Looking from space to tidal flats: Integrating remote sensing and deep learning for mapping sediment and macrozoobenthos

Abstract

Tidal flats are dynamic coastal areas that support biodiversity, including migratory birds, fish, and benthic communities. However, their ecological stability is increasingly threatened by natural processes and anthropogenic pressures such as climate change and coastal development. Effective conservation and management require continuous monitoring and a thorough understanding of key environmental variables. Traditional field-based monitoring provides valuable data but is costly and labor-intensive. Remote sensing offers a complementary approach, though challenges such as low spectral contrast limit its effectiveness. Deep learning techniques, such as convolutional neural networks (CNNs), enhance remote sensing applications by extracting meaningful features using autoencoders. These autoencoder-extracted features complement spectral imagery and improve the prediction of environmental and ecological variables. Object-Based Image Analysis (OBIA) refines spatial pattern analysis by grouping pixels into spatially contiguous units. Integrating OBIA with autoencoder-extracted features enhances predictive accuracy by incorporating both spatial and spectral information. Despite the advancements, challenges remain in adapting deep-learning models across different spatial, temporal, and dataset conditions requires further investigation. Additionally, analyzing the seasonal and spatial variability of the environmental and ecological variables is crucial for optimizing monitoring strategies, given the dynamic nature of tidal flat ecosystems. This PhD thesis integrates deep learning, OBIA, and remote sensing to improve tidal flat monitoring. It explores how autoencoder-extracted features enhance environmental and ecological predictions (Chapter 2), how the combination of autoencoder-extracted features and OBIA improves spatial analysis (Chapter 3), and what factors influence model transferability across different conditions (Chapter 4). Furthermore, it analyzes the seasonal and spatial variability to enhance environmental monitoring (Chapter 5). Chapter 2 demonstrates that autoencoder-extracted features significantly improve environmental and ecological predictions. These autoencoder-extracted features capture complex patterns in tidal flat ecosystems that spectral data alone cannot, leading to more accurate assessments of macrozoobenthos and sediment properties. Chapter 3 investigates the integration of autoencoder-extracted features with OBIA. The results show that combining both methods provides a more detailed and spatially meaningful representation of tidal flat environments, improving the prediction of sediment, macrozoobenthic properties as well as the classification of species distributions. Chapter 4 explores the transferability of deep learning models across different spatial, temporal and dataset conditions. Findings indicate that while deep learning models perform well when transferred over datasets and temporal settings, their generalization across varied locations remains challenging. Chapter 5 analyzes seasonal and spatial variability in tidal flat ecosystems. The study reveals that sediment and macrozoobenthic properties did not exhibit seasonal variance, whereas chlorophyll-a concentration displayed clear seasonal patterns. Spatially, sediment properties showed highest stability, followed by macrozoobenthic communities and chlorophyll-a concentrations. The findings emphasize the importance of spatial sampling for sediment and macrozoobenthic properties as well as a combined seasonal and spatial sampling design for chlorophyll-a. Overall, this PhD thesis contributes to the development of effective methodologies for tidal flat monitoring, complementing traditional field-based approaches. By integrating deep learning, OBIA, and remote sensing, the research enhances the ability to predict and monitor environmental and ecological changes.