Global Mangrove Watch: Timeseries of Mangrove Extent

The Global Mangrove Watch (GMW; https://www.globalmangrovewatch.org) v4.1 timeseries layers were generated using the full Landsat and Sentinel-2 timeseries and available JAXA L-band data (JERS-1, ALOS PALSAR, ALOS-2 PALSAR-2). The methodology is built on knowledge gained from developing the previous GMW layers (GMW v2.0, GMW v3.0, and GMW v4.0). Building on the GMW v4.0 10m baseline for defining the reference data, the GMW v4.1 timeseries used multiple baseline years (1985, 1990, 1995, 2000, 2005, 2010, 2015, 2020, 2025), with classifications created using an XGBoost classifier and composites of Landsat and Sentinel-2 imagery. Areas of missing data due to the availability of Landsat imagery were copied through from the nearest year. These baselines were visually checked, and for 1985, 1990, and 1995, areas of poor quality were copied using the base quality and the earliest available layer. Subsequently, these baselines were filtered into a timeseries using a Markov chain with probabilities defined for the transitions. To create the annual timeseries, every Landsat and Sentinel-2 scene with less than 70% cloud cover was classified using 130 random forest classifiers in Google Earth Engine and summarised on an annual basis with the number of valid pixel observations and the number of times the pixel was classified as mangroves. These classifications were combined with thresholded JAXA L-band SAR data and the baseline classifications using a likelihood ratio methodology to define a probability of a pixel being mangroves for each year in the timeseries. The subsequent timeseries was then filtered using Bayesian updating and Markov chains, with the output probabilities thresholded to define the final mangrove extent maps. Throughout, extensive visual quality assurance (QA) checks were integrated, and the layers were updated based on the feedback.

The accuracy assessment for the mangrove extent estimated a global F1-Score of 0.9276 (0.9220 - 0.9330), with an omission of 6.357% (5.648 - 7.133) and a commission of 8.093% (7.093 - 9.128). This accuracy assessment was based on 32,569 reference points distributed across 120 geographic sites covering the full range of years within the analysis. Therefore, these accuracy statistics represent the average for the product and will vary locally in both time and geography. Where errors are identified and can be fed back to the developers, those errors can be improved in future versions of the GMW timeseries.

It is important to be aware that the timeseries is dependent on the availability of satellite imagery, and before 1999/2000, the availability of imagery can be significantly reduced. This problem increases the further back in time you go. To provide full coverage on an annual basis throughout the timeseries, this has required some infilling of data from other years, particularly in the 1990s and 1980s.

These data and statistics are available through the Global Mangrove Watch portal (https://www.globalmangrovewatch.org), along with further mangrove related data layers and products, such as blue carbon and restoration potential.