Litterfall, the process by which plants shed leaves, flowers, fruits and twigs due to ageing, environmental stress, wind and weathering, is far more than botanical housekeeping. It is the primary conduit through which organic matter returns from plants to soil, feeding decomposers, enriching sediments and driving nutrient cycling. In mangrove ecosystems, litterfall is widely used as an essential measure of productivity and plays a critical role in assessing carbon accumulation rates, organic matter export and nutrient availability. Mangrove detritus is recognised as a significant carbon source for coastal food webs in tropical regions, nourishing the detrital food web and sustaining fish assemblages that in turn attract higher predators such as elasmobranchs. Understanding the quality and quantity of this litter is therefore central to grasping how mangrove forests function, how much blue carbon they sequester and how they will respond to shifting climatic conditions.

The research team selected three mangrove habitats along the coastlines of Ras Al Khaimah, dominated by the resilient grey mangrove Avicennia marina within lagoon systems at Khor, Saraya and Rams, covering a total of 474.85 hectares. The UAE’s climate is classified as hyper-arid, characterised by extremely high summer temperatures and low mean precipitation, with daily maximum temperatures in July and August reaching up to 44 degrees Celsius. From April 2021 to May 2022, the study monitored litter mass through 12 monthly collection events, deploying 34 litterfall traps designed according to the ICP Forests Level II network, a pan-European forest monitoring programme. Each trap consisted of a nylon mesh bag with a pore size of 0.5 millimetres stitched to a square iron frame, placed one metre above the ground beneath mangrove trees to avoid tidal flushing. Three control traps were positioned in areas with no canopy cover. Collected litter was oven-dried at 70 degrees Celsius for 48 hours until constant weight, weighed on a precision balance and ground for carbon estimation through loss-on-ignition at 550 degrees Celsius.

The results revealed a striking seasonal signature. Annual litterfall ranged from 0.107 to 0.768 megagrams per hectare, with the lowest values observed in February and the highest in August. The average monthly litterfall rate was 0.2607 megagrams per hectare, and significant differences amongst monthly rates were confirmed through the Kruskal-Wallis test. Seasonal analysis showed that litterfall production was highest in summer, averaging 0.405 megagrams per hectare, followed by spring at 0.320 and lowest in winter at 0.175. Summer contributed 47.8 percent of the total annual litterfall, spring 32.2 percent and winter just 20.0 percent, with all seasons differing significantly. Across the entire 474.85-hectare study area, annual litterfall was estimated at 139.605 megagrams, with carbon input from litterfall estimated at 44.41 megagrams. The median carbon content per gram of dry litter was 0.319 grams, and average monthly litterfall carbon input was 0.083 megagrams per hectare, significantly higher in summer months at 0.129 compared to 0.056 in winter.

Two scenarios could explain the pronounced summer peak. First, summer is the UAE’s flowering and fruiting season for Avicennia marina, a pattern consistent with studies from Qatar and Kuwait, where flowering typically starts in May and June and reproductive peaks coincide with the hottest months. Second, the plant may respond to water stress during high-temperature periods. Increased evaporation and low rainfall raise salinity levels along the Arabian Gulf coast, impacting conditions in semi-closed lagoons. Under such stress, transpiration becomes metabolically costly, potentially necessitating canopy thinning as an adaptive response. The Gulf’s arid and hypersaline coastal region represents a habitat where most marine species have already reached their tolerance limits, making these mangroves an exceptional case study in plant resilience at the edge of survival.

To link these ground observations to satellite data, the researchers acquired all available Sentinel-2A and Sentinel-2B imagery from March 2021 to February 2022 through Google Earth Engine, filtering scenes with less than five percent cloud cover and computing monthly averages of spectral bands. From these, they derived ten distinct vegetation indices, each targeting a specific aspect of vegetation dynamics: greenness through NDVI, EVI, SAVI and TVI; structure through RVI, LAI and RFDI; water status through MSI and NDMI; and senescence through PSRI. The Normalised Difference Vegetation Index quantifies vegetation greenness and vigour, while the Normalised Difference Moisture Index tracks vegetation water content using near-infrared and shortwave-infrared reflectance. The Plant Senescence Reflectance Index indicates vegetation stress and senescence through the pigment balance between chlorophyll and carotenoids. All index rasters were standardised to a uniform 10-metre resolution, litterfall data were interpolated to match the satellite pixel size, and 315 point values were extracted, correlated pixel-by-pixel with the index rasters over 12 months, yielding 37,800 observations.

Spearman rank correlation analysis revealed generally weak linear relationships between litterfall and the vegetation indices. RVI showed a very weak positive correlation of 0.13, while LAI and EVI were negatively correlated at -0.46 each, and SAVI at -0.35. Moderate positive correlations were observed for TVI at 0.44 and NDMI at 0.39, while PSRI exhibited a weak positive correlation of 0.17. The pairwise correlation matrix amongst the indices themselves revealed several strongly intercorrelated pairs, including LAI-EVI at 0.999 and RVI-RFDI at 0.998, indicating high redundancy due to shared spectral sensitivity to canopy greenness, a natural consequence of indices derived from similar multispectral bands. Predictors with correlations of 0.8 or above were filtered to minimise multicollinearity, retaining the index with the stronger correlation to monthly litterfall.

Because simple correlation could not capture the complex, non-linear interdependencies at play, the team turned to random forest regression, an ensemble learning method that generates large quantities of relatively uncorrelated regression trees using randomly assigned subsets of features. The dataset was split into 75 percent training and 25 percent testing sets, and hyperparameters were optimised using the Out Of Bag root mean square error from the training data. The full model demonstrated strong predictive performance, with an initial R² of 0.78 and an RMSE of 6.71, improving to R² = 0.81 following tuning with RandomizedSearchCV. The optimal configuration used 1000 trees with three features considered per split. A reduced six-index model achieved nearly identical performance, with a test R² of 0.78 improving to 0.79 after tuning. In the full model, NDMI accounted for the highest variable importance at 0.264, followed by PSRI at 0.203 and TVI at 0.156, though the reduced model produced a different ranking, indicating that importance estimates are sensitive to the composition of the predictor set.

The dominance of NDMI, PSRI and TVI carries ecological meaning. NDMI has recorded strong correlations with field measures of vegetation parameters, including plant area index and canopy gap fraction, underscoring its effectiveness in monitoring vegetation health and moisture content. PSRI has proven to outperform NDVI in tracking canopy senescence dynamics, providing clear advantages for detecting early senescence by assessing the chlorophyll-to-carotenoid ratio in leaves and fruits. That moisture and senescence indices, rather than classic greenness measures, proved most informative suggests that litterfall in these hyper-arid mangroves is governed more by water stress and pigment breakdown than by canopy vigour alone. Random forest algorithms are less vulnerable to the multicollinearity plaguing these datasets, as bootstrap aggregation and randomised feature selection at each split reduce inter-tree correlation, allowing the inclusion of multiple partially redundant indices to enhance the model’s ability to capture subtle spectral variations.

The study is not without limitations. Random forest regression cannot predict outcomes outside the range of training data, meaning pest outbreaks or extreme weather producing litterfall beyond historical records would elude the model. The approach is region-specific, as both the remote-sensing indices and ground-truth litterfall data were derived from local conditions, and previous studies of random forest spatial transfer indicate that accuracy losses can be substantial when moving to unseen regions or resolutions. Reliable replication elsewhere would require compatible vegetation-index layers, local ground-truth data for calibration, retraining under region-specific canopy, tidal and edaphic conditions, and thorough documentation of preprocessing steps. Without appropriate spatial or temporal validation, performance metrics may be overly optimistic. Nevertheless, the findings provide a basis for developing a future framework for large-scale litterfall assessment, one that could facilitate broader analyses of carbon and nutrient fluxes and ecosystem health by integrating field measurements, spatial analysis and remotely sensed vegetation indices. As the IPCC projects with high confidence that heatwaves will intensify in the Arabian Gulf, such satellite-based monitoring of the region’s mangroves may soon become indispensable for conservation amid shifting climatic conditions.

Subject of Research: Litterfall dynamics and remote sensing-based carbon modelling in hyper-arid Arabian Gulf mangrove forests

Article Title: Seasonal dynamics of litterfall production and carbon input in arid mangroves: Modelling aboveground litterfall in Arabian Gulf mangroves using remotely sensed and ground data

Article References: Gopi, K., Ksiksi, T., & Henderson, A. C. (2026). Seasonal dynamics of litterfall production and carbon input in arid mangroves: Modelling aboveground litterfall in Arabian Gulf mangroves using remotely sensed and ground data. One Ecosystem, 11, Article e146607. https://doi.org/10.3897/oneeco.11.e146607

Image Credits: AI Generated

DOI: 10.3897/oneeco.11.e146607

Keywords: mangroves, litterfall, blue carbon, remote sensing, Sentinel-2, random forest, Avicennia marina, Arabian Gulf, vegetation indices, carbon input, United Arab Emirates, machine learning