Statistical Deep Learning for Spatial and Spatio-Temporal Data

Wikle, Christopher K.; Zammit‐Mangion, Andrew

doi:10.48550/arxiv.2206.02218

Cited by 4 publications

(5 citation statements)

References 84 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The application of spatio-temporal models to this unique dataset reveals that consideration of interactions between space and time is essential in order to predict recovery patterns and investigate the fine-scale variability of coral dynamics (Appendix S1). Such interactions are challenging to compute, even using Bayesian approaches, but improvements in the field of computational science and applied statistics will ease their inclusion in future statistical modelling frameworks (Wikle and Zammit-Mangion, 2022). Moreover, these computational improvements will enable the scaling-up of the approach to more than one reef.…”

Section: Spatio-temporal Modelling For Coral Reef Datamentioning

confidence: 99%

Fine-scale interplay between decline and growth determines the spatial recovery of coral communities within a reef

Vercelloni¹,

Roelfsema²,

Kovács³

et al. 2023

Preprint

View full text Add to dashboard Cite

As coral reefs endure increasing levels of disturbance, understanding patterns of recovery following disturbance(s) is paramount to assessing the sustainability of these ecosystems. Given the slow dynamics of coral reefs and the increasing frequency of environmental pressures, management strategies focus on understanding recovery patterns to drive efforts and actively promote the recovery of key coral populations. However, the fine spatial scale heterogeneity of coral dynamics challenges our capacity to understand recovery patterns at large spatial scales and guide effective management actions. In this study, we developed a spatio-temporal statistical model to estimate the long-term trajectories of branching, plate and massive corals at fine-spatial scales and predict their recovery patterns at unobserved locations within a reef. We parameterized the model using repeated and georeferenced observations from 783 locations during 16 years at Heron Reef (Great Barrier Reef, Australia). We then developed indicators of recovery that capture the interplay between coral growth and relative decline from disturbance(s) across time, space and growth morphology. Our results reveal that successful recoveries, expressed in terms of probability, are associated with minimum growth rate thresholds of 4.3\% and 6.4\% (absolute growth, y-2) for branching and plate corals in reef locations that were impacted by disturbance(s) at medium-high levels and historically abundant. As a product of the data revolution, predictive maps from statistical models support the development of new indicators that can support the identification of areas of concern to prioritise management intervention. They should be used into larger spatially explicit modelling framework for decision-making in reef conservation and restoration.

show abstract

Section: Spatio-temporal Modelling For Coral Reef Datamentioning

confidence: 99%

Fine-scale interplay between decline and growth determines the spatial recovery of coral communities within a reef

Vercelloni¹,

Roelfsema²,

Kovács³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Another limitation is the need for high computing power to employ complex Bayesian statistical models. Exciting advancements in statistical ecology that aim to address these challenges include the use of deep learning (Wikle and Zammit-Mangion 2022) and a hybrid modelling framework (Sainsbury-Dale et al 2021) to speed up estimation of the spatio-temporal structure from data. These advancements will allow models to include more detailed interactions between coral communities and to be scaled up to additional reefs and regions.…”

Section: Scaling-up Detection Of Spatial Recoverymentioning

confidence: 99%

Fine‐scale interplay between decline and growth determines the spatial recovery of coral communities within a reef

Vercelloni,

Roelfsema,

Kovacs

et al. 2023

Ecography

View full text Add to dashboard Cite

As coral reefs endure increasing levels of disturbance, understanding recovery patterns of reef‐building hard corals is paramount to assessing the sustainability of these ecosystems. At local scales, coral recovery slows down; however, it's unclear how this trend propagates across spatial scales due to the inherent complexity of coral dynamics. In this paper, we aimed to learn about fine scale heterogeneity of coral dynamics and explore implications for assessing coral recovery at larger spatial scales. We developed a spatio‐temporal statistical model to estimate long‐term trajectories of three types of corals and predict their recovery patterns at unobserved locations within a reef. Then, model predictions were used to derive metrics that capture the interplay between coral growth and decline from disturbance(s) across time, space and growth morphology. This model is developed in the context of a substantive case study at Heron Reef using a high spatio‐temporal resolution dataset. Our results revealed that successful coral community recoveries took place in different habitats of Heron Reef and associated with various reasons. Branching corals recovered in the southern slope, due to fast growth in locations that were previously abundant. Plate corals flourished in the northern slope due to fast growth, despite a large decline and low baseline cover. They also recovered in the southern slope but in this case there was both a low decline and baseline cover. At Heron Reef, the recovery of coral communities followed specific conditions that were acting at a fine scale in a complex and heterogeneous way within habitat. This implies that capturing the variability of fine‐scale coral dynamics is an important first step to detect accurate signals of coral recovery at larger spatial scales. The approach proposes here can be further extend to the scale of a reef and beyond enabling assessment of recovery patterns representative at management scales.

show abstract

“…Likewise, the dependency observed in space or time has triggered intellectual curiosity in the deep learning community. Wikle and Zammit-Mangion (2022) provide a recent overview of statistical and AI approaches for deep learning spatial and spatio-temporal data. Convolutional neural networks (CNNs), recurrent neural networks (RNNs), and their variants are the primary approaches in the deep learning literature for such data.…”

Section: Introductionmentioning

confidence: 99%