both inside and outside of forests at national scales in East Africa. The studies cover all ecosystems in East Africa, including tropical rainforests, other natural forests, woodlands, shrublands, savannas, grasslands, farmlands, wetlands, mangroves, bare or sparse vegetation, and built-up and urban environments.The first study starts with a particular focus on Trees Outside Forests (TOF). It showcases how TOF can be mapped and monitored using the emerging new methods integrating deep learning with high resolution satellite images, to optimise their potential as part of natural climate solutions. The second study uses a similar approach but with a focus on producing comprehensive maps of location, size (i.e. crown area), biomass, and carbon stock of individual trees both inside and outside forests at the national scale of Rwanda. The study shows that the manually delineated forest maps have missed over 38% of individual trees scattered across open landscapes, which contain over 25% of the estimated total national carbon stock. The study also demonstrates the transferability of the approach in East Africa, beyond Rwanda. The third study demonstrates how the mapping and monitoring of individual trees can support the national climate change mitigation goals such as the “net zero by 2050”. It shows that for about one decade, smallholder farmers planted and maintained at least 3 trees per plot, which contributed to about 50 million new trees within farmlands nation-wide. Furthermore, under three different scenarios, the study presents options to “reach new tree densities for optimal restoration of both agroforestry areas and degraded forests”. It further provides information on ecological and socioeconomic benefits associated with each option, and illustrates how each option would support the national carbon neutrality goal in Rwanda.The fourth study, which is a preliminary draft of a manuscript under preparation, adapts deep patch-level regression principles from the field of computer vision, to develop an approach which estimates the count of individual overstory trees across all ecosystem types at national scale in Rwanda, Burundi, and Tanzania. The approach uses a composite of PlanetScope, Sentinel-1, and Sentinel-2 satellite images at 3 m resolution. This allows the model to take advantage of tree-specific visual cues as observed in PlanetScope, a broad array of multi-spectral features from Sentinel-2, and structure and roughness features from Sentinel-1, and relates them to the count of trees within an area. The approach includes young and small trees not necessarily visible at the resolution of input images, which lays a foundation for future work to provide information such as the survival rate of newly planted trees. Although this is one of the key performance indicators for tree restoration projects, conventional methods have been challenged to provide such information accurately, mainly due to the small size of the newly planted trees. The study estimated a total count of about 6.9 billion individual overstory trees across in Rwanda, Burundi, and Tanzania, and only 53.3% of the estimated count is found in “tree cover” class using WorldCover classification (Zanaga et al., 2022), while the rest is found in other ecosystems dominated by open landscapes.
M3 - Ph.D. thesis
BT - Remote Sensing of Trees
PB - Department of Geosciences and Natural Resource Management, Faculty of Science, University of Copenhagen
ER -
ID: 400371593