Indonesia is the largest archipelagic country in the world, which consists of more than 17,380 islands1, and has the second-longest coastline in the world2, approximately 81,290 km (Kusumadewi 2025). This geographical characteristic offers a significant opportunity to develop coastal ecosystems, particularly for mangrove forests, as one of the key strategic elements in climate change mitigation (Basyuni et al. 2024; Dharmayasa et al. 2025; Hilmi et al. 2025).

Indonesia harbors the largest mangrove ecosystem in the world, covering approximately 3.36 million ha (about 25% of the world’s total mangrove area) (Bibi et al. 2019; Giri et al. 2011; Priyashanta and Taufikurrahman 2020). However, the Indonesian Peatland and Mangrove Restoration Agency (BRGM) reveals that about 637,000 ha of mangroves are in a critical state as a result of land conversion for aquaculture, oil palm plantations, urbanization, and agricultural extensification activities3. In addition, many coastal zones with high potential for mangrove establishment remain underutilized, while existing regulations largely emphasize rehabilitation of degraded mangroves rather than strategic optimization of new suitable areas4.

Mangrove ecosystems provide a wide range of ecosystem services, including shoreline protection from erosion, biodiversity conservation, and livelihood support for coastal communities, alongside their crucial role in climate regulation (Garmaeepour et al. 2025; Govindasamy 2011). Mangroves sequester three to five times more carbon than tropical terrestrial forests due largely to a significant accumulation of below-ground carbon (Ahmed et al. 2022; Ali et al. 2025; Basyuni et al. 2024; Kauffman and Donato 2012). This environmental service emphasizes the importance of mangroves in meeting global targets for climate mitigation, as well as supporting Indonesia’s Nationally Determined Contributions (NDC) (Martin et al. 2024; Prihatiningtyas et al. 2024).

Climate change remains a pressing global challenge that threatens human and ecological systems5. According to the IPCC (2023), the global temperature has already increased by 1.1°C above pre-industrial levels due to anthropogenic (human-caused) emissions of carbon dioxide (CO₂). If trends continue, global temperatures are expected to rise to the 1.5°C threshold within two decades, leading to intensified storms, unpredictable rainfall, flooding, and sea-level rise (Bolan et al. 2024; Zommers and Singh 2014). Despite their proven carbon sequestration capacity, mangroves in Indonesia remain underutilized as a large-scale climate mitigation strategy.

Previous studies have mapped and identified potential for mangrove restoration areas using varied datasets and methodologies. Mangrove extent mapping has advanced significantly through remote sensing technologies, including UAV-LiDAR approaches that enable precise individual tree detection in dense mangrove stands (Nandika et al. 2025), and satellite-based products such as the Global Mangrove Watch (GMW) (Bunting et al. 2022). Beyond documenting existing cover, habitat suitability modeling has emerged as a complementary approach for identifying restoration potential. Unlike extent mapping, which documents where mangroves currently exist, suitability modeling identifies where ecological and environmental conditions are suitable for mangrove establishment and growth, including areas that have been deforested or remain ecologically unrealized. A suitability model is therefore designed to extend beyond the current species distribution, capturing restoration potential rather than present-day cover. Studies by Syahid et al. (2020) and Sahraei et al. (2024) have employed this approach using tidal zonation and the Multi-Criteria Decision-Making (MCDM) method, respectively. More recently, machine learning approaches have emerged as powerful tools for mangrove suitability modeling, with studies by Hu et al. (2020); Jaffé et al. (2025); Lifeng et al. (2024) successfully applying these methods. Nevertheless, many of these efforts still rely heavily on site-specific field data for model training and validation, which limits scalability, incurs high costs, and constrains replicability at national or global levels.

This study addresses these limitations by mapping mangrove habitat suitability across the coast of Indonesia, employing GMW 2020 as a globally accessible, open-access dataset for training labels, where mangrove presence pixels are classified as ecologically suitable and absence pixels as non-suitable. The analysis integrates GMW with remote sensing–derived climate, biophysical, and socio-environmental predictors within a Random Forest machine learning framework. This approach enables accurate, scalable, and cost-efficient habitat suitability mapping by reducing dependence on extensive field surveys. In doing so, this study aims to contribute to more targeted spatial-based mangrove management and support climate change mitigation strategies such as the Nationally Determined Contribution (NDC) or Reducing Emissions from Deforestation and Forest Degradation (REDD+) programs, while also providing a policy basis for optimizing the utilization of Indonesia’s coastal areas.

Indonesia—an archipelagic nation situated off the coast of mainland Southeast Asia in the Indian and Pacific Oceans—is the largest archipelago in the world, encompassing a maritime area of approximately 3,257,357 km² (Jompa and Murdiyarso 2022). The country lies along the equator and experiences a tropical climate (BPS-Statistics Indonesia 2024). In this study, the analysis covers the entire Indonesian coastline, with the spatial extent delineated by a 5 km buffer seaward and landward from the shoreline, resulting in a 10 km-wide coastal zone of interest. This buffer-based approach was selected to capture the transitional zone between marine and terrestrial systems, which is ecologically significant for mangrove development and rehabilitation. The 5 km buffer distance was selected based on established mangrove habitat mapping that uses 5 km shoreline buffer intervals to ensure observable differences in coastal landscape characteristics (Yancho et al. 2020), and follows coastal management practice that conceptualizes the coastal buffer zone as the transition area between land and sea, encompassing supratidal, intertidal, and subtidal domains where waves, tides, sediment, and coastal ecosystems interact (Balke and Friess 2016; Lan and Hsu 2021; Ma et al. 2020). In addition, mangrove-based coastal protection designs also consider protective mangrove and foreshore width from hundreds of meters up to kilometers and explicitly recommend additional buffer space to accommodate future landward migration and accretion under sea-level rise (van Bijsterveldt et al. 2023; Xie and Schwarz 2022). The study area is shown in Fig. 1.

This study used a variety of spatial datasets obtained from publicly available sources to support the modeling of mangrove habitat suitability along the Indonesian coastline. The datasets were compiled with careful consideration of their spatial resolution and temporal relevance to ensure robust model performance. A 250 m spatial resolution was selected as the modeling grid to ensure compatibility with the soil texture datasets, which were subsequently overlaid with the suitability map to identify appropriate mangrove species for planting based on local soil characteristics—an approach aimed at supporting ecologically optimal and sustainable mangrove restoration as well as afforestation. All datasets were standardized to this 250 m grid within GEE using the sampleRegions() function. In addition, the categorical variable (Land Cover) was resampled using nearest neighbor interpolation to preserve class integrity, while the continuous data were resampled using bilinear interpolation—which estimates pixel values based on the weighted average of the four nearest pixels and is widely recommended for continuous spatial data to minimize abrupt value transitions (Arif and Akbar 2005; Q. Wang et al. 2022). Table 1 summarizes all datasets used in this study, including their resolution, data type, purpose, and respective references.

The input datasets used in this study can be categorized into three groups based on their role in the workflow. First, the GMW dataset served as the source of labeled samples for model training and testing. Second, the national coastline dataset was used solely for study area delineation—specifically to generate a 5 km seaward and landward coastal buffer as the spatial extent of the analysis. Third, eight raster datasets, including temperature, precipitation, elevation, slope, tidal, land cover, nighttime light, and population density, were used as predictor variables in the Random Forest classification model. In addition, soil texture datasets were not included as model predictors but were used in a post-classification overlay to provide essential information on the suitability of specific mangrove species to local edaphic conditions. The selection of mangrove species was guided by the literature review of Syahid et al. (2023), which emphasized the role of soil texture as a determinant in identifying potential areas for mangrove rehabilitation in Southeast Asia. By linking soil characteristics with species preference, species-specific restoration recommendations can be developed more effectively, reflecting the distinct ecological requirements of each mangrove species. Fig. 2. presents the spatial distribution of soil texture across coastal Indonesia, while Table 2 lists the potential mangrove species associated with each soil type.

This study applied an integrated approach of remote sensing and machine learning to assess mangrove habitat suitability across Indonesia’s coastal zone. Fig. 3. presents the workflow diagram outlining the analytical framework, which consisted of three main components.

The Random Forest algorithm was employed as the primary modeling method to classify mangrove habitat suitability using remote sensing–based predictors. Random Forest was selected due to its robustness in handling heterogeneous predictor variables, its ability to capture non-linear relationships (Breiman 2001), and its effectiveness in reducing overfitting (Shelestov et al. 2017). All analyses were implemented in the Google Earth Engine (GEE) platform, which enables large-scale integration of environmental, biophysical, and socio-economic datasets. Training and validation samples were derived from the GMW 2020 dataset, where pixels indicating mangrove presence were labeled as suitable (class 1), and absence as non-suitable (class 0). To address class imbalance and minimize potential classification bias, the number of suitable and non-suitable samples was balanced prior to model training. The balanced dataset was then randomly split into 70% for training and 30% for testing. A series of hyperparameter tuning experiments was conducted to identify the configuration with the best predictive performance. Variations in the number of trees, minimum leaf population, and bagging fraction were systematically tested. The final configuration—100 trees, a minimum leaf population of 5, a bag fraction of 0.8, no restriction on maximum nodes, and a fixed random seed of 42—was selected as it yielded the lowest classification error while maintaining stable and consistent results across trials. The resulting model produced a binary suitability map with two classes (0 = non-suitable; 1 = suitable) at a spatial resolution of 250 meters.

The performance of the Random Forest model was assessed using a combination of classification metrics derived from the confusion matrix and feature importance. The confusion matrix (Table 3) compares predicted and observed classes and is composed of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). From these components, several evaluation metrics were calculated as Eq. 1 through 5.

 (1)

 (2)

 (3)

 (4)

 (5)

where Pe denotes the expected accuracy, calculated from the marginal distributions of the confusion matrix.

These metrics provide a comprehensive evaluation of model performance. Overall accuracy reflects the proportion of correctly classified samples, precision assesses the reliability of positive predictions, recall measures the model’s sensitivity in detecting suitable areas, specificity evaluates its ability to correctly classify non-suitable areas, while the F1-score balances precision and recall under potentially imbalanced data conditions. The Cohen’s Kappa coefficient (k) is particularly important as it accounts for agreement occurring by chance, thus offering a more robust assessment compared to simple accuracy. To prevent bias and detect possible overfitting, performance metrics were compared between the training and testing datasets. Substantial discrepancies would indicate poor model generalization.

To visualize the contribution of each variable, we assessed feature importance, which was calculated from the relative contribution of each predictor to reducing impurity in the decision trees. In this study, the contribution of each variable to the mangrove suitability model was assessed using the Mean Decrease in Impurity (MDI, or Gini importance) method from the Random Forest algorithm. Variables with higher importance scores are interpreted as dominant factors influencing mangrove habitat suitability. This information not only enhances the interpretability of the model but also provides valuable ecological insights for targeted conservation planning.

The identification of suitable mangrove species was carried out by integrating the predicted suitability map with soil texture distribution to ensure ecologically sound recommendations. Species allocation was guided by the framework of Syahid et al. (2023), which establishes linkages between soil characteristics and species-specific ecological preferences across Southeast Asia. By explicitly associating habitat suitability with species preferences, the results provide actionable insights for large-scale restoration planning.

The performance of the mangrove suitability model was evaluated using multiple accuracy metrics derived from the confusion matrix. The dataset was randomly split into 70% for training (n = 1,203 samples) and 30% for testing (n = 503 samples). The final Random Forest model achieved an overall testing accuracy of 97.81%, indicating a very strong predictive capability. To assess the balance between precision and recall, the F1-score was calculated and yielded a value of 97.82%, confirming the model’s robustness in classifying both suitable and non-suitable areas. Cohen’s Kappa coefficient reached 0.96, which reflects an almost perfect agreement between predicted and actual labels, well above the commonly accepted threshold of 0.80 for high model reliability. The overall error rate was only 2.19%, suggesting minimal misclassification.

The confusion matrix (Table 4) demonstrates that the model correctly classified the majority of samples in both categories. For the suitable class, 97.24%, indicating that nearly all suitable habitats were correctly identified. For the non-suitable class, 98.41%, confirming the model’s ability to reliably exclude unsuitable areas. Only 4 false negatives and 7 false positives were recorded, showing that the model rarely misclassified unsuitable areas as suitable, and vice versa. These results highlight the effectiveness of the Random Forest classifier in integrating environmental and anthropogenic predictors for large-scale mangrove suitability mapping. The high agreement metrics and very low error rates suggest that the model provides a reliable spatial representation of potential mangrove habitats across Indonesia’s coastal zone.

The contribution of each predictor to the mangrove suitability model was assessed using the MDI or Gini importance method from the Random Forest algorithm. Fig. 4 shows the variable importance results of the Random Forest algorithm in this study. Analysis of variable importance shows that land cover is the most influential variable (31.81%), followed by population density (19.74%) and nighttime light (17.79%), reflecting human pressure on coastal ecosystems. The dominance of these variables in the model aligns with broader evidence that anthropogenic pressures are one of the primary drivers of mangrove habitat changes. A study by Thomas et al. (2017) indicates that land use conversion accounts for approximately 50% of mangrove loss in Southeast Asia, with Indonesia representing one of the most affected regions. At finer scales, a pattern consistent with findings from other tropical mangrove systems, such as the Sundarbans, demonstrates that human disturbance variables are also contributors in habitat suitability models and suggests that restoration areas should be located in regions with minor anthropogenic disturbance (Mondal et al. 2026; Sahana et al. 2022).

Among the environmental variables, temperature contributed most substantially (14.66%), reflecting the role of thermal gradients in defining the physiological boundaries of mangrove growth at the national scale, as temperatures that are too high or too low can lead to mangrove mortality (Osland et al. 2018). Elevation (5.86%), precipitation (4.71%), slope (2.8%), and tidal (2.63%) exhibit lower predictive contribution, despite their well-documented ecological significance for mangrove growth (Ellison 2021; Osland et al. 2017; Sagala et al. 2024). This does not imply that these variables are ecologically irrelevant; rather, it reflects a known limitation of large-scale machine learning models in which variables with limited spatial variation within the study area contribute less to model discrimination, even when they are ecologically critical.

Fig. 5. presents the spatial distribution of mangrove suitability across Indonesia’s coastal zone, derived from the Random Forest classification model using multiple environmental and socio-ecological predictors. The result of the mangrove suitability area has a spatial resolution of 250 m.

The model identified a total of 10,091,548.41 ha, or approximately 10.09 million ha, of land as suitable for mangrove growth. However, the spatial distribution of these areas is highly heterogeneous across the archipelago, with a spatial extent approximately 3 times larger than the existing mangrove cover based on GMW 2020. To assess the spatial correspondence between the predicted suitability and existing mangrove distribution, a spatial overlay analysis was conducted between our model output and GMW 2020, as shown in Fig. 6. Of the total GMW 2020 extent in Indonesia (2,935,637.99 ha), approximately 357,936.27 ha falls outside the 10 km coastal study area, which represents inland or offshore mangrove stands beyond the ecologically defined transitional coastal zone. Within the study area, the effective GMW 2020 extent is 2,577,701.72 ha, of which 1,741,370.62 ha (~67.5%) spatially overlaps with our predicted suitable areas, confirming that the model correctly identifies the majority of ecologically optimal mangrove zones. The remaining ~32.5% of GMW within the study area not captured by our model is attributed to two main factors: (1) resolution mismatch, GMW operates at 25 m SAR resolution while our model aggregates at 250 m, where narrow or fragmented mangrove belts are spatially averaged below the ecological signal threshold; and (2) ecological marginality, GMW records mangrove canopy presence regardless of health status, while our model predicts conditions for establishment and optimal growth, meaning that degraded or stress-affected stands may register in GMW but fall below our suitability threshold.

Several regions traditionally recognized as mangrove strongholds, such as Sembilang National Park (South Sumatra), Tanjung Puting National Park (Central Kalimantan), and Bintuni Bay (West Papua), were consistently classified as highly suitable, reflecting the strong ecological validity of the model outputs. Beyond existing conservation zones, the model also revealed substantial tracts of suitable land in regions with limited mangrove coverage or ongoing anthropogenic pressures, including Cilacap (Central Java), Nunukan (North Kalimantan), North Luwu (South Sulawesi), Kotabaru (South Kalimantan), and Merauke (South Papua). In addition to the national-level distribution, the spatial analysis was further disaggregated by province to identify priority regions for mangrove rehabilitation and expansion (Table 5). The results indicate that South Papua represents the largest share of suitable land, covering approximately 1.12 million ha, followed by Maluku (0.90 million ha) and Riau (0.84 million ha). Other provinces with substantial potential include West Papua (0.68 million ha), West Kalimantan (0.61 million ha), and Papua (0.58 million ha). Both eastern Indonesia (Papua and Maluku regions) and western Indonesia (Sumatra and Kalimantan) emerge as key hotspots, suggesting that opportunities for large-scale mangrove development are not geographically confined to a single island group. This uneven but extensive distribution underscores the importance of adopting region-specific management strategies that consider ecological conditions, socio-economic contexts, and governance capacity in each province.

In addition, translating national-scale suitability maps into on-the-ground restoration programs requires bridging the gap between broad spatial prediction and site-level ecological assessment. In this regard, emerging UAV-based monitoring tools, such as the LiDAR-derived individual tree detection approach demonstrated by Nandika et al. (2025) for dense Rhizophora stylosa stands in the Seribu Islands, illustrate the complementary role that high-resolution remote sensing can play in validating and refining suitability predictions at the stand level. Integrating national-scale models with such fine-grained verification approaches represents a promising direction for operationalizing future restoration planning that can be optimized by policymakers or restoration planners.

Soil texture is a critical physical property that strongly influences mangrove growth potential, as it determines water and oxygen retention capacity, soil fertility, and the stability of organic matter and decomposer activity (Dewiyanti et al. 2021). The overlay between the mangrove habitat suitability map and soil texture distribution revealed that most suitable coastal areas in Indonesia are dominated by soils with higher clay content, particularly Sandy Clay Loam (4,894,655.52 ha) and Clay Loam (4,134,043.31 ha). This finding is consistent with Dutta Roy et al. (2025), who reported that clay-rich soils, including silty clay, sandy clay, and clay loam, are the most favorable for mangrove growth due to their superior capacity to retain water and nutrients compared to sandy soils.

Fig. 7 illustrates the spatial distribution of mangrove-suitable areas overlaid with soil texture classes, showing the dominant soil conditions within predicted suitable mangrove habitats across Indonesia’s coastal zone. This result was generated by spatially overlaying the predicted mangrove habitat suitability map (as shown in Fig. 5), whereby only pixels classified as suitable were retained and overlaid with their corresponding soil texture class. The map serves as the basis for species-specific recommendations for mangrove planting. Based on these results, the most significant areas for mangrove ecosystem development are located on clay loam and sandy clay loam soils. According to Syahid et al. (2020), clay loam soils are particularly suitable for species such as Rhizophora stylosa, Rhizophora mucronata, Rhizophora apiculata, Xylocarpus granatum, Phoenix paludosa, Excoecaria agallocha, Bruguiera gymnorhiza, Ceriops decandra, and Avicennia alba. In contrast, sandy clay loam soils are more favorable for species including Bruguiera gymnorhiza, Kandelia candel, Avicennia officinalis, Sonneratia apetala, Xylocarpus mekongensis, Aegiceras corniculatum, and Heritiera fomes. These insights are essential for developing species- and soil-based strategies in mangrove restoration and cultivation programs. Ultimately, these findings are expected to support more effective and optimized approaches for the sustainable development of coastal ecosystems.

This study successfully mapped the suitability of mangrove habitats along the Indonesian coast by integrating remote sensing, environmental, biophysical, and socio-economic data with a machine learning approach. However, several limitations should be acknowledged.

(1) The habitat suitability model relied exclusively on globally available predictors for the year 2020, which assumes that the biophysical and socio-environmental conditions of coastal areas remain constant during that period. Such temporal assumptions may overlook interannual variability, extreme climatic events, or recent land-use changes that could significantly alter local habitat suitability.

(2) The model depended on the GMW dataset as the source of training labels for mangrove presence and absence classification. Although GMW provides consistent global coverage, its reliability may be constrained in regions with complex land-cover mosaics or mixed vegetation types, potentially leading to under- or over-estimation of mangrove extent.

(3) A further limitation concerns the 250 m spatial resolution adopted in this study, which was set to match the soil texture dataset in the analytical framework. At this resolution, each pixel represents 6.25 ha of coastal landscape, meaning that narrow mangrove belts, fragmented tidal fringe patches, and linearly distributed stands—which are ecologically significant but spatially confined—may be spatially averaged below the detectable signal threshold. This resolution constraint likely contributes to systematic underestimation of suitable areas along coastlines with complex morphology, and the predicted suitable areas should therefore be interpreted as reflecting ecologically realizable potential at the landscape scale, and complementary fine-resolution analyses are encouraged for site-level verification in morphologically complex coastal settings.

Addressing these limitations, future research should prioritize the integration of additional biophysical parameters such as soil salinity, pH, water quality indices, and sediment dynamics, which are known to strongly influence mangrove growth and survival. Incorporating such variables would allow for a more nuanced and ecologically comprehensive representation of mangrove habitat suitability. At broader scales, this approach can be extended to regional or global levels by leveraging open-source datasets, enabling cross-country comparisons across tropical regions and strengthening the global relevance of the findings. Furthermore, scenario-based modeling (e.g., under projected sea-level rise, climate change, or increasing anthropogenic pressures) is essential to assess the future dynamics of mangrove suitability. Such forward-looking assessments would provide more robust evidence for adaptive management, restoration planning, and climate change mitigation strategies, such as Indonesia’s NDC and REDD+ commitments.

This study successfully mapped mangrove habitat suitability along the Indonesian coastline by integrating remote sensing data with environmental, biophysical, and socio-economic predictors using a Random Forest machine learning approach. The developed model demonstrated strong performance, achieving an overall accuracy of 97.81% and a Kappa coefficient of 0.96, and estimated a total suitable area of ~10.09 million ha. Land cover, population density, and nighttime light emerged as the most influential variables, underscoring the importance of anthropogenic factors alongside biophysical drivers. The spatial distribution analysis revealed that the largest suitable areas are concentrated in South Papua, Maluku, and Riau Provinces, with dominant soil types of sandy clay loam and clay loam supporting the growth of major mangrove species. These findings not only confirm the ecological validity of the model but also highlight the significant potential for mangrove restoration and rehabilitation in Indonesia. Additionally, the results provide a scientific basis for more targeted spatial planning in conservation, restoration, and sustainable coastal ecosystem management, while contributing to Indonesia’s NDCs and the REDD+ programs. Looking ahead, the approach can be scaled up to regional and global levels using harmonized open-source datasets, enabling its application across other tropical countries and supporting global climate change mitigation strategies.

1. BIG. (2024). Pulau Indonesia Bertambah Jadi 17.380, Mengapa Angkanya Berubah Setiap Tahun? Sipulau BIG. https://sipulau.big.go.id/news/112. Ministry of ESDM. (2009). Kapal Survei Geomarin III Sebagai Sebuah Jawaban |The Geomarin III Survey Ship as an Answer. https://www.esdm.go.id/id/media-center/arsip-berita/kapal-survei-geomarin-iii-sebagai-sebuah-jawaban3. Siregar, E. (2021). BRGM: Kerusakan ekosistem mangrove kategori kritis capai 637.000 ha. ANTARA 2025. https://jambi.antaranews.com/berita/455534/brgm-kerusakan-ekosistem-mangrove-kategori-kritis-capai-637000-ha4. Regulation of the Minister of Environment and Forestry of the Republic of Indonesia Number: P.105/MENLHK/SETJEN/KUM.1/12/2018 Concerning Procedures for Implementation, Supporting Activities, Provision of Incentives, as Well as Guidance and Control of Forest, Jakarta: Kementerian Lingkungan Hidup dan Kehutanan 1 (2018).5. WHO. (2023). Climate change. https://www.who.int/news-room/fact-sheets/detail/climate-change-and-health