Next Article in Journal
Exploring Impacts of Land Use and Cover Changes on Ecosystem Services on the Qinghai-Xizang Plateau
Previous Article in Journal
ω − k MUSIC Algorithm for Subsurface Target Localization
Previous Article in Special Issue
Assessing Mangrove Forest Recovery in the British Virgin Islands After Hurricanes Irma and Maria with Sentinel-2 Imagery and Google Earth Engine
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 17
Issue 16
10.3390/rs17162837
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Characterizing Role of Spatial Features in Improving Mangrove Classification—A Case Study over the Mesoamerican Reef Region

by
Suvarna M. Punalekar
1,2,*,
A. Justin Nowakowski
2,
Steven W. J. Canty
2,
Craig Fergus
1,
Qiongyu Huang
1,
Melissa Songer
1 and
Grant M. Connette
1
1
Conservation Ecology Center, Smithsonian’s National Zoo and Conservation Biology Institute, Front Royal, VA 22630, USA
2
Smithsonian Environmental Research Center, Edgewater, MD 21037, USA
*
Author to whom correspondence should be addressed.
Remote Sens. 2025, 17(16), 2837; https://doi.org/10.3390/rs17162837
Submission received: 11 July 2025 / Revised: 4 August 2025 / Accepted: 5 August 2025 / Published: 15 August 2025
(This article belongs to the Special Issue Remote Sensing in Mangroves IV)
Browse Figures
Review Reports Versions Notes

Abstract

Mangrove forests are among the world’s most vital coastal ecosystems. Mapping mangrove cover from local to global scales using spectral data and machine learning models is a well-established method. While non-spectral contextual datasets (spatial features) have also been incorporated into such models, the contribution of these additional features to improving mangrove mapping remains underexplored. Using the Mesoamerican Reef Region as a case study, we evaluate the effectiveness of incorporating spatial features in binary mangrove classification to enhance mapping accuracy. We compared an aspatial model that includes only spectral data with three spatial models: two included features such as geographic coordinates, elevation, and proximity to coastlines and streams, while the third integrated a geostatistical approach using Inverse Distance Weighted (IDW) interpolation. Spectral inputs included bands and indices derived from Sentinel-1 and Sentinel-2, and all models were implemented using the Random Forest algorithm in Google Earth Engine. Results show that spatial features reduced omission errors without increasing commission errors, enhancing the model’s ability to capture spatial variability. Models using geographic coordinates and elevation performed comparably to those with additional environmental variables, with storm frequency and distance to streams emerging as important predictors in the Mesoamerican Reef region. In contrast, the IDW-based model underperformed, likely due to overfitting and limited representation of local spectral variation. Spatial analyses show that models incorporating spatial features produced more continuous mangrove patches and removed some false positives in non-mangrove areas. These findings highlight the value of spatial features in improving classification accuracy, especially in regions with ecologically diverse mangroves across varied environments. By integrating spatial context, these models support more accurate, locally relevant mangrove maps that are essential for effective conservation and management.

1. Introduction

Mangrove forests are crucial coastal wetland ecosystems located in tropical and subtropical regions, providing key ecosystem services such as coastal protection, carbon sequestration, and biodiversity support [1,2,3,4,5]. Recognizing their importance for the well-being of coastal communities and beyond, global and national policies have been implemented to conserve and restore these forests [6,7]. Remote sensing-based mapping, especially at the global scale, has been crucial in monitoring the loss and recovery of these ecosystems over the last three decades [8,9,10,11,12].
Recent advancements in satellite sensor technology, improved spatial and temporal resolutions of data, and cloud-based platforms hosting analysis-ready satellite data (e.g., Google Earth Engine) have led to significant enhancements in global mangrove data products [10,12,13,14]. Several regional and national maps have been developed which account for structural as well as species-specific differences in canopies [15,16,17]. National or regional maps are often used for local decision-making processes, due to their localized calibration and alignment with national definitions of mangroves.
Land-cover classifications, including mangrove extent mapping, both globally and regionally, have relied heavily on machine learning algorithms [8,10,18,19,20]. These approaches are typically implemented with spectral bands and indices (referred to collectively as spectral features in this paper) from optical and microwave satellite sensors. In addition to these datasets, non-spectral ancillary datasets (referred to as spatial features) have also been utilized to provide spatial context to classification models. These spatial features are either directly integrated in model development or used for post-processing of maps into more refined and detailed versions of habitat types. Examples of these features include geographic coordinates of locations in the training data [21,22,23], proximity to other geographic features [24], and spatial lags or texture matrices [25,26,27]. Datasets characterizing environmental conditions like climate, elevation, slopes, and biogeography have also been utilized as contextual information in remote sensing-based classifications [28,29,30,31,32]. These features are included in classification with the intention to support algorithms like Random Forest to identify conditions where a target land-cover class may exhibit spatial variation. The additional information embedded in spatial features can enhance classification models by capturing similarities and differences within spectral characteristics, especially across certain distance-based gradients.
Studies also recommend using these non-spectral features to capture spatial relationships within spectral data, particularly those influenced by spatial autocorrelation—the tendency for nearby locations to exhibit similar values for a given variable [23,33,34,35]. In this context, some studies have incorporated spatial features derived from geospatial interpolation methods such as Inverse Distance Weighting (IDW) and Kriging to represent spatial relationships within training data. For example, integration of spatially interpolated spectral reflectance values alongside actual reflectance was found to increase the classification accuracies for tree species [36].
Despite many studies utilizing ancillary datasets in classification models, their actual impact on model performance—including in mangrove extent classification—has not been thoroughly examined. In case of binary classifications, wherein the target class shows inherently patchy distribution, training data can be highly clustered. This may result in maps with highly clustered localized errors if spatial variation across the target class is not captured [37]. It remains to be tested whether having additional spatial variables alleviate such localized errors, especially under limited training data situations.
Mangrove forests, despite their simpler canopies compared to many other tropical forests, exhibit significant variation in structure and distributions that are influenced by factors such as tidal dynamics, salinity and sediment gradients, storm frequency, water quality, herbivory and competition [38,39,40]. The distributions of different mangrove and non-mangrove species across coastal zones are influenced by multiple environmental gradients, including distance from coast or tidal range. Considering these local variations can be crucial for mapping mangrove extent as well as biophysical attributes [32,41,42]. Impacts of nutrient enrichment and frequent storm disturbances on mangrove canopies and their spectral reflectance have been captured in various studies [43,44,45]. In mangrove mapping, elevation and proximity to coast are among the most used spatial features [12,14,16,46]. Other spatial datasets, such as distance to freshwater streams and storm frequencies, however, have not been tested widely in classification schemes. The main concerns, in general, with using spatial features are the risks of model overfitting and the limited generalizability of the model outside the defined geographic range [47,48].
This study was undertaken to evaluate the effectiveness of various non-spectral spatial features in improving mangrove classification models. Given the local variation dynamics, mangroves make for a good case study in assessing the impact of spatial features for classification modeling. We focused on different sets of spatial variables and one geostatistical method—IDW—due to their relatively straightforward implementation within Google Earth Engine [49]. We hypothesized that incorporating these spatial features would enhance model accuracy. To test this, we examined the robustness of each feature set under data-sparse conditions by systematically reducing the size of the training dataset. Additionally, we assessed the spatial autocorrelation in the training data and model errors to explore whether spatial dependencies in the training data influenced model performance across the different feature sets. Our case study was conducted in the Mesoamerican Reef region, chosen for its ecological significance and the structural diversity of its mangrove forests.

2. Materials and Methods

2.1. Study Area

The Mesoamerican Reef (MAR) region (Figure 1) stretches for over 1000 km, beginning at the Yucatán Peninsula in Mexico (21.56°N; 087.09°W) and extending to the northeastern coast of Honduras (14.97°N; 083.16°W), encompassing the coastlines of Belize and Guatemala. This region harbors key coastal ecosystems, including mangroves, seagrasses, and the largest coral barrier reef in the western hemisphere [50,51,52]. The coastline features a variety of habitats, such as estuarine riverbeds, creeks, atolls, lagoons, and beaches. Along the shores and offshore islands, mangrove ecosystems are found either as dense, continuous canopies or in fragmented patches, sometimes forming fringes along coastlines and river channels. The dominant species in these mangrove forests are Rhizophora mangle (red mangrove), Avicennia germinans (black mangrove), and Laguncularia racemosa (white mangrove). In the case of Honduras and Guatemala, the mangrove-associated species Conocarpus erectus (buttonwood mangrove) is functionally recognized as a mangrove species [53]. The mangrove forests vary with heights ranging from short canopies (under 2 m) to medium (2 to 5 m) and tall trees (5 to >30 m). In some inland areas of eastern Mexico and northern Belize, distinctive sparse mangrove shrubs are intermixed with other coastal wetlands.
The mangroves in the MAR region are an important source of natural resources and livelihoods for communities along the coastline. These ecosystems are crucial for commercial fisheries and tourism, as well as for natural protection against increasing threats of tropical cyclones [45]. Nevertheless, these mangrove ecosystems are increasingly under threat from human impacts (e.g., coastal development) and environmental (e.g., sea level rise) stressors [53], which modify the extent as well as health of these forests. Recognizing the critical role of mangroves, national and regional policies are being developed and implemented for better management and monitoring [55].
For this mangrove mapping project, we considered all islands as well as areas within 20 km inland from the coastline, and at an altitude less than 100 m above sea level. While most mangroves in the region are distributed within ~5 km of the coastline, mangrove shrubs and fringes bordering rivers and other waterways are found much farther inland. The coastline was defined using the public Global Self-consistent, Hierarchical, High-resolution Geography (GSHHG) layer [56]. Similarly, the offshore area covered by open ocean was masked out using Normalized Difference Water Index (NDWI) while keeping all the islands within the region of interest. A threshold of 0.1 was set on the median composite of NDWI (January to June, 2019 and 2020) to mask out water.

2.2. Machine Learning Model Development

We used the Random Forest implementation in the Google Earth Engine platform for the classification model. Random Forest is a commonly used algorithm for landcover classifications using Earth observation datasets due to its flexibility to work with highly correlated variables and its non-parametric nature [57,58,59]. We primarily used the Sentinel-2 and Sentinel-1 Earth observation satellite datasets at 10 m spatial resolution (Table 1). The extent maps were developed for the composite period of 2019–2020. Figure 2 illustrates the technical workflow implemented. Further details about the components of the algorithm development are described below.

2.2.1. Generation of Calibration Dataset

The model calibration dataset was developed by identifying areas of mangrove forest based on the locations of field surveys conducted in September 2021 [60], as well as visual assessment of high-resolution satellite imagery, particularly from Google Earth and Bing Maps, and multi-year false color composites (SWIR-NIR-Red) from Sentinel-2 data. GPS coordinates from the field surveys, which focused on soil and vegetation sampling, served as reliable indicators of mangrove-dominated canopies. However, due to the limited number of GPS points and their patchy, clustered distribution, it became necessary to supplement the training data with additional reference locations derived from high-resolution satellite image view (Google Earth and Bing Maps). Polygons were drawn around the GPS points and in areas with predominantly mangrove cover. Similarly, polygons were drawn in non-mangrove regions to capture a range of other land-cover types. Random points were generated within each polygon, with a maximum of 30 points per polygon, while ensuring that no two points were closer than 30 m apart. This approach increased the number of training points while preventing the selection of adjacent pixels, ultimately resulting in a total training dataset of 2880 mangrove and 2737 non-mangrove pixels. Care was taken to ensure that the training data for both mangrove and non-mangrove classes were evenly distributed across the geographic extent of the study area. The mangrove training points were also used to study the pattern of spatial autocorrelation across the MAR region using correlograms.

2.2.2. Feature Sets for Model Development

Different feature spaces (Table 1, Figure 2) were defined for model development to evaluate the impact on classification accuracies. The Non-spatial feature set consisted solely of satellite sensor-derived inputs, representing a feature space where no additional features were used. It included median reflectance values for specific Sentinel-2 bands [61], covering the period from January to June in both 2019 and 2020. The first half of the year was chosen given the predominantly cloud-free conditions in the region. Additionally, four spectral indices (Table 1) were used during the same period to capture the distinct spectral characteristics of mangroves in the Short-Wave Infrared (SWIR), red-edge, and Near-Infrared (NIR) bands. These indices were intended to help differentiate mangroves from other land-cover types, such as terrestrial vegetation, other wetlands, urban areas, bare surfaces, and water bodies. To account for vegetation phenology and the tidal cycle’s effect on mangrove reflectance, the 80th and 20th percentiles of these composite indices were utilized while excluding noisy pixels. Finally, we also included synthetic aperture radar data from the Sentinel-1 satellite. Sentinel-1 has dual polarization mode with one band having vertical transmission and horizontal reception, while the other has vertical transmission and vertical reception. We generated median composites of backscatter values from both Vertical–Horizontal (VH) and Vertical–Vertical (VV) polarizations [62]. The significance of backscatter data has been particularly noted in distinguishing rough mangrove canopies from shorter wetlands, bare surfaces, and water bodies [10,63,64].
Table 1. Description of four feature sets with key characteristics and references associated with each feature.
Table 1. Description of four feature sets with key characteristics and references associated with each feature.
Feature SetSourceKey Characteristics/ProcessingReferences
Non-spatialSentinel-2 surface reflectance bands (Level-2A)
(blue, green, red, rededge-1, rededge-2, rededge-3, NIR, rededge-4, SWIR-1, SWIR-2)
GEE collection: COPERNICUS/S2_SR_HARMONIZED		Median composite calculated with all scenes captured from January to June of 2019 and 2020 and having less than 30% cloud cover.
Spatial resolution: 10 m		Bunting et al. [11,12]
Cissell et al. [46]
ESA, Copernicus [61,62]
Hu et al. [63]
Yancho et al. [14]
Normalized Difference Vegetation Index (NDVI),
Red-Edge Chlorophyll Index (CLrededge),
Normalized Difference Water Index (NDWI),
Normalized Difference Moisture Index (NDMI)		80th and 20th percentiles composite calculated over January to June of 2019 and 2020.
Spatial resolution: 10 m
Sentinel-1 VH and VV backscatter
GEE collection: COPERNICUS/S1_GRD		Median composite calculated over January to June of 2019 and 2020, and with all scenes captured in ascending orbit and interferometric mode. Noisy pixels (backscatter < −30 dB) were removed.
Spatial resolution: 10 m
Spatial-1All features in non-spatial set
Latitude and Longitude for pixel center Farr et al. [65]
Bunting et al. [11,12]
Cissell et al. [46]
NASA, USGS, JPL Caltech [66]
Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM)
GEE collection: Version 3		Original ~30 m resolution data resampled to 10 m.
Spatial-2All features in non-spatial set + DEM
Distance to Coast (CoastD)
Coastline defined by GSHHG (Version 2.3.7 Full Resolution, Level 1)		Bi-directional Euclidean distance in m from the coastline.
Spatial resolution: 100 m,
resampled to 10 m.		Wessel and Smith [56]
Distance to Streams (StreamD)
Streams defined by WWF Free Flowing Rivers 		Bi-directional Euclidean distance in m from streams (river order 6 or lower).
Spatial resolution: 100 m,
resampled to 10 m.		Grill et al. [67]
Storm frequency (Storms)
Storm track line data from IBTrACS		Number of occurrences of cyclones between 1990 and 2016 recorded as density of storm track lines per km.
Spatial resolution: 1 km,
resampled to 10 m.		Knapp et al. [68]
Spatial-3All features in non-spatial set
Difference between IDW interpolated surface and actual median surface reflectance for Sentinel-2 bands. These features were named with ‘Dif’ appended to the band name (e.g., BlueDif, GreenDif).IDW interpolation was obtained at 100 spatial resolution, resampled to 10 m.Johnson et al. [36,69]
The Spatial-1 feature set included the previous Non-spatial features but also incorporated the location-specific features latitude, longitude [70], as well as elevation. In the Spatial-2 feature set, spatial features were chosen to characterize the environmental space across the region. These included proximity to the coast, distance from rivers and streams, and storm frequency, while also maintaining elevation from Spatial-1 and all spectral features as in Non-spatial set. Mangroves in areas that experience frequent storms often exhibit shorter growth compared to those in more sheltered regions [38]. This structural variation could result in significant differences in spectral reflectance and backscatter among types of mangroves with differing physical characteristics [10,71]. The Spatial-3 feature set followed the methodology outlined by Johnson et al. [36]. In this approach, interpolated median mangrove spectral reflectance surfaces were generated for each Sentinel-2 band (Table 1) using all mangrove training pixels. The interpolation was performed at a 100 m spatial resolution using the IDW technique within Google Earth Engine. Final feature values in this set were derived by calculating the differences between the actual median reflectance values and these interpolated surfaces. The actual median reflectance bands, along with other Sentinel-1 and -2 based features, were retained as in Spatial-1 and -2 feature sets (Table 1).

2.2.3. Spatial Cross-Validation Strategy for Model Calibration

A spatial cross-validation approach was used to evaluate the potential impact of geographic biases in the training data on model performance. Given the concern that using features such as geographic coordinates can introduce inaccuracies when models are applied to new areas [47,48], we implemented this approach to assess the extent to which such biases may affect model accuracy. This was performed using a grid-based method, where a 20 × 20 km grid was superimposed over the study area, and each training point was assigned a grid cell identity based on its location. There were a total of 116 grid cells. For each iteration, all points from one grid cell were left out, while the model was trained on points from the remaining cells. The excluded points were then classified by the trained model. Once all grid cells were left out, calibration accuracies were calculated. This process was repeated for each feature set (Table 1). During each iteration of spatial cross-validation, feature importance was evaluated using the default method in Google Earth Engine, which measures the decrease in Gini Impurity Index [58], with random feature selection within decision trees. The number of trees (ntrees) for the Random Forest algorithm was set to 500. Final models for each feature set were trained on the full dataset and subsequently used to generate the final maps.

2.3. Map Accuracy Assessment

The approach we followed for map accuracy assessment is comparable to other mangrove mapping studies [12,46]. We generated 6000 random points for the map accuracy assessment, which were distributed in a 2:1 ratio (non-mangrove to mangrove classes), using the Spatial-2 feature set. We used the ACaTaMa plugin (version 24.6) [72] in QGIS (version 3.40.9) [73] to assign ‘true’ labels to these points, by inspecting high resolution Google Earth and Bing Maps imagery, as well as False Color Composites generated from multiple-year (2019, 2020, 2021) Sentinel-2 images. We were able to assign labels to 5932 points in total, while the remaining 68 points were left unlabeled due to uncertainty in their identification. Finally, a confusion (error) matrix [74] was generated using these 5932 points, and overall accuracies, omission errors, and commission errors were calculated for maps produced by models built with each feature set.

2.4. Effect of Data Scarsity on Map Accuracies

We evaluated the effectiveness of the feature sets for model development under conditions of sparse training data by systematically reducing the size of the training datasets. This was achieved by using the 116 grid cells as devised for the spatial cross-validation model building strategy described in Section 2.2.3. From the 116 cells, 20 to 100 grid cells were randomly sampled with intervals of 20. Each selection model was built using all the points within the chosen grid cells. The performance of this model was assessed using an independent validation dataset of 5932 randomly distributed and ‘ground truth’ points (see Section 2.3). We used the Random Forest algorithm from the SciKit-Learn library [75] within Python Conda environment (Python 3.9.13) for this experiment. The parameter ntrees was set to 500, consistent with the setup in the Google Earth Engine implementation.

3. Results

3.1. Model and Map Accuracy Assessment

The overall accuracy for model calibration following the spatial cross-validation approach (Section 2.2.3) and for map validation (Section 2.3) were both greater than 90% (Table 2). The overall accuracy for the model calibration step ranged from 2% to 3% lower than the corresponding overall accuracy based on map validation, except for Spatial-3, where it was 4% higher than the map validation.
The most striking result was that Spatial-3 (IDW-based approach) had the lowest model calibration errors (1% omission error, 3% commission error) but the highest omission errors (17%) for map validation. The Non-spatial set had a comparable commission error to other feature sets (3% for map validation), but a 7% omission error. Spatial-1 (with geographic locations and elevation) performed best, considering both omission and commission errors corresponding to map validation.

3.2. Feature Importance

The feature importance plots (Figure 3) showed that the SWIR-1 band was the most important feature across all the feature sets used. The SWIR-2 band, as well as the SWIR-1-based spectral index NDMI, were also among the most important features in all sets. In the case of Spatial-1, latitude and longitude were identified as the second and fifth most important features, respectively. Similarly, for Spatial-2, storm frequency (Storms) and distance to streams (StreamD) were recognized as being in the top five most used features in the model. Distance to coasts (CoastD) was ranked as less important compared to Storms and StreamD.
In the case of Spatial-3, all the difference bands (calculated as the difference between IDW-interpolated median reflectance and actual median reflectance) were ranked higher in feature importance than the actual median reflectance. SWIR-1 and SWIR-2 were the exceptions; however, the corresponding difference bands were marked as important features, indicating that in the decision tree building, it could be the combinations of these features (both difference and actual) that played a significant role.

3.3. Effect of Data-Size on Map Accuracies

The average omission errors improved considerably with increasing data size (i.e., the increasing number of grid cells) for all feature sets, except Spatial-3 (Figure 4). In the case of this feature set, the mean error remained around 17% despite the increase in grid cells up to 100. Spatial-1 and Spatial-2 had very similar omission errors (1 to 2% difference) when the number of grid cells was set to 70 and above, and these errors were marginally lower than the Non-spatial set. In general, omission errors fluctuated less with the increase in number of grid cells for all feature sets.

3.4. Comparing Spatial Feature Output Maps with Non-Spatial Feature Output

Compared to the map generated using the Non-spatial feature set, those with spatial features showed both greater and lower numbers of pixels classified as mangroves across different areas. We summarized differences between spatial and Non-spatial feature sets in the area (ha) of mangrove cover within 10 × 10 km grid cells, separately as negative (Figure 5) and positive (Figure 6) differences in mapped mangrove area. Spatial-1 and Spatial-2 showed similar spatial patterns in terms of additions or removal of mangrove pixels with respect to the non-spatial map. In contrast, Spatial-3 showed greater reductions in mangrove pixels, especially in Belize and Mexico. Spatial-3 had consistently more grid cells showing reductions above 25 ha per 100 km2 compared to the other two sets (Figure 5). Using all three spatial feature sets removed some misclassifications in inland forested areas (Figure 7a–c). However, for Spatial-3, in comparison with Spatial-1 and 2, some big patches of mapped mangrove forests also shrunk considerably leaving out areas which could potentially be under sparse cover (Figure 7d–f).
When examining area increases as a function of feature set, Spatial-3 added the least area compared to the Non-spatial map (Figure 6). Spatial-1 showed more grid cells with increases above 25 ha per 100 km2 than the other sets. Additions mostly remained below 100 ha per 100 km2 area for all sets. The largest additions for all sets were primarily in areas close to the border between Mexico and Belize. Some of these areas have sparse cover of shrubby mangroves. Reductions along the coasts of southern Belize, Guatemala, and Honduras were generally below 25 ha across all feature sets. Additions of mangrove pixels mapped in model sets with spatial features were mainly observed around the patches of mangroves mapped in the Non-spatial set. Spatial-1 and Spatial-2 especially captured more sparse cover mangroves in the marginal areas between mangroves and non-mangroves (Figure 7g–h). These expansions were generally much thinner or even absent for Spatial-3 (Figure 7i).

3.5. Spatial Autocorrelation in Spectral Data and Model Errors

The degree of spatial autocorrelation across mangrove training points was quantified using spline correlograms [76]. A spline correlogram is a spatial correlation plot which accounts for distance between pairs of points. It is a generalization to a continuous function of the typical binned spatial correlogram based on Moran’s I [33,77]. The spline- and spatial correlograms are asymptotically equivalent [76]. The spline correlograms show that there was a higher degree of positive correlation in mangrove surface reflectance values among the training points close to each other (<50 km), and there was positive spatial autocorrelation at distances up to 250 km for most bands (Figure 8). For mangrove points far apart (>300 km), the correlation in Green and Red bands dropped below −0.5. For NIR and SWIR-1 bands, negative correlation was observed but remained above −0.5 in general. Lower negative values particularly in Red and Green bands implied that some patches of mangroves were distinctly different in terms of spectral reflectance in these bands. These could potentially be the open, scrubby mangroves found inland away from the coast in Yucatán and Northern Belize, in contrast to dense, continuous coastal mangrove patches in the southern part of the study region. The observed correlation patterns could be due to highly localized variations across mangrove canopies, which manifested differently for each band [41]. In general, spatial autocorrelation was an important factor for locations within ~50 km from each other and for most spectral bands.
To assess whether the spatial autocorrelation in spectral data influenced the model predictions, we also plotted spline correlograms for the residual errors obtained from each model (Figure 9). For this, the residual probabilities were calculated for all validation points by assuming a mangrove probability of 1 for a validated true mangrove point. Deviance residuals were calculated from raw residuals to account for the binomial distribution of errors resulting from a binary classification model. None of the models, including the Non-spatial model, had pervasive autocorrelation in the residuals, as was seen in the reflectance values for the training data (Figure 8). The results indicate that in neither of the maps were the error points clustered in certain geographical areas.

4. Discussion

4.1. Role of Spatial Features in Improving Classifications

Mangrove cover maps within the MAR region improved when using spatial features, compared to a Non-spatial model which lacked such features. The incorporation of geographical coordinates and elevation data (Spatial-1), and environmental variables (distance to coast, distance to stream, storm frequency and elevation in Spatial-2) resulted in mangrove cover maps with lower omission errors, while commission errors remained virtually unchanged (Table 2). The omission errors in Spatial-1 were 4% lower than the Non-spatial model. For Spatial-2 the difference was 2%. Spatial-1 and Spatial-2 produced similar maps overall. The reduction in omission errors, compared to the Non-spatial model, suggests that spatial features provided local context required to account for spectral variations across the region.
Features such as geographical coordinates, elevation, and coastline datasets have previously been utilized in mangrove mapping [10,14]. However, our analysis revealed that storm frequency and distance to rivers and streams were also valuable features in the study region (Figure 3). This finding indicates an important influence of storm disturbances and freshwater inflows on mangrove structure and composition within the MAR region. While storm severity is widely recognized as a major factor driving mangrove loss and productivity [43,78], our study also highlights the utility of storm frequency in mapping mangrove extent. However, distance to the coast was less important than the other two features in Spatial-2, possibly because elevation was also included in the model. Furthermore, mangrove fringes and open scrubby mangroves in the MAR region are often found much further inland.
The local context provided by the additional features included in the RF algorithm allows the classification to better account for the local similarities among mangrove (and non-mangrove) pixels [70,79]. The role of spatial features in contextual modeling varies depending on the type of feature [70]. For environmental features, local context represents specific conditions, such as areas prone to frequent storms having dwarf canopies compared to those in sheltered areas [80]. In contrast, geographic coordinates account for proximity among pixels, helping to directly capture similarity of pixel characteristics in multi-dimensional space [70]. In both cases, for Spatial-1 and Spatial-2, the features have mainly been useful to resolve confusions in the transition zones between mangrove and non-mangroves (Figure 7).
In contrast to Spatial-1, Spatial-2, and the Non-spatial model, Spatial-3 performed poorly. In principle, IDW-based variables were expected to assist the classification in a similar way to the incorporation of geographic coordinates, by capturing similarities among pixels in close proximity. It has indeed been shown that including an IDW-based spectral difference matrix improved tree species mapping in forestry blocks [36,69]. However, this approach was ineffective for mangroves. This discrepancy may stem from the diverse land-cover types surrounding mangroves and the structural variability within mangrove canopies. The spectral differences derived from IDW-interpolated mangrove reflectance may not have adequately captured the localized spatial pattern in mangrove reflectance.
Our assessment of map accuracy (Table 2) and visual appraisal of maps (Figure 5, Figure 6 and Figure 7) provided deeper insights than the spatial cross-validation model accuracy assessment. While spatial cross-validation accuracy did not reveal the suboptimal performance of Spatial-3, map validation clearly exposed high omission errors. This finding supports the argument by Wadoux et al. [81] that spatial cross-validation strategies may not accurately reflect overall performance across an entire region of interest. While spatial cross-validation can be useful for evaluating spatial autocorrelation in model calibration accuracy, it does not necessarily address how to represent it within the model-building process unless additional measures, such as incorporating spatial context, are implemented.

4.2. Nature of Modifications in the Mangrove Map with Additional Spatial Features

Spatial comparisons of the maps revealed that areas classified as mangroves in Spatial-1 and Spatial-2 were essentially expanded or dilated versions of mangrove patches identified in the Non-spatial model (Figure 7). Conversely, some small non-mangrove areas that were mistakenly classified as mangroves in the Non-spatial model were corrected in the spatial models. These alterations resemble post-classification refinement techniques, such as morphological filtering [82]. However, unlike purely mathematical operations, the integration of spatial features introduced meaningful physical context to these modifications. These changes stem from the additional information provided in transitional or mixed pixels between mangrove and non-mangrove land-cover types. This finding is consistent with observations by Dobbertin and Biging [27], who noted that incorporating texture was especially beneficial in areas with low spatial autocorrelation. In such transitional zones, the marginal value added by spatial features becomes particularly significant. Nevertheless, given the low spatial autocorrelation in model residuals across all feature sets (Figure 9), it can be concluded that spatial autocorrelation was not a major confounding factor in any of the tested models. All models that adequately captured local and regional mangrove distributions were able to mitigate the potential influence of spatial autocorrelation.
Our findings also underscore the critical importance of training data in effective mangrove mapping when using spatial features (Figure 4). Models trained on limited or geographically biased datasets showed similar accuracies regardless of the inclusion of spatial features. This supports earlier research emphasizing the role of training data characteristics in classification outcomes [27,32,83]. Without a robust and well-distributed training dataset, none of the tested feature sets consistently produced maps with lower omission errors than the Non-spatial model. Spatial features can effectively capture localized spectral variation only when supported by adequate and representative training data.
While the overall improvement in map accuracy with the addition of spatial features in Spatial-1 and Spatial-2 was modest, their influence was pronounced in specific localities. The spatial difference maps showing additions and removals of mangrove pixels revealed that classification outcomes in certain regions were more responsive to spatial features (Figure 5 and Figure 6). These localized changes may be particularly relevant for studies focused on fragmentation or habitat suitability, especially when the analysis area is smaller than the full extent of the mapped region.

5. Conclusions

In this study, we evaluated the effectiveness of various feature sets for classifying mangrove cover in the MAR region. Our results demonstrate that incorporating spatial features—such as geographic coordinates, elevation, distance to streams, and storm frequency—can improve mangrove classification compared to models relying solely on spectral data. These improvements were most apparent in the reduction in omission errors. Although overall changes in map accuracy were modest, the inclusion of spatial features led to substantial localized adjustments in mangrove classification across the region. We also caution against uncritical use of geospatial techniques, such as IDW interpolation, in mangrove extent mapping. Our models for this method appeared overfitted, resulting in overly constrained mangrove extents relative to other outputs. While addressing spatial autocorrelation is important—particularly to avoid spatial bias in model predictions—overcorrecting with methods like IDW can be counterproductive. Lastly, we underscore the critical role of training data in building representative and reliable models. Without well-distributed and comprehensive training datasets, spatial feature sets may fail to yield consistently accurate or generalizable classification outcomes.

Author Contributions

Conceptualization, S.M.P., G.M.C.; methodology, S.M.P., G.M.C.; software, S.M.P.; validation, S.M.P.; formal analysis, S.M.P., A.J.N.; investigation, S.M.P.; resources, S.W.J.C., G.M.C.; data curation, S.M.P., C.F. and A.J.N.; writing—original draft preparation, S.M.P.; writing—review and editing, S.W.J.C., G.M.C., A.J.N., C.F., Q.H., and M.S.; visualization, S.M.P., A.J.N.; supervision, G.M.C., S.W.J.C.; project administration, G.M.C.; funding acquisition G.M.C., S.W.J.C., A.J.N., C.F., Q.H. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the NASA Land-Cover and Land-Use Change (LCLUC) Program [grant number: 80NSSC23K0527].

Data Availability Statement

The developed maps, training datasets, and Google Earth Engine will all be open access and available on the FigShare repository 10.6084/m9.figshare.28396094.

Acknowledgments

We are very grateful to Hannah Morrissette from the Smithsonian Environmental Research Center for providing the ground survey locations for this research. We also thank Susan Kotikot (Penn State University, USA) and Jordan Cissell (Samford University, USA) for sharing their experience with mangrove mapping within Google Earth Engine platform. We are very thankful to Pete Bunting (Aberystwyth University, UK) for providing feedback on the accuracy assessment strategy.

Conflicts of Interest

The authors declare no conflicts of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

References

  1. Donato, D.C.; Kauffman, J.B.; Murdiyarso, D.; Kurnianto, S.; Stidham, M.; Kanninen, M. Mangroves among the Most Carbon-Rich Forests in the Tropics. Nat. Geosci. 2011, 4, 293–297. [Google Scholar] [CrossRef]
  2. Carrasquilla-Henao, M.; Juanes, F. Mangroves Enhance Local Fisheries Catches: A Global Meta-Analysis. Fish Fish. 2017, 18, 79–93. [Google Scholar] [CrossRef]
  3. Himes-Cornell, A.; Grose, S.O.; Pendleton, L. Mangrove Ecosystem Service Values and Methodological Approaches to Valuation: Where Do We Stand? Front. Mar. Sci. 2018, 5, 376. [Google Scholar] [CrossRef]
  4. Bimrah, K.; Dasgupta, R.; Hashimoto, S.; Saizen, I.; Dhyani, S. Ecosystem Services of Mangroves: A Systematic Review and Synthesis of Contemporary Scientific Literature. Sustainability 2022, 14, 12051. [Google Scholar] [CrossRef]
  5. Owuor, M.; Santos, T.M.T.; Otieno, P.; Mazzuco, A.C.A.; Iheaturu, C.; Bernardino, A.F. Flow of Mangrove Ecosystem Services to Coastal Communities in the Brazilian Amazon. Front. Environ. Sci. 2024, 12, 1329006. [Google Scholar] [CrossRef]
  6. Friess, D.A.; Rogers, K.; Lovelock, C.E.; Krauss, K.W.; Hamilton, S.E.; Lee, S.Y.; Lucas, R.; Primavera, J.; Rajkaran, A.; Shi, S. The State of the World’s Mangrove Forests: Past, Present, and Future. Annu. Rev. Environ. Resour. 2019, 44, 89–115. [Google Scholar] [CrossRef]
  7. Golebie, E.J.; Aczel, M.; Bukoski, J.J.; Chau, S.; Ramirez-Bullon, N.; Gong, M.; Teller, N. A Qualitative Systematic Review of Governance Principles for Mangrove Conservation. Conserv. Biol. 2022, 36, e13850. [Google Scholar] [CrossRef]
  8. Giri, C.; Ochieng, E.; Tieszen, L.L.; Zhu, Z.; Singh, A.; Loveland, T.; Masek, J.; Duke, N. Status and Distribution of Mangrove Forests of the World Using Earth Observation Satellite Data. Glob. Ecol. Biogeogr. 2011, 20, 154–159. [Google Scholar] [CrossRef]
  9. Thomas, N.; Bunting, P.; Lucas, R.; Hardy, A.; Rosenqvist, A.; Fatoyinbo, T. Mapping Mangrove Extent and Change: A Globally Applicable Approach. Remote Sens. 2018, 10, 1466. [Google Scholar] [CrossRef]
  10. Bunting, P.; Rosenqvist, A.; Lucas, R.M.; Rebelo, L.-M.; Hilarides, L.; Thomas, N.; Hardy, A.; Itoh, T.; Shimada, M.; Finlayson, C.M. The Global Mangrove Watch—A New 2010 Global Baseline of Mangrove Extent. Remote Sens. 2018, 10, 1669. [Google Scholar] [CrossRef]
  11. Bunting, P.; Rosenqvist, A.; Hilarides, L.; Lucas, R.M.; Thomas, N.; Tadono, T.; Worthington, T.A.; Spalding, M.; Murray, N.J.; Rebelo, L.-M. Global Mangrove Extent Change 1996–2020: Global Mangrove Watch Version 3.0. Remote Sens. 2022, 14, 3657. [Google Scholar] [CrossRef]
  12. Bunting, P.; Rosenqvist, A.; Hilarides, L.; Lucas, R.M.; Thomas, N. Global Mangrove Watch: Updated 2010 Mangrove Forest Extent (v2.5). Remote Sens. 2022, 14, 1034. [Google Scholar] [CrossRef]
  13. Wang, L.; Jia, M.; Yin, D.; Tian, J. A Review of Remote Sensing for Mangrove Forests: 1956–2018. Remote Sens. Environ. 2019, 231, 111223. [Google Scholar] [CrossRef]
  14. Yancho, J.M.M.; Jones, T.G.; Gandhi, S.R.; Ferster, C.; Lin, A.; Glass, L. The Google Earth Engine Mangrove Mapping Methodology (GEEMMM). Remote Sens. 2020, 12, 3758. [Google Scholar] [CrossRef]
  15. Acosta-Velázquez, J.; Ochoa-Gómez, J.; Vázquez-Lule, A.; Guevara, M. Changes in Mangrove Coverage Classification Criteria Could Impact the Conservation of Mangroves in Mexico. Land Use Policy 2023, 129, 106651. [Google Scholar] [CrossRef]
  16. Chen, C.-F.; Son, N.-T.; Chang, N.-B.; Chen, C.-R.; Chang, L.-Y.; Valdez, M.; Centeno, G.; Thompson, C.A.; Aceituno, J.L. Multi-Decadal Mangrove Forest Change Detection and Prediction in Honduras, Central America, with Landsat Imagery and a Markov Chain Model. Remote Sens. 2013, 5, 6408–6426. [Google Scholar] [CrossRef]
  17. Lymburner, L.; Bunting, P.; Lucas, R.; Scarth, P.; Alam, I.; Phillips, C.; Ticehurst, C.; Held, A. Mapping the Multi-Decadal Mangrove Dynamics of the Australian Coastline. Remote Sens. Environ. 2020, 238, 111185. [Google Scholar] [CrossRef]
  18. Maurya, K.; Mahajan, S.; Chaube, N. Remote Sensing Techniques: Mapping and Monitoring of Mangrove Ecosystem—A Review. Complex Intell. Syst. 2021, 7, 2797–2818. [Google Scholar] [CrossRef]
  19. Talukdar, S.; Singha, P.; Mahato, S.; Shahfahad; Pal, S.; Liou, Y.-A.; Rahman, A. Land-Use Land-Cover Classification by Machine Learning Classifiers for Satellite Observations—A Review. Remote Sens. 2020, 12, 1135. [Google Scholar] [CrossRef]
  20. Thakur, S.; Mondal, I.; Ghosh, P.B.; Das, P.; De, T.K. A Review of the Application of Multispectral Remote Sensing in the Study of Mangrove Ecosystems with Special Emphasis on Image Processing Techniques. Spat. Inf. Res. 2020, 28, 39–51. [Google Scholar] [CrossRef]
  21. Li, L. Geographically Weighted Machine Learning and Downscaling for High-Resolution Spatiotemporal Estimations of Wind Speed. Remote Sens. 2019, 11, 1378. [Google Scholar] [CrossRef]
  22. Liu, X.; Kounadi, O.; Zurita-Milla, R. Incorporating Spatial Autocorrelation in Machine Learning Models Using Spatial Lag and Eigenvector Spatial Filtering Features. ISPRS Int. J. Geo-Inf. 2022, 11, 242. [Google Scholar] [CrossRef]
  23. Mascaro, J.; Asner, G.P.; Knapp, D.E.; Kennedy-Bowdoin, T.; Martin, R.E.; Anderson, C.; Higgins, M.; Chadwick, K.D. A Tale of Two “Forests”: Random Forest Machine Learning Aids Tropical Forest Carbon Mapping. PLoS ONE 2014, 9, e85993. [Google Scholar] [CrossRef] [PubMed]
  24. McMahon, C.A.; Roberts, D.A.; Stella, J.C.; Trugman, A.T.; Singer, M.B.; Caylor, K.K. A River Runs through It: Robust Automated Mapping of Riparian Woodlands and Land Surface Phenology across Dryland Regions. Remote Sens. Environ. 2024, 305, 114056. [Google Scholar] [CrossRef]
  25. Atkinson, P.M. Spatially Weighted Supervised Classification for Remote Sensing. Int. J. Appl. Earth Obs. Geoinf. 2004, 5, 277–291. [Google Scholar] [CrossRef]
  26. Chen, L.; Wang, Y.; Ren, C.; Zhang, B.; Wang, Z. Assessment of Multi-Wavelength SAR and Multispectral Instrument Data for Forest Aboveground Biomass Mapping Using Random Forest Kriging. For. Ecol. Manag. 2019, 447, 12–25. [Google Scholar] [CrossRef]
  27. Dobbertin, M.; Biging, G.S. A Simulation Study of the Effect of Scene Autocorrelation, Training Sample Size and Sampling Method on Classification Accuracy. Can. J. Remote Sens. 1996, 22, 360–367. [Google Scholar] [CrossRef]
  28. Ayala-Izurieta, J.E.; Márquez, C.O.; García, V.J.; Recalde-Moreno, C.G.; Rodríguez-Llerena, M.V.; Damián-Carrión, D.A. Land Cover Classification in an Ecuadorian Mountain Geosystem Using a Random Forest Classifier, Spectral Vegetation Indices, and Ancillary Geographic Data. Geosciences 2017, 7, 34. [Google Scholar] [CrossRef]
  29. Kwong, I.H.Y.; Wong, F.K.K.; Fung, T.; Liu, E.K.Y.; Lee, R.H.; Ng, T.P.T. A Multi-Stage Approach Combining Very High-Resolution Satellite Image, GIS Database and Post-Classification Modification Rules for Habitat Mapping in Hong Kong. Remote Sens. 2022, 14, 67. [Google Scholar] [CrossRef]
  30. Maselli, F.; Rodolfi, A.; Bottai, L.; Romanelli, S.; Conese, C. Classification of Mediterranean Vegetation by TM and Ancillary Data for the Evaluation of Fire Risk. Int. J. Remote Sens. 2000, 21, 3303–3313. [Google Scholar] [CrossRef]
  31. Punalekar, S.M.; Hurford, C.; Lucas, R.M.; Planque, C.; Chognard, S. Hierarchical-Modular Framework for Habitat Mapping through Systematic and Informed Integration of Remote Sensing Data with Contextual Information. Ecol. Inform. 2024, 82, 102714. [Google Scholar] [CrossRef]
  32. Selvaraj, J.J.; Gallego Pérez, B.E. Estimating Mangrove Aboveground Biomass in the Colombian Pacific Coast: A Multisensor and Machine Learning Approach. Heliyon 2023, 9, e20745. [Google Scholar] [CrossRef]
  33. Griffith, D. What Is Spatial Autocorrelation? Reflections on the Past 25 Years of Spatial Statistics. L’Espace Géograph. 1992, 21, 265–280. [Google Scholar] [CrossRef]
  34. Griffith, D.A.; Chun, Y. Spatial Autocorrelation and Uncertainty Associated with Remotely-Sensed Data. Remote Sens. 2016, 8, 535. [Google Scholar] [CrossRef]
  35. Nikparvar, B.; Thill, J.-C. Machine Learning of Spatial Data. ISPRS Int. J. Geo-Inf. 2021, 10, 600. [Google Scholar] [CrossRef]
  36. Johnson, B.; Tateishi, R.; Xie, Z. Using Geographically Weighted Variables for Image Classification. Remote Sens. Lett. 2012, 3, 491–499. [Google Scholar] [CrossRef]
  37. Campbell, J.B. Spatial Correlation Effects upon Accuracy of Supervised Classification of Land Cover. Photogramm. Eng. Remote Sens. 1981, 47, 355–363. [Google Scholar]
  38. Alongi, D.M. Mangrove Forests: Resilience, Protection from Tsunamis, and Responses to Global Climate Change. Estuar. Coast. Shelf Sci. 2008, 76, 1–13. [Google Scholar] [CrossRef]
  39. Alongi, D.M. Impact of Global Change on Nutrient Dynamics in Mangrove Forests. Forests 2018, 9, 596. [Google Scholar] [CrossRef]
  40. MacKenzie, R.; Sharma, S.; Rovai, A.R. Chapter 12—Environmental Drivers of Blue Carbon Burial and Soil Carbon Stocks in Mangrove Forests. In Dynamic Sedimentary Environments of Mangrove Coasts; Sidik, F., Friess, D.A., Eds.; Elsevier: Amsterdam, The Netherlands, 2021; pp. 275–294. ISBN 978-0-12-816437-2. [Google Scholar]
  41. Kamal, M.; Phinn, S.; Johansen, K. Characterizing the Spatial Structure of Mangrove Features for Optimizing Image-Based Mangrove Mapping. Remote Sens. 2014, 6, 984–1006. [Google Scholar] [CrossRef]
  42. Luo, Z.; Sun, O.J.; Xu, H. A Comparison of Species Composition and Stand Structure between Planted and Natural Mangrove Forests in Shenzhen Bay, South China. J. Plant Ecol. 2010, 3, 165–174. [Google Scholar] [CrossRef]
  43. Lagomasino, D.; Fatoyinbo, T.; Castañeda-Moya, E.; Cook, B.D.; Montesano, P.M.; Neigh, C.S.R.; Corp, L.A.; Ott, L.E.; Chavez, S.; Morton, D.C. Storm Surge and Ponding Explain Mangrove Dieback in Southwest Florida Following Hurricane Irma. Nat. Commun. 2021, 12, 4003. [Google Scholar] [CrossRef]
  44. Simpson, L.T.; Canty, S.W.J.; Cissell, J.R.; Steinberg, M.K.; Cherry, J.A.; Feller, I.C. Bird Rookery Nutrient Over-Enrichment as a Potential Accelerant of Mangrove Cay Decline in Belize. Oecologia 2021, 197, 771–784. [Google Scholar] [CrossRef]
  45. Amaral, C.; Poulter, B.; Lagomasino, D.; Fatoyinbo, T.; Taillie, P.; Lizcano, G.; Canty, S.; Silveira, J.A.H.; Teutli-Hernández, C.; Cifuentes-Jara, M.; et al. Drivers of Mangrove Vulnerability and Resilience to Tropical Cyclones in the North Atlantic Basin. Sci. Total Environ. 2023, 898, 165413. [Google Scholar] [CrossRef] [PubMed]
  46. Cissell, J.R.; Canty, S.W.J.; Steinberg, M.K.; Simpson, L.T. Mapping National Mangrove Cover for Belize Using Google Earth Engine and Sentinel-2 Imagery. Appl. Sci. 2021, 11, 4258. [Google Scholar] [CrossRef]
  47. Meyer, H.; Reudenbach, C.; Wöllauer, S.; Nauss, T. Importance of Spatial Predictor Variable Selection in Machine Learning Applications—Moving from Data Reproduction to Spatial Prediction. Ecol. Model. 2019, 411, 108815. [Google Scholar] [CrossRef]
  48. Meyer, H.; Pebesma, E. Predicting into Unknown Space? Estimating the Area of Applicability of Spatial Prediction Models. Methods Ecol. Evol. 2021, 12, 1620–1633. [Google Scholar] [CrossRef]
  49. Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-Scale Geospatial Analysis for Everyone. Remote Sens. Environ. 2017, 202, 18–27. [Google Scholar] [CrossRef]
  50. Gress, E.; Voss, J.D.; Eckert, R.J.; Rowlands, G.; Andradi-Brown, D.A. The Mesoamerican Reef. In Mesophotic Coral Ecosystems; Loya, Y., Puglise, K.A., Bridge, T.C.L., Eds.; Springer International Publishing: Cham, Switzerland, 2019; pp. 71–84. ISBN 978-3-319-92735-0. [Google Scholar]
  51. Wilson, R.; Burke, L.; Lambert, L.J. A Situational Analysis of Mangroves in the Mesoamerican Reef System; Seatone Consulting: Larkspur, CA, USA; WRI: Washington, DC, USA, 2015; pp. 1–27. [Google Scholar]
  52. Arreola, A.; Eugenia, M. The Mesoamerican Barrier Reef System. In Marine Transboundary Conservation and Protected Areas; Routledge: Abingdon, UK, 2016; ISBN 978-1-315-72427-0. [Google Scholar]
  53. Canty, S.W.J.; Preziosi, R.F.; Rowntree, J.K. Dichotomy of Mangrove Management: A Review of Research and Policy in the Mesoamerican Reef Region. Ocean Coast. Manag. 2018, 157, 40–49. [Google Scholar] [CrossRef]
  54. Bing Maps. Microsoft Corporation. 2024. Available online: https://www.bing.com/maps (accessed on 31 January 2024).
  55. Rivas, A.B.; González, C.; Canty, S.; Rodríguez Olivet, C.; Flamenco, X.; González, M.J.; Escobedo, M. Regional Strategy for Mangrove Management, Conservation, Restoration and Monitoring in the Mesoamerican Reef 2020–2025; Mesoamerican Reef Fund: Guatemala City, Guatemala, 2020. [Google Scholar] [CrossRef]
  56. Wessel, P.; Smith, W.H.F. A Global, Self-Consistent, Hierarchical, High-Resolution Shoreline Database. J. Geophys. Res. Solid Earth 1996, 101, 8741–8743. [Google Scholar] [CrossRef]
  57. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  58. Rodriguez-Galiano, V.F.; Ghimire, B.; Rogan, J.; Chica-Olmo, M.; Rigol-Sanchez, J.P. An Assessment of the Effectiveness of a Random Forest Classifier for Land-Cover Classification. ISPRS J. Photogramm. Remote Sens. 2012, 67, 93–104. [Google Scholar] [CrossRef]
  59. Pelletier, C.; Valero, S.; Inglada, J.; Champion, N.; Dedieu, G. Assessing the Robustness of Random Forests to Map Land Cover with High Resolution Satellite Image Time Series over Large Areas. Remote Sens. Environ. 2016, 187, 156–168. [Google Scholar] [CrossRef]
  60. Morrissette, H.K.; Baez, S.K.; Beers, L.; Bood, N.; Martinez, N.D.; Novelo, K.; Andrews, G.; Balan, L.; Beers, C.S.; Betancourt, S.A.; et al. Belize Blue Carbon: Establishing a National Carbon Stock Estimate for Mangrove Ecosystems. Sci. Total Environ. 2023, 870, 161829. [Google Scholar] [CrossRef] [PubMed]
  61. European Space Agency. Copernicus Harmonized Sentinel-2 MSI: MultiSpectral Instrument, Level-2A. Google Earth Engine. 2015. Available online: https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2 (accessed on 1 October 2024).
  62. European Space Agency. Copernicus Sentinel-1 SAR GRD: C-Band Synthetic Aperture Radar Ground Range Detected, Log Scaling. 2014. Available online: https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S1_GRD (accessed on 1 October 2024).
  63. Hu, L.; Xu, N.; Liang, J.; Li, Z.; Chen, L.; Zhao, F. Advancing the Mapping of Mangrove Forests at National-Scale Using Sentinel-1 and Sentinel-2 Time-Series Data with Google Earth Engine: A Case Study in China. Remote Sens. 2020, 12, 3120. [Google Scholar] [CrossRef]
  64. Lucas, R.; Van De Kerchove, R.; Otero, V.; Lagomasino, D.; Fatoyinbo, L.; Omar, H.; Satyanarayana, B.; Dahdouh-Guebas, F. Structural Characterisation of Mangrove Forests Achieved through Combining Multiple Sources of Remote Sensing Data. Remote Sens. Environ. 2020, 237, 111543. [Google Scholar] [CrossRef]
  65. Farr, T.G.; Rosen, P.A.; Caro, E.; Crippen, R.; Duren, R.; Hensley, S.; Kobrick, M.; Paller, M.; Rodriguez, E.; Roth, L.; et al. The Shuttle Radar Topography Mission. Rev. Geophys. 2007, 45, RG2004. [Google Scholar] [CrossRef]
  66. NASA Shuttle Radar Topography Mission Global 1 arc Second. NASA EOSDIS Land Processes DAAC. Google Earth Engine. Available online: https://developers.google.com/earth-engine/datasets/catalog/USGS_SRTMGL1_003 (accessed on 1 October 2024).
  67. Grill, G.; Lehner, B.; Thieme, M.; Geenen, B.; Tickner, D.; Antonelli, F.; Babu, S.; Borrelli, P.; Cheng, L.; Crochetiere, H.; et al. Mapping the World’s Free-Flowing Rivers. Nature 2019, 569, 215–221. [Google Scholar] [CrossRef] [PubMed]
  68. Knapp, K.R.; Kruk, M.C.; Levinson, D.H.; Diamond, H.J.; Neumann, C.J. The International Best Track Archive for Climate Stewardship (IBTrACS): Unifying Tropical Cyclone Data. Bull. Am. Meteorol. Soc. 2010, 91, 363–376. [Google Scholar] [CrossRef]
  69. Johnson, B.; Tateishi, R.; Kobayashi, T. Remote Sensing of Fractional Green Vegetation Cover Using Spatially-Interpolated Endmembers. Remote Sens. 2012, 4, 2619–2634. [Google Scholar] [CrossRef]
  70. Behrens, T.; Schmidt, K.; Viscarra Rossel, R.A.; Gries, P.; Scholten, T.; MacMillan, R.A. Spatial Modelling with Euclidean Distance Fields and Machine Learning. Eur. J. Soil Sci. 2018, 69, 757–770. [Google Scholar] [CrossRef]
  71. Hidayah, Z.; Utama, R.Y.S.; As-Syakur, A.R.; Rachman, H.A.; Wiyanto, D.B. Mapping Mangrove above Ground Carbon Stock of Benoa Bay Bali Using Sentinel-1 Satellite Imagery. IOP Conf. Ser. Earth Environ. Sci. 2024, 1298, 012013. [Google Scholar] [CrossRef]
  72. Llano, X. SMByC-IDEAM. AcATaMa—QGIS Plugin for Accuracy Assessment of Thematic Maps, Version 24.6. 2024. Available online: https://github.com/SMByC/AcATaMa (accessed on 1 October 2024).
  73. QGIS. Association QGIS Geographic Information System. 2024. Available online: https://qgis.org/ (accessed on 1 October 2024).
  74. Congalton, R.G. A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data. Remote Sens. Environ. 1991, 37, 35–46. [Google Scholar] [CrossRef]
  75. Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; et al. Scikit-Learn: Machine Learning in Python. J. Mach. Learn. Res. 2011, 12, 2825–2830. [Google Scholar]
  76. BjØrnstad, O.N.; Falck, W. Nonparametric Spatial Covariance Functions: Estimation and Testing. Environ. Ecol. Stat. 2001, 8, 53–70. [Google Scholar] [CrossRef]
  77. Moran, P.A.P. The Interpretation of Statistical Maps. J. R. Stat. Society. Ser. B (Methodol.) 1948, 10, 243–251. [Google Scholar] [CrossRef]
  78. Rasquinha, D.N.; Mishra, D.R. Tropical Cyclones Shape Mangrove Productivity Gradients in the Indian Subcontinent. Sci. Rep. 2021, 11, 17355. [Google Scholar] [CrossRef]
  79. Hengl, T.; Nussbaum, M.; Wright, M.N.; Heuvelink, G.B.M.; Gräler, B. Random Forest as a Generic Framework for Predictive Modeling of Spatial and Spatio-Temporal Variables. PeerJ 2018, 6, e5518. [Google Scholar] [CrossRef]
  80. Cohen, M.C.L.; Yao, Q.; de Souza, A.V.; Liu, K.; Pessenda, L.C.R. Hurricanes Are Limiting the Mangrove Canopy Heights in the Gulf of Mexico. Sci. Total Environ. 2024, 927, 172284. [Google Scholar] [CrossRef]
  81. Wadoux, A.M.J.-C.; Heuvelink, G.B.M.; de Bruin, S.; Brus, D.J. Spatial Cross-Validation Is Not the Right Way to Evaluate Map Accuracy. Ecol. Model. 2021, 457, 109692. [Google Scholar] [CrossRef]
  82. Yıldırım, I.; Ersoy, O.K.; Yazgan, B. Improvement of Classification Accuracy in Remote Sensing Using Morphological Filter. Adv. Space Res. 2005, 36, 1003–1006. [Google Scholar] [CrossRef]
  83. Chen, D.; Stow, D. The Effect of Training Strategies on Supervised Classification at Different Spatial Resolutions. Photogramm. Eng. Remote Sens. 2002, 68, 1155–1162. [Google Scholar]
Remotesensing 17 02837 g001
Figure 1. Map of the study area illustrating the extent of the Mesoamerican Reef region across four countries, displayed over a Bing Maps satellite view [54].
Figure 1. Map of the study area illustrating the extent of the Mesoamerican Reef region across four countries, displayed over a Bing Maps satellite view [54].
Remotesensing 17 02837 g001
Remotesensing 17 02837 g002
Figure 2. Methodological flowchart set up for the model calibration and mapping using different feature sets. DEM—Digital Elevation Model, CoastD: Distance to coast; StreamD: Distance to river and streams; Storms: Storm frequency, IDW: Inverse Distance Weighted.
Figure 2. Methodological flowchart set up for the model calibration and mapping using different feature sets. DEM—Digital Elevation Model, CoastD: Distance to coast; StreamD: Distance to river and streams; Storms: Storm frequency, IDW: Inverse Distance Weighted.
Remotesensing 17 02837 g002
Remotesensing 17 02837 g003
Figure 3. Average, with standard deviation, of feature importance during the spatial cross-validation phase of the Random Forest model development. Description of features: Median composites of Sentinel-2 bands: Blue, Green, Red, NIR (Near-Infrared), SWIR-1 (Short-Wave Infrared band 1) and SWIR-2 (Short-Wave Infrared band 2); NDMI: Normalized-Difference Moisture Index, NDVI: Normalized-Difference Vegetation Index, CLrededge: Chlorophyll rededge, NDWI: Normalized-Difference Water Index, subscript p80 and p20 denote 80th and 20th percentile, respectively; VH and VV: Median composite of Sentinel-1 Vertical–Horizontal and Vertical–Vertical backscatter polarization bands; DEM: Digital Elevation Model; CoastD: Distance to coast; StreamD: Distance to river and streams; Storms: Storm frequency. ‘Dif’ in Spatial-3 denotes Difference bands. DEM onwards features on the x-axis are additional features integrated in different feature sets.
Figure 3. Average, with standard deviation, of feature importance during the spatial cross-validation phase of the Random Forest model development. Description of features: Median composites of Sentinel-2 bands: Blue, Green, Red, NIR (Near-Infrared), SWIR-1 (Short-Wave Infrared band 1) and SWIR-2 (Short-Wave Infrared band 2); NDMI: Normalized-Difference Moisture Index, NDVI: Normalized-Difference Vegetation Index, CLrededge: Chlorophyll rededge, NDWI: Normalized-Difference Water Index, subscript p80 and p20 denote 80th and 20th percentile, respectively; VH and VV: Median composite of Sentinel-1 Vertical–Horizontal and Vertical–Vertical backscatter polarization bands; DEM: Digital Elevation Model; CoastD: Distance to coast; StreamD: Distance to river and streams; Storms: Storm frequency. ‘Dif’ in Spatial-3 denotes Difference bands. DEM onwards features on the x-axis are additional features integrated in different feature sets.
Remotesensing 17 02837 g003
Remotesensing 17 02837 g004
Figure 4. Profiles of average omission errors in the mangrove classification with respect to increasing number of grid cells used in the model development. The errors were calculated for an independent set of validation points. The shaded portions represent standard deviations across 20 iterations for each number of grid cells.
Figure 4. Profiles of average omission errors in the mangrove classification with respect to increasing number of grid cells used in the model development. The errors were calculated for an independent set of validation points. The shaded portions represent standard deviations across 20 iterations for each number of grid cells.
Remotesensing 17 02837 g004
Remotesensing 17 02837 g005
Figure 5. Comparisons of Spatial-1, 2 and 3 output maps with Non-spatial output map to highlight areas (grid cells of 10 × 10 km grid) showing reduced number of mangroves pixels mapped.
Figure 5. Comparisons of Spatial-1, 2 and 3 output maps with Non-spatial output map to highlight areas (grid cells of 10 × 10 km grid) showing reduced number of mangroves pixels mapped.
Remotesensing 17 02837 g005
Remotesensing 17 02837 g006
Figure 6. Comparisons of Spatial-1, 2 and 3 output maps with Non-spatial output map to highlight additions in mangrove cover areas summarized over 10 km grid.
Figure 6. Comparisons of Spatial-1, 2 and 3 output maps with Non-spatial output map to highlight additions in mangrove cover areas summarized over 10 km grid.
Remotesensing 17 02837 g006
Remotesensing 17 02837 g007
Figure 7. Zoomed in sections of map showing differences in Spatial-1, 2 and 3 maps with respect to Non-spatial mangrove map. Maps corresponding to (a,d) Spatial-1, (b,e) Spatial-2, (c,f) Spatial-3 models are overlaid on the Non-spatial map to show removal of pixels. Whereas the Non-spatial map is overlaid on maps corresponding to (g) Spatial-1, (h) Spatial-2 and (i) Spatial-3 to show additions of pixels.
Figure 7. Zoomed in sections of map showing differences in Spatial-1, 2 and 3 maps with respect to Non-spatial mangrove map. Maps corresponding to (a,d) Spatial-1, (b,e) Spatial-2, (c,f) Spatial-3 models are overlaid on the Non-spatial map to show removal of pixels. Whereas the Non-spatial map is overlaid on maps corresponding to (g) Spatial-1, (h) Spatial-2 and (i) Spatial-3 to show additions of pixels.
Remotesensing 17 02837 g007
Remotesensing 17 02837 g008
Figure 8. Spline correlograms for mangrove training points for four Sentienl-2 bands, (a) Green, (b) Red, (c) NIR, and (d) SWIR-1. Medians values of bands were calculated over the period of January to June of 2019 and 2020 each.
Figure 8. Spline correlograms for mangrove training points for four Sentienl-2 bands, (a) Green, (b) Red, (c) NIR, and (d) SWIR-1. Medians values of bands were calculated over the period of January to June of 2019 and 2020 each.
Remotesensing 17 02837 g008
Remotesensing 17 02837 g009
Figure 9. Spline correlograms for validation points for four Sentienl-2 bands, (a) Non-spatial, (b) Spatial-1, (c) Spatial-2, and (d) Spatial-3. These were calculated using deviance residuals with true probability for true mangrove points set to 1.
Figure 9. Spline correlograms for validation points for four Sentienl-2 bands, (a) Non-spatial, (b) Spatial-1, (c) Spatial-2, and (d) Spatial-3. These were calculated using deviance residuals with true probability for true mangrove points set to 1.
Remotesensing 17 02837 g009
Table 2. Model and map accuracy assessment results for mangroves—including overall accuracy (OA), omission error (OM), and commission error (COM)—corresponding with four sets of features.
Table 2. Model and map accuracy assessment results for mangroves—including overall accuracy (OA), omission error (OM), and commission error (COM)—corresponding with four sets of features.
Feature SetModel Calibration Accuracies (%) from Spatial Cross-Validation StrategyMap Validation Accuracies (%)
OAOMCOMOAOMCOM
Non-spatial 94579773
Spatial-196459833
Spatial-295569753
Spatial-3981394172
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Punalekar, S.M.; Nowakowski, A.J.; Canty, S.W.J.; Fergus, C.; Huang, Q.; Songer, M.; Connette, G.M. Characterizing Role of Spatial Features in Improving Mangrove Classification—A Case Study over the Mesoamerican Reef Region. Remote Sens. 2025, 17, 2837. https://doi.org/10.3390/rs17162837

AMA Style

Punalekar SM, Nowakowski AJ, Canty SWJ, Fergus C, Huang Q, Songer M, Connette GM. Characterizing Role of Spatial Features in Improving Mangrove Classification—A Case Study over the Mesoamerican Reef Region. Remote Sensing. 2025; 17(16):2837. https://doi.org/10.3390/rs17162837

Chicago/Turabian Style

Punalekar, Suvarna M., A. Justin Nowakowski, Steven W. J. Canty, Craig Fergus, Qiongyu Huang, Melissa Songer, and Grant M. Connette. 2025. "Characterizing Role of Spatial Features in Improving Mangrove Classification—A Case Study over the Mesoamerican Reef Region" Remote Sensing 17, no. 16: 2837. https://doi.org/10.3390/rs17162837

APA Style

Punalekar, S. M., Nowakowski, A. J., Canty, S. W. J., Fergus, C., Huang, Q., Songer, M., & Connette, G. M. (2025). Characterizing Role of Spatial Features in Improving Mangrove Classification—A Case Study over the Mesoamerican Reef Region. Remote Sensing, 17(16), 2837. https://doi.org/10.3390/rs17162837

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop