Background and purpose: Monitoring the structural characteristics of mangrove forests is of particular importance for understanding habitat conditions and planning the conservation and management of these ecosystems. Canopy density is one of the important measures of stand structure and mangrove establishment status. Field measurement of this characteristic in tidal environments is associated with limitations such as difficult accessibility, the extensive spatial extent of habitats, and high sampling costs. Sentinel-2 multispectral imagery provides an opportunity for extensive monitoring of these habitats; however, mapping performance can depend on spatial resolution, spectral characteristics of the data, and mangrove patch structure. The main objective of this study was to map the canopy density of the Mahshahr mangrove forests using Sentinel-2 imagery and the Random Forest algorithm. model performance was compared at spatial resolutions of 10 and 20 m, and the most effective bands and spectral indices for distinguishing low-, medium-, and high-density classes were investigated.

Materials and Methods: The study area was located in the Mahshahr mangrove habitats, situated within the Musa Estuary and the Mahshahr Special Economic Zone. A Sentinel-2 surface reflectance image with Level-2A processing level acquired on June 26, 2023, was selected, with low cloud cover and coinciding with low-tide conditions. The mangrove extent was extracted from the 2023 land-cover and land-use map and was used as a common spatial mask. Nine spectral bands and four spectral indices were derived from the Sentinel-2 image and entered into the Random Forest model as predictor variables. The indices included the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index-2 (NDWI2), Mangrove Vegetation Index (MVI), and Combined Mangrove Recognition Index (CMRI). To evaluate the effect of spatial scale on model performance, two datasets were prepared at 10 and 20 m resolutions.

Reference data were generated through visual interpretation of Google Earth Pro imagery acquired on 26 June 2023. Circular sample plots with a radius of 30 m were used to determine the percentage of canopy cover. Values below 30% were defined as low density, 30–70% as medium density, and above 70% as high density. From a 310 collected samples, 280 valid samples—including 59 low-density, 147 medium-density, and 74 high-density plots—were retained for model training after matching with the mangrove mask.

For each spatial resolution, a Random Forest model with 1,500 trees was trained, and its performance was evaluated based on out-of-bag samples. Overall accuracy, the confusion matrix, and variable importance based on the mean decrease in accuracy criterion were extracted. After selecting the appropriate model, the initial map was refined through interpretation of very-high-spatial-resolution imagery.

Results and discussion: The results showed that the Random Forest model performed better at the 20-m resolution than at the 10-m resolution. The out-of-bag overall accuracies of the 10- and 20-m models were 90.18% and 95%, respectively. The better performance of the 20-m model was likely due to closer agreement between the processing scale and the native resolution of the red-edge and shortwave infrared bands and reduced within-class fine-scale variability.

The confusion matrix showed that the low-density class had greater separability, whereas the greatest confusion occurred between the medium- and high-density classes. Variable-importance analysis showed that B5, NDVI, and B4 contributed most to class discrimination, highlighting the importance of vegetation greenness and the red-edge region. Map refinement showed that the greatest spatial uncertainty was concentrated in low-density areas and along patch edges. After refinement, the final mangrove area was estimated at 513.80 ha, of which 63.66% was low density, 31.86% was medium density, and 4.48% was high density. This pattern indicated that, at the time of image acquisition, the Mahshahr habitat was predominantly composed of open and discontinuous canopies.

Conclusion: The integration of Sentinel-2 bands, spectral indices, and the Random Forest algorithm provided an effective approach for mapping the canopy density of the Mahshahr mangroves with suitable accuracy. The 20-m model, with an overall accuracy of 95%, was selected for the final map. The variables B5, NDVI, and B4 made the greatest contributions to distinguishing density classes. The dominance of the low-density class indicated a predominantly open and discontinuous canopy structure in the habitat. Therefore, mapping canopy density using Sentinel-2 can not only represent the spatial pattern but also identify areas with greater uncertainty that require field verification.