Monitoring Saltmarsh Erosion Using Cumulative Sums of Sentinel-1 Timeseries
1. Introduction: Saltmarshes are coastal wetlands that flood and drain with the tides and are an integral part of the interface between marine and terrestrial systems. They provide many important ecosystem services, but regularly suffer losses due to erosion despite conservation efforts [1]. There is a large body of work involving monitoring wetlands using Synthetic Aperture Radar (SAR), however most studies focus on wetland mapping and water dynamics, as opposed to change detection of grassland salt marshes [2]. This study aims to demonstrate that a new algorithm, the cumulative sums of Sentinel-1 (S-1) time series SAR images (CuSum-SAR) [3] can be used to effectively monitor erosion and accretion in a dynamic saltmarsh environment.
2. Methodology: this study uses a section of the Solway Firth saltmarsh in Scotland as a case study. It is known that this marsh is undergoing a process of erosion, which has been accentuated in the past recent years. Optical Sentinel-2 (S-2) data, provided by the European Space Agency (ESA) through the Copernicus program, was used to visually locate areas of the marsh where erosion had occurred between 2017 and 2020. Due to cloud cover, a continuous assessment of the region is not possible. Additionally, identifying automatically erosion using these optical images is challenging. However, due to the ease of interpretation, these images were used for the validation of our algorithm. Areas which had demonstrated either erosion or regrowth (accretion) were identified. A time series of S-1 (GRD) images from 2017 through to 2020 were obtained and used for the saltmarsh monitoring analysis, using the CuSum-SAR, an approach previously used for forest and floodplain wetland monitoring [3, 4]. The method subtracts a reference mean from each image in the time series and then cumulatively sums these difference images. This has the effect of amplifying persistent changes (increasing the power of the detector for those changes) and smoothing out those which do not persist [3]. The thresholds were set by fitting a Generalized Gaussian distribution to the historical values of CuSum_SAR and setting the threshold using a Constant False Alarm rate of 0.01. ROC (Receiver Operating Characteristics) curves, F-score and Cohen Kappa statistics are used to evaluate the performance of the detection method. In addition, a visualization method which illustrates the temporal and spatial persistence of erosion and accretion over a 3 year period is also presented, as this may be of interest to conservation management bodies.
3. Results: Preliminary results show that the algorithm can accurately detect erosion and accretion of the saltmarsh. An overally accuracy of 0.60, F-score of 0.61, and Cohen Kappa score of 0.38 indicate that whilst there is considerable agreement between the visual inspection of S-2 and S-1 obtained erosion/accretion data, there are also some inconsistencies. Areas where no discernable change was observed in the S-2 data comprise 14% of the data, yet 31% of all misclassifications occurred in these zones, which indicates the possibility that the S-1 cumulative sum change detection approach may be able to distinguish erosion/accretion at the subpixels level, i.e. at a higher resolution than the 10m S-2 data, although further work with higher resolution ground truth data will be required to confirm this.
4. Conclusion: this study demonstrates the potential of S-1 SAR data for monitoring of erosion and accretion in a dynamic saltmarsh environment. Whilst further validation and optimization is needed to bring this to an operational level, the capability of the CuSum algorithm to automatically identify erosion is clear and may prove to be a valuable tool for assessing conservation efforts in coastal marsh environments.
Acknowledgment: This ground work and validation for this project was done with the ground support from Suzanne McIntyre, Nature Scot,.
References:
[1] N. M. Foster, M. D. Hudson, S. Bray, and R. J. Nicholls, ‘Intertidal mudflat and saltmarsh conservation and sustainable use in the UK: A review’, Journal of Environmental Management, vol. 126, pp. 96–104, 2013, doi: https://doi.org/10.1016/j.jenvman.2013.04.015.
[2] S. Adeli et al., ‘Wetland Monitoring Using SAR Data: A Meta-Analysis and Comprehensive Review’, Remote Sensing, vol. 12, no. 14, 2020, doi: 10.3390/rs12142190.
[3] J. Ruiz-Ramos, A. Marino, C. Boardman, and J. Suarez, ‘Continuous Forest Monitoring Using Cumulative Sums of Sentinel-1 Timeseries’, Remote Sensing, vol. 12, no. 18, 2020, doi: 10.3390/rs12183061.
[4] J. Ruiz-Ramos, A. Marino, A. Berardi, A. Hardy and M. Simpson, ‘Characterization of Natural Wetlands with Cumulative Sums of Polarimetric SAR Timeseries’, 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, 2021, pp. 5899-5902, doi: 10.1109/IGARSS47720.2021.9554249.
Due to the harsh conditions of the Red Sea, mangroves are one of the few trees that grow in this region. Mangroves are one of the most effective ecosystems for fixing and storing carbon and benefit not only coastal and marine ecosystems, but also society through the provision of food and coastal protection. For these reasons, mangrove afforestation has gained attention in the last few years.
In this work, we mapped the actual extent of mangroves in the Red Sea, modelled the potential area for mangrove afforestation in the region and estimated the actual and potential carbon sequestration rate (CSR) of mangroves using literature values. We used Sentinel-2 imagery to map mangrove forests and the Maximum Entropy species distribution model to evaluate potential areas for mangrove afforestation.
We mapped the actual mangrove extent to be ~171 km2 and identified a suitable area of ~354 km2 for potential mangrove afforestation along the Red Sea coast. The actual CSR of the Red Sea mangroves was estimated at 1007 ± 175 Mg C yr-1, while the potential CSR was calculated at 2089 ± 365 Mg C yr-1. Our results highlight the positive trend of mangrove growth in the region that has been occurring for the past three decades. However, they also show the low CSR of Red Sea mangroves and the importance of focusing efforts on the conservation of existing mangrove forests in the Red Sea region. These results could improve future activities aimed at improving the health of existing mangroves and those aimed at mangrove afforestation in the Red Sea.
Justdiggit believes in the power of nature and in cooling down the planet together by regreening degraded land and bringing back vegetation. Together with our partners in Tanzania we inspire and activate local farmers to start regreening their own land by using a combination of agroforestry and rainwater harvesting techniques. We've empowered over 500,000 people to bring back over 9 million trees on smallholder farmland in the Dodoma and Singida regions of Tanzania.
To keep track of all these trees we've been collecting data on the ground, using drones and via satellites. Together with our partners of the VU Amsterdam University and Lynxx we have developed machine learning models to map individual trees, including very young ones, on fields of smallholder farmers.
We use a three-step approach to train the machine learning models and predict individual trees:
- First, we trained a model to detect trees on extremely detailed (5 cm resolution) drone imagery which we have for about 4% of the villages that we work in. Each drone image covers an area of approximately 100 hectares.
- The predicted trees from the drone model are used as training data for a model to detect trees on high resolution Planet SkySat imagery (50 cm resolution) in these same villages, but which covers a much larger area (about 2500 ha each).
- Ultimately, we combine all these insights to detect changes in tree cover for the whole region of Dodoma (over 4 million hectares) using free Planet NICFI (5 m resolution) and Sentinel (10 m resolution) data.
Based on the results of the regional datasets we can detect changes in tree cover and tree count for the villages that we are active in with the program and compare these with villages where we are not active to distinguish the impact of our ecosystem restoration program.
Since we use image segmentation models, we are not only able to detect the number of trees brought back by our programs, but we can also derive other statistics such as crown size distribution for these trees. Combined with field measurements and data from literature this provides us insights in the amount of carbon sequestered by the more than 9 million trees that we brought back in Tanzania.
The Minimal Sampling Classifier – MiSa.C - is an innovative classification webservice for remote sensing image raster data that allows a classification of complex surfaces over space and time by considering only a minimal amount of training samples as reference input data. Based on only one reference sample per class, the tool incorporates images statistics and machine learning to automatically create an enlarged set of unbiased and comprehensive training data, which is used as foundation for the creation of a large number of models representing the classification results. Users with expert knowledge for at least a small, representative region of the entire classification area can then evaluate the model outputs in a step-wise approach, that provides the flexibility to fine-tune each classification decision by thresholding the number of models used as basis for the final classification. A big surplus of MiSa.C is the ability to process multiple source imagery together that are provided in one 4D data-cube (areal, spectral, time domain), e.g. a combination of optical imagery, radar imagery and a digital elevation model or else.
Originating from the GFZ Potsdam German Research Centre for Geosciences in-house software development called “Habitat Sampler”, MiSa.C is an example for successful technology transfer from an innovative open source code to a user-friendly Software-as-a-Service (SaaS) webservice. The development was undertaken by the Helmholtz Innovation Lab FERN.Lab, which is part of the research centre. The lab planned and implemented the webservice in close cooperation with the lead scientist and developed an intuitive graphical user interface suitable for all kind of users coming from both the public and the private sector.
MiSa.C is especially designed for close monitoring of locally delimited regions that are dominated by mixed-pixel areas with typical spectral patterns and phenological changes over the time such as habitats and other ecosystem areas. We demonstrate applications of the Non-Governmental Organisation Heinz-Sielmann-Stiftung where MiSa.C is used to classify different habitat types in former military areas, which are not allowed to access nowadays due to ammunition load. Results of the change of habitat types over the time allow caretakers and decision makes to fulfil their report duties and to carefully plan their future restoration measurements.
Further examples show forest applications, such as tree species detection and classifications of crops within the agriculture domain. In an outlook, the authors also encourage users to use MiSa.C as an automatic training data generator that provides the necessary source data for other supervised, AI-based classifiers.
Habitat fragmentation poses one of the biggest threats to biodiversity. When natural habitats get isolated into small patches, species may have difficulty moving between them to search for resources and to mate, capping the possibility of geneflow between populations. Such is the extent of this problem that the first key message of the U.N. 2021-2030 “decade of restoration” report is that countries need to deliver on their existing commitments to restore 1 billion hectares of degraded land. Despite an increasing body of literature there is still much to be understood on how to reconnect nature.
We propose a way to solve one of the open problems to counter habitat fragmentation —the restoration of so-called stepping stones or corridors— by addressing one of its more evident issues: adequately selecting where to implement such actions under a constrained budget. This is non-trivial because given a large landscape showing some degree of fragmentation, the sheer amount of existing habitat patches makes the ways in which they may be reconnected a computationally complex problem. To tackle this, we propose to use Deep Reinforcement Learning (DRL), one of the most promising paradigms for finding optimal strategies in decision making problems with a high dimensional state-space and low prior knowledge.
DRL very generally works as follows: an agent starts in a given state within a (virtual) environment. It then gathers an initial observation. At each time step, the agent has to perform an action. What follows are three consequences: (i) the agent gets a reward, (ii) the environment transitions to a next state and, (iii) the agent obtains a new observation. The agent has an associated model (deep neural net) that, as more training steps take place, learns how to best act in a specific state. In this presentation we will describe how to specify a spatially-explicit environment in which an artificial agent will use DRL to propose stepping stones suitable for ecosystem restoration in order to maximally improve connectivity in a region, assuming costs associated with restoring different LCLU classes (e.g. water > urban > agriculture). This approach will be showcased on a land cover map of the state of Tabasco (Mexico) produced with Sentinel-2 imagery and tailored to the specific needs of Panthera Onca (Jaguars).
Access to high resolution data to support sustainable development activities, particularly for conservation and deforestation has often been limited by barriers of cost and licensing. Yet the benefit of higher resolution data provides opportunities for improved reporting, monitoring changes or high cadence updates not afforded by public sources alone.
This was one of the reasons the Norwegian Ministry of Climate and Environment through NICFI (Norway’s International Climate and Forest Initiative) funded the Global tropical forest program initiative to, for the first time ever, enable users to access high resolution data without these usual barriers. The program started in October 2020 and is led by KSAT who with it’s partners Planet and Airbus provide free access to high resolution monitoring and archival mosaics across the tropical forest region (45 M sq km) for all users as well as historical SPOT5, 6 and 7 scenes over specific areas for selected users.
The NICFI satellite data program focuses on the purpose of reducing and reversing the loss of tropical forests and is designed to have as broad access as possible to ensure it is useful for as many groups as possible. This presentation will showcase the impact of the first 18 months of the program with a focus on how the imagery is being used by different user groups. From Government, NGO to research, media and commercial users wherever they are can access the data to support improved reporting, classification, mapping and monitoring in support of saving our tropical forests. It will also highlight how facilitation of different mechanisms to view and analyse the data through both open and commercial tools including Global Forest Watch, QGIS and Google Earth Engine allows users different ways to access the data using tools or platforms familiar to them.