Large river systems are a vital resource providing a multitude of ecosystem benefits, yet these rivers are under increasing pressure from anthropogenic activity and global climate change, threatening our use of these resources. There is a substantial disparity between the importance of river systems and our ability to monitor them, to support sustainable management. Many methods of monitoring are point-source based which can be expensive and lack spatial coverage resulting in discontinuity. Satellite Earth Observation has been shown to overcome this spatial limitation to monitoring river processes. However, until recently satellites, with sufficient spatial resolution, have lacked the revisit time to become dynamic monitoring tools for sustainable management. This thesis explores the potential of constellations of multiple satellites to provide high spatiotemporal measurements of these large river systems by focusing on three first order river variables; Surface water extent, hydrology, and river water temperature. From these first order variables, more characteristics of the river system and management scenarios can be derived. This is carried out to provide the foundation for a potential Earth System Digital Twin (ESDT) which is a system for near real-time monitoring of a river system to support management decisions. Firstly, water surface extent ‘masks’ were required to measure the distribution of river water which varies dynamically with discharge and channel changes. To do this, the 130+ CubeSat constellation ‘PlanetScope’ was employed to provide almost daily imagery of the entire globe at 3 m spatial resolution. Traditional classification methods for optical imagery are inhibited by the high radiometric variability of these data, through its combining imagery from different times of day, atmospheric conditions, altitudes, and solar angles. In Chapter 3 artificial intelligence (AI) was used to develop a novel Convolutional Neural Network Supervised Classification which incorporates two different model formats to separate much of the radiometric variability out of the results. This model was tested on 36 rivers in 12 global biomes providing a median F1 accuracy score of 0.93. The difference was more marked in mean F1 score where there were fewer very poorly classified images, resulting in a mean F1 score 0.05 higher than the closest comparative model. This displayed both the generalisability of the model and the types of rivers and landscapes in which it did not perform well. Additional imagery was shown to improve transferability of the model to a new location with just 5 additional site-specific training images significantly improved classification accuracy (F1 increased from 0.81 to 0.90, p2500 PlanetScope images. These results clearly demonstrated the ability of the methodology to monitor the inter-annual dry season as well as the extreme 2023 droughts. This methodology also enabled concurrent identification of the impacts of reduced river stage on shipping channels which significantly narrowed during the dry season by an average of 750 m. Moreover, 65% of AOIs formed isolated pools of river water (which could lead to ecological strandings). Finally, the third first order component, river water surface temperature, is inherently linked to anthropogenic activity where river water is used for power generation. Two issues currently prevent water surface temperature monitoring by satellites, which has been studied extensively, from being operationalised. Firstly, thermal infrared satellites often lack the temporal resolution to monitor a component as dynamic as water temperature. Therefore, a constellation of constellations approach was undertaken where no single satellite provided more than 45% of the available data, combining imagery from multiple different thermal satellites in a modularised format which enables future satellites to be added to this quasi-constellation. Secondly, the higher spatial resolution PlanetScope imagery was used as an independent water mask to reduce the potential error of mixed land and water thermal pixels from returning erroneous results. This demonstrated the methodologies effectiveness at identifying the downstream thermal pollution effects from nuclear, coal, and hydropower generation with 2oC temperature spikes which persisted variable distances downstream based on individual facets of these river systems. These chapters are used as the basis to produce the foundation from which an ESDT can be built by providing high spatio-temporal methods for transferable monitoring of large river systems. Rivers are highly variable therefore the methodology for each component cannot be effective in all river systems, but the individual chapters are tested in a variety of locations to assist the user in applying them only when appropriate. In the synthesis, these components were assessed against 5 key tenets to understand how they can be used within an ESDT and what is required to produce an operationalised end product. Despite the improvements made, there are limitations on the minimum width of a river that can be measured, and the improved temporal resolutions may still overlook dynamics within these observed components which cannot be overcome without development of additional satellite systems. Their modular methodology can be continually built upon with additional satellite launches and new artificial intelligence methods. The thesis calls for further application of these methods to complement regular in-situ monitoring to assist sustainable management.
Samuel Valman (Thu,) studied this question.