2.1. Study area

The study area spans 35.28°N to 36.00°N and 75.71°W to 77.12°W along the eastern US Coast, covering 644,892 ha of the Atlantic Coastal Plain in North Carolina, United States (Fig 1). This area includes a series of low-relief drowned river valleys with a chain of barrier islands surrounding the country’s second-largest estuary, the Albemarle-Pamlico. It is right below the US mid-Atlantic coast, an area known as a global sea-level rise hotspot [6], spanning strong gradients in salinity, topography, and accelerating relative sea level rise rate (RSLR). Specifically, this area is experiencing a rapid rate of relative sea-level rise, estimated at 3.5 mm yr^-1 in 2010 and 5.14 mm yr^-1 in 2021 (S6 Fig). Across the study area, hurricanes and SLR are driving saltwater intrusion and increased inundation, resulting in forest dieback (i.e., ghost forest, see Fig 1(a)) and ultimately leading to marsh encroachments [7,9,46,47]. The study area encompasses both private and protected areas [48], including the Alligator River National Wildlife Refuge, Pocosin Lakes National Wildlife Refuge, Mattamuskeet National Wildlife Refuge, Buckridge Coastal Reserve, Palmetto Peartree Preserve, and Swanquarter National Wildlife Refuge. Freshwater, shrub-dominated pocosins are unique, acidic, peat-rich wetlands found in the coastal plain, known for their dense vegetation and critical role in water storage [49]. Additionally, the eastern part of the study area is mostly covered by woody or forested wetlands; as you move inland, it increasingly consists of upland forests. Farmland also accounts for 27% of the study area.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. The study area map (bounded in the north by the Albemarle Sound and in the south by the Pamlico Sound) shows the digital elevation model of the North Carolina coastal plain and the protected area (brown strip diagonal lines).

The image inserts show ghost forests, which increase light availability for shrub growth and marshlands.

https://doi.org/10.1371/journal.pone.0357346.g001

2.2. Image acquisition and preprocessing

We collected and preprocessed satellite imagery from multiple sensors to classify land cover. We selected all images for the growing and leaf-off seasons from Landsat (30-meter resolution) and Sentinel-2 (10-meter resolution) using a cloud cover filter threshold of less than 2% (see Table 1). The filtered images spanning multiple years (1985, 2010, 2021) and overlapping with the study area were downloaded from the Google Earth Engine catalog. We didn’t account for intermediate time steps before 2010, given the relatively slow processes of coastal forest change, especially since a major portion of forest change in this study area became more visible with the coarse resolution of Landsat images post-2010 [4,9]. Images collected included the filtered Landsat 8 and 5 Level 2, Collection 2 Surface Reflectance for six spectral bands (Blue, Green, Red, Near Infrared, Short-wave infrared (SWIR1, SWIR2)) bands, as well as the Harmonized Sentinel-2 Surface Reflectance images (Blue (B2), Green (B3), Red (B4), Near Infrared (B8), SWIR 1 (B11), SWIR(B12). We utilized the ancillary quality assessment datasets to process all Landsat images by masking out pixels related to cloud, cloud shadow, missing data, snow, and ice specific to each satellite sensor. Specifically, residual clouds, shadows, and haze were filtered using the SR_ATMOS_OPACITY (the aerosol optical thickness value in Landsat surface reflectance data, indicating the clarity or haziness of the atmosphere when the image was captured) for Landsat 5 by thresholding to exclude those with values less than 0.3. Conversely, the ‘SR_AEROSOL’ band was employed to eliminate pixels labeled as high probability of haze in Landsat 8 [50]. Images were also visually verified to ensure they were not inundated and of high quality. To be consistent, two images (representing bi-season) were acquired to generate the annual land cover classification map: one in May-September (growing season) and the other in December–February (leaf off) – the period when inter-vegetation contrast is at its highest [9,23,32,51]. The 2019 and 2011 images were used to supplement the validation and testing data, given our small study area.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Data acquisition description for the winter and summer seasons.

https://doi.org/10.1371/journal.pone.0357346.t001

Multispectral indices, including the normalized difference vegetation index (NDVI), Enhanced Vegetation Index (EVI), modified normalized difference water index (mNDWI), tasseled cap transformation, and Modified Soil-Adjusted Vegetation Index (mSAVI), were calculated and stacked over the spectral bands (see Table 2).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Summary of metrics from the Landsat 5, 8, and Sentinel 2 images.

https://doi.org/10.1371/journal.pone.0357346.t002

To collect training samples for the stacked bands, Esri and Google Earth imagery, collected on March 8, 2021, and April 30, 2010, were used as a reference for visual inspection. Additionally, the NC Orthophoto imagery (6 inches) for the years 2020 and 2010 [52] and the 2019 NAIP Imagery were also used for visual reference [53]. The classes included forests, ghost forests, shrubs, marshes, water, and cultivated land. Samples were taken for each class for the years 2021, 2011, and 2010 (see Fig 2 for a description of the classes). We obtained additional training data for the shrub and ghost forest classes from previous research conducted in the Albemarle Peninsula [7,9]. Specifically, as shown in Fig 2 (used for training data), each point characterized as a ghost forest contained approximately 20–40 visible snags or fallen tree trunks per Landsat pixel, with some bare-soil reflectance [9]. The forest here refers to upland forests and forested wetlands (i.e., freshwater swamps) with closed-canopy covers.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Examples of land cover classes mapped and visualized in a false color image (NIR, Red, and Green).

Source: NC Orthophoto Imagery, NC OneMap (2020).

https://doi.org/10.1371/journal.pone.0357346.g002

2.3. Vegetation phenology and phenological metrics

As many specific phenological events cannot be directly observed at the spatial resolution of satellite imagery, vegetation dynamics are characterized using a descriptor called “land surface phenology (LSP)” [54], which is estimated based on time series vegetation indices such as NDVI derived from satellite data [55]. We extracted phenological metrics from the time-series NDVI computed from satellite imagery, after smoothing using the Savitsky-Golay filter [56] to reduce uncertainties in the phenological signals. The Savitsky-Golay function has been widely used to reduce noise in the vegetation index [57,58]. The model used for this is Phenolopy [59], a pixel-based phenological extraction Python library. This study adopted the Chen and Kirwan (2022) [32] method, which involves stacking a 3-year NDVI image to account for prominent cloud cover. Firstly, we download Landsat surface reflectance images from Google Earth Engine by filtering to select only images with less than 60% cloud cover and 35% cloud cover for the Sentinel images. These images include both winter and summer images, which are used for classification. Then, we computed a 3-year NDVI (e.g., the previous year – 2020, the year of interest – 2021, and the following year – 2022) for all the masked images (i.e., excluding water, haze, clouds, and cloud shadow). We further stacked all NDVI data in Day of Year (DOY) to produce enough monthly data points for modeling yearly phenological metrics [32]. On average across all 3-year NDVI stacks, we have 36 ± 5 images per 3-year for Landsat and 51 ± 9 images per 3-year for Sentinel-2.

Lastly, we extracted phenological metrics from smooth NDVI data. These metrics include the start of the growing season (SOGS) – the day of the year when the rising edge of the fitted curve reaches 20% of its seasonal amplitude, length of the growing season (LOGS) – the period between the start and the end of the growing season; base NDVI (B_NDVI) value – mean value of the lowest NDVI values to the left and right of the Peak of the Season, the peak of the growing season (POGS) – the day of the year when NDVI of the fitted curve reaches its maximum greenness; peak NDVI (P_NDVI) value – maximum value in the NDVI timeseries; and amplitude of NDVI (A_NDVI) – the difference between the maximum values and the base NDVI values per pixel (see S1 Fig for the graphical description of each phenological metrics) [32,55,60].

We further compared the phenological metrics across different land cover classes using a one-way ANOVA and the Kruskal-Wallis test [61]. We stack the six extracted phenology metrics on top of the initial spectral and multispectral indices for their corresponding satellite imagery (see Table 2 for a summary of the metrics).

2.4. Training, validation, and test datasets

Training data for land cover analysis involves selecting a sample of pixels from the image and using it to establish thresholds for delineating specific land covers on the ground [67]. We extracted and stacked spectral, multispectral, and biophysical indices computed from the satellite imagery and the DEM, and then normalized them using Z-score normalization techniques [27]. A supervised deep learning model typically requires a substantial amount of training data samples. However, it is often very time-consuming and costly to label the observed data and create training samples for each LULC class [38]. Therefore, a series of training, test, and validation polygons were drawn across each class using Google Earth Imagery as the base map for the years 2010 and 2021, which were converted to raster based on the classes. The datasets for training, testing, and validation were generated by randomly selecting pixel patches from each corresponding train, test, and validation raster, with patches spaced at least 90 meters (i.e., 3 pixels) apart, ensuring class stratification across all satellite sensor types. This homogeneous patch sampling ensures that pixels are only derived from a region of consistent land cover. Specifically, for the co-available year (2021), the shape of each of the patches sampled and fitted into the model were as follows: (a) Landsat-8 full feature model is 7 (width) x 7 (height) x 26 (Bands), and 20 bands for the non-phenology models; (b) while Landsat-5 was trained with just the full feature model; and (c) Sentinel-2 full feature model is 21 (width) x 21 (height) x 26 (Bands), and 20 bands for the non-phenology models). This totaled up to 600 image patches per class, except for the Ghost Forests (310), for each satellite. The training, validation, and test data were split by a ratio of 7:1.5:1.5, stratified by class per image sensor, with the training set for the Landsat-5 model from 2010 imagery, while 2021 imagery for both Landsat-8 and Sentinel-2 models, except for the ghost forest classes, supplemented by 2019 samples, while 80% of the test and validation sets were from the 2019 imagery.

2.5. Classification method

CNN is a hierarchical feature detector, inspired by biological principles, with a classification capability based on contextual information [68]. It is capable of learning highly abstract features and efficiently identifying objects through weight sharing and feature extraction [38,69–71]. Due to this characteristic, CNN is well-suited for processing multiband remote-sensing image data, where pixels are arranged in a regular pattern [38]. Fully trained full-feature CNN models were validated using the validation datasets and tested using the testing sets obtained from previous steps to classify forest classes and identify areas of decline.

The architecture includes four convolutional layers, each applying 32–64 filters [27,72] with a kernel size of 3 × 3 (see Fig 3). An L2 regularization of 0.01 was used to prevent overfitting, given the large feature space relative to spatial context. The training was terminated at the epoch (each being a complete pass through the training data) corresponding to the lowest validation loss for each of the sensors (see S2 Fig), with a batch size of 32 batches [72] and a learning rate of 0.0001 based on the cross-validation performance. The filter, also known as the weight vectors, slides over the input vector (i.e., stacked image patches) to generate the feature map [73]. This method, which involves sliding the filter horizontally and vertically, is known as a convolution operation [74]. This operation extracts X features from the input image within a single layer, producing X filters and X associated feature maps [74]. The output a_ij in the next layer for location (i,j) is computed after applying the convolution operation using the formula given by [71] as shown below:

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. The architecture of the Convolutional Neural Network (CNN) used for image classification consists of multiple convolutional layers followed by a fully connected output layer for final prediction.

https://doi.org/10.1371/journal.pone.0357346.g003

(1)

where X is the input given to the layer, W is the filter or kernel that moves over the input, b is the bias, * signifies the convolution operation, and σ indicates the non-linearity added in the network. The layer that carries out the classification was converted to a convolutional layer. The convolution kernel is chosen so that its dimensions match those of the input layer [71].

To evaluate the classification results across the various satellite imagery (given the difference in temporal, spectral, and spatial resolution), we computed the confusion matrix table, which was used to calculate the Accuracy (Eqn. 2), User Accuracy (Eqn. 3), Producer Accuracy (Eqn. 4), and F1 score (Eqn. 5) [27,32], as follows:

(2)

(3)

(4)

(5)

A confusion matrix compares predicted and observed class labels, and it summarizes model outcomes into four categories: true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). The user’s accuracy (Precision) is the proportion of correct predictions for each predicted class in the test data. In contrast, the producer’s accuracy (Recall) represents the proportion of correct predictions for each actual class [75], while the F1 score is the harmonic mean of both.

An image patch to be classified is fed into the input layer, and the output is the predicted land cover class label for each pixel, computed using the extracted features from the image. Furthermore, we utilized the permutation feature importance (PFI) to evaluate the significance of each feature in the model. PFI measures the degree to which a trained model relies on each feature [76]. After training, the values of a single feature are randomly shuffled, and a drop in model performance indicates the importance of that feature. If the error (Mean Absolute Error (MAE), i.e., the average magnitude of errors between predicted values and actual (observed) values) increases significantly when a feature is shuffled, it indicates that the feature is important. Since shuffling the features leads to an increase in the error metric, it means that these features are contributing positively to the model’s performance [77]. The size of the values (how far above zero they are) indicates the level of importance. Larger MAE values indicate that the feature plays a more significant role in the model’s predictions.

(6)

Where I is used to index the features, j represents the total number of features, 1_j is the importance of the j-th feature, S is the baseline performance of the model (i.e., accuracy), is the score after permuting the j-th feature in the k-th repetition, and K is the number of permutations, which in this case is 20. We calculate the average MAE over the total 20 runs and display the variation as a standard deviation.

We also conducted post-processing of our final classified map to remove obvious classification errors. Additionally, to facilitate cross-product comparisons of coastal forests, we compared our results with the National Land Cover Database (NLCD) for 1985 and 2021. NLCD uses a complex system comprising approximately 20 land-cover classes [78]. For comparison, we consolidated NLCD pixels labeled evergreen forest, deciduous forest, shrub/scrub, mixed forest, and woody wetland as forests. We considered the NLCD emergent herbaceous wetlands to be marshes [32]. In our study, we also merged the forest, ghost forest, and shrub classes into a single class to represent forests for comparison.

2.6. Drivers of forest retreats

We examined coastal forest loss between 1985 and 2021 by identifying pixels classified as forest in 1985 that have since transitioned to other land cover types, such as shrub, wetland, or ghost forest. This is done only for the Landsat Imagery because it’s long-term, while Sentinel-2 was used only for the 2021 image comparison with Landsat 8. Using the slope and elevation computed from the CONED DEM [66], we assessed the topographical impact on forest loss and the formation of ghost forests. We also tested the relationship between distance to channel, salinity, climatic variables, and forest loss. We examined three major pathways of forest change (i.e., forest-to-ghost forest, forest-to-shrub, and forest-to-marsh). For each pathway, we used logistic regression to assess the factors driving the dynamic patterns of coastal forest change (see Table 3). A total of 1,000 points were randomly generated across all pixels/areas, stratified by change types (i.e., binary change), using a 500-meter spatial exclusion buffer to reduce autocorrelation as employed by Ury et al. (2021) [9]. Other environmental predictors, such as salinity, distance to channel, and climate variables, were also employed. A k-fold cross-validation approach was employed to fit a binomial logistic regression model in which binary coastal forest change is modeled as a function of the predictors, including distance to water, salinity, relative sea level rise, and elevation (Table 3).

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Potential predictors of forest change.

https://doi.org/10.1371/journal.pone.0357346.t003

3.1. Comparison of phenological metrics between landcover classes based on the preprocessing steps

Coastal land cover types exhibit contrasting patterns of land surface phenology (Fig 4). For example, as obtained from the training data, the Peak NDVI value for both Landsat and Sentinel images reaches 0.88 ± 0.02 in the forest, which is higher than that of marsh (0.57 ± 0.08), shrub (0.72 ± 0.04), and ghost forest (0.83 ± 0.02) vegetation. Another thing that stands out in Fig 4 is the annual variability of the forest class (D) for the Landsat imagery, which differs from that of the other classes and may be related to differences in reflectance values between Landsat 5 and Landsat 8.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Annual vegetation patterns estimated from training data for marsh (A), shrub (B), ghost forest (C), and forest (D) were estimated using Landsat NDVI time series, while E-H represent the Sentinel 2 NDVI trend.

Plots show the mean DOY NDVI ± SD data at three time steps.

https://doi.org/10.1371/journal.pone.0357346.g004

Additionally, the Landsat Forest (103 ± 27 DOY) has the earliest start of growing season (SOGS) value, and the shrub (123 ± 23 DOY) has the latest SOGS value, while the other classes are in between. We also observed a distinct difference between the satellites’ SOGS values for ghost forests, with the start of the growing season for the Sentinel-2 images occurring 17 days earlier than the Landsat images, and the forest starting 2 days earlier. Differences are evaluated using a one-way ANOVA, and the results indicate that all phenology metrics differ among the four vegetation classes for the Sentinel-2 and Landsat 8 images (see Table 4). The ANOVA results show that the interclass difference at the start of the growing season is significant (p ≤ 0.001), as well as for all other metrics. However, all other metrics have a low R² (R² ≤ 0.11, i.e., while the difference between classes is statistically significant, only 10−21% of the variability in SOGS, POGS, A_NDVI can be explained by class) except for the Base NDVI (R² = 0.79) and Peak NDVI (R² = 0.86) for Landsat 8 images indicating that a large proportion of the variation these metrics can be explained by the class difference. A similar pattern was observed in the extracted metrics from the Sentinel-2 images, but a lower R-squared value (R² ≤ 0.019) in SOGS, POGS, and LOGS with A_NDVI (R² = 0.28), Base NDVI (R² = 0.59), and Peak NDVI (R² = 0.83). There is also a variation in the length of the growing season for all land cover types except ghost forests (approximately 225 days) between the two satellites (see Table 4). However, the base NDVI, Amplitude of NDVI, and Peak NDVI metrics for both satellite sensors are similar.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Phenological results for four classes estimated from the training samples.

https://doi.org/10.1371/journal.pone.0357346.t004

3.2. Classification results and evaluation

We assessed the model’s accuracy using the user’s, producer’s, and F1 score metrics, as well as the loss curve. Firstly, as shown in S2 Fig, the training and validation accuracies increased rapidly during the initial epochs and plateaued at around 0.94, indicating that the model effectively learned the underlying patterns for both Sentinel 2 and Landsat. Correspondingly, the training and validation loss decreased steadily, stabilizing at approximately 0.4 by the final epoch. The close alignment between the training and validation curves for both Sentinel-2 and Landsat-8, with minimal divergence, suggests that the model achieved high predictive performance and good generalization on unseen datasets without significant overfitting.

For the Landsat dataset, the classification results for all land cover types demonstrated a high classification accuracy (weighted average accuracy = 94% & F1 Score = 93%), indicating that most instances of each land cover class were correctly classified into their true class with few misclassifications. For the Ghost Forest class, the results show the presence of false negatives, instances where areas that should have been classified as Ghost Forest were instead misclassified as forest (11%) and Shrub (5.5%), suggesting underestimation and potential overestimation by the forest samples, with about 5.1% false positives in the model estimation. Additionally, 8% of the marsh class was misclassified as water.

Similar to Landsat, the producer’s accuracy for Sentinel imagery ranged from 90% for ghost forest to 99% for water, emphasizing the model’s overall consistency. Sentinel achieved an overall accuracy of 97%, indicating a high level of agreement (see Table 5). However, 11% of ghost forest pixels were misclassified as shrub, while 7% of water pixels were misclassified as marsh. There is an 8% false positive for the ghost forest class, indicating an overestimation of ghost forest areas, i.e., non-ghost forest areas (forests, shrubs, and cultivated) were classified as such

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. The confusion matrix showing the classification accuracy of land-cover maps produced by the Landsat 8 and Sentinel 2 satellites in 2021 using all 26 features.

https://doi.org/10.1371/journal.pone.0357346.t005

The Landsat 5 Classifier is trained using 2010 and 2011 images, achieving an overall accuracy of 94% with an F1 score of 93.8%. Landsat 5 has the lowest accuracy for users and producers in the Ghost Forest class compared to Landsat 8 and Sentinel-2, which could be a result of a very small training dataset for the Ghost Forest class.

Here, the result indicates that the phenological metrics have the highest importance on the model prediction, especially the Base NDVI (B_NDVI), amplitude of NDVI (A_NDVI), and peak growing season (POGS) for the Landsat 8 image, followed by the winter spectral indices, with the NIR band leading to an approximately 0.10 mean absolute error in classification accuracy when shuffled (Fig 5). On a per-class level, the results show that multispectral indices (mNDWI) aid in distinguishing water from every other class, while a combination of Amplitude of NDVI, mNDWI, winter red, and blue helps in separating marsh and every other class, as indicated by the model fit error. For the Ghost Forest class, winter blue and peak NDVI help in separating from other vegetation classes (see S3 Fig).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. (a) User’s Accuracy and (b) Producer’s accuracy of classification results from Landsat 5, Landsat 8, and Sentinel 2; (c) Landsat 8’s and (d) Sentinel-2’s permutation feature importance value – the red and blue colors represent the spectral indices for winter and summer, respectively while green represents multispectral indices gray for topographical and orange represents the phenological indices.

https://doi.org/10.1371/journal.pone.0357346.g005

In contrast, in the Sentinel-2 full feature prediction, multispectral indices have the most significant impact on the prediction, specifically the modified Normalized Difference Water Index (mNDWI), resulting in a 0.17 reduction in accuracy when shuffled. The Peak growing season (POGS), Amplitude, and Base NDVI metrics also contribute significantly to the model prediction, with an average of 0.085 increase in error when removed from the model. On a per-class level, mNDWI distinctly aids in the separation of water and marsh from every other class, while wintergreen, short-wave infrared, and peak NDVI aid in separating ghost forests and other vegetation classes. TCP greenness, the start of the growing season, and Base NDVI have the highest separation power for the forest class (see S4 Fig). Lastly, for both satellites, phenology is the major feature that aids in the separability of shrub from other classes (see S3–S4 Figs). The non-phenology model (i.e., excluding phenology) shows that without phenology, Sentinel-2 held only a marginal advantage over Landsat 8 (test accuracy 0.967 vs. 0.948; gap = 0.019), while Sentinel-2’s advantage increases over Landsat 2 (0.972 vs. 0.941; gap = 0.031). Sentinel-2’s accuracy improved with the addition of phenology, while Landsat 8’s accuracy declined (McNemar’s test, p < 0.001 for both sensors).

3.3. Forest change analysis

Given the limited time series data available from Sentinel-2, the forest change analysis primarily focused on Landsat Imagery. In the study area of 624,892 ha, 48.8% (304,729 ha) was forest in 1985, which reduced to 43.2% (269,970 ha) in 2021 (see Fig 6c). Over 36 years, the Albemarle Peninsula study area, which includes private land as well as Pocosin Lakes National Wildlife Refuge and other preserved areas, experienced a net loss of 21.1% in forest cover, or approximately 64,220 ha. Approximately 10% (30,336 ha) were converted to cultivated land. Additionally, 33,884 ha (53%), i.e., 940 ha yr^-1 transitioned to either shrub (13423 ha, 372 ha yr^-1), marsh (9690 ha, 269 ha yr^-1), water (1364 ha, 38 ha yr^-1), and ghost forest (9405 ha, 261 ha yr^-1) by 2021. Specifically, 23,876 ha of forest were lost between 2010 and 2021 to marsh, ghost forest, and shrub, which is 1.5 times higher than the 16,968 ha loss between 1985 and 2010.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. (a) 1985 Coastal land cover, (b) 2021 land cover map, (c) pattern of coastal land cover change, (d-i) Comparison of classified data to existing land cover products.

The maps presented in g-i plots are our maps generated in 1985 and 2021, followed by the differenced map (i.e., area of forest-marsh gain or loss). The corresponding NLCD products are shown in d-f, followed by the differenced map, which shows the area of marsh-forest gain and loss from 1985 to 2021.

https://doi.org/10.1371/journal.pone.0357346.g006

Overall, we found land cover change was dynamic across our study area between 1985 and 2021 (Fig 6). We quantified a 28% increase in marsh between 1985 (26228 ha) and 2021 (33611 ha). Forests accounted for 43% of the marsh area increase between 2010 and 2021, and 60% of the marsh increase between 1985 and 2021. Much of the remaining increase in marsh was shrubland in 1985. Ghost forests had a net gain of 7561 ha (687 ha yr^-1) between 2010 and 2021, which is 2.5 times higher than the estimated net gain in ghost forest area between 1985 and 2010 (3087 ha). Our results show that forest-shrub transition (i.e., shrubland gain) was higher relative to forest-marsh and forest-ghost forest loss between 1985 and 2021, with a net gain of 24% (3638 ha) between 1985 and 2010 and a net gain of 10% (1877 ha) between 2010 and 2021. Lastly, we found some areas of forest gain, and cultivated land accounted for 47% of reforestation, mostly abandoned farmlands, followed by shrubland (21%) between 1985–2021.

We compared our results to existing products, such as the NLCD data. As presented in Fig 6, there is a close overlap in the forest-wetland boundary area loss. However, our study shows approximately 71% more area of forest-wetland loss compared to the NLCD estimated loss.

3.4. Patterns and predictors of forest change

In our analysis, each coefficient (estimate) represents the change in log odds of the outcome for a 1-unit increase in the predictor, holding other variables constant. Distance to the channel was particularly predictive of transitions between forest and shrub, and between forest and marsh. The result of the logistic regression model showed that for each additional meter from the channel, the odds of forest loss to marsh decreased by approximately 0.16% (β = −0.0016, p < .0001) and the odds of forest loss to shrubs decreased by 0.13% (β = −0.0013, p < .0001), see Fig 7. Salinity was one of the most significant predictors of ghost forest formation as well as forest loss to marsh. With a significant coefficient (p < .01), the model showed that every unit increase in salinity leads to a 0.0628 unit increase in the log odds of forest loss to ghost forest (β = 0.0628, p < .05) and even more loss to shrub (β = 0.1637, p < .0001). Like the distance to the channel, a 1-meter increase in elevation leads to a 1.54 unit decrease in the log odds of forest-to-marsh change (β = −1.536, p < 0.0001). Although slightly weaker statistically, the log odds of forest transition increased by 9% for ghost forests (β = 0.0906, p < 0.05) and 16% for shrubs (β = 0.1654, p < 0.05) per unit decrease in elevation, respectively. Ghost forest formation was significantly and positively influenced by RSLRR (β = 8.120, p < .0001) as was forest-marsh loss (β = 7.69, p < .0001). An increasing trend in the annual maximum temperature (slope) was observed, with the slope of annual maximum temperature strongly and positively influencing forest loss to ghost forest (β = 72, p < .0001), indicating that the warming trend leads to higher forest loss. However, it did not have any significant effects on the observed shrub and marsh transgression.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Predicted probability of forest loss for each driver.

The lines represent model-predicted probabilities of each change type across all environmental drivers with 95% confidence intervals (shaded areas).

https://doi.org/10.1371/journal.pone.0357346.g007

4.1. Phenological insights from coastal vegetation mapping with Landsat and Sentinel2

With accelerating rates of forest loss resulting in expansive ghost forests and shrub zones, our study demonstrates how remote sensing combined with CNN models can improve separability between vegetation classes and map land cover with spatial precision. Our findings support the use of DL models and phenology to improve land cover mapping and change analysis using Landsat and Sentinel imagery. We observed that both sensors demonstrated high classification performance across all classes. However, Sentinel-2 (F1 score = 91%) showed slightly higher accuracy for the ghost forest class compared to Landsat-8 (F1 score = 88%), possibly due to its higher resolution, distinct phenological attributes, and more training samples.

Our results demonstrate that combining winter and summer composite images with phenological and multispectral indices significantly improves the classifier’s ability to distinguish among the shrub, forest, and ghost forest categories. This improvement is due to seasonal differences in reflectance, which are better captured when images from both seasons are included, rather than using only summer images (see Fig 5 and S3 Fig). Accordingly, our approach to accounting for the phenological differences of coastal vegetation enables improved mapping accuracy, with phenological metrics contributing 36% (Fig 5) to the Sentinel 2 model, following the contribution of multispectral indices (37%) and 40% to the Landsat 8 model accuracy (Fig 5), followed by Winter indices at 29%. Earlier research indicated that incorporating phenology yields higher accuracy than traditional bi-seasonal methods, suggesting that land surface phenology is crucial for constraining intra-class variations in complex landscapes [27,32,33,45] and for separability between spectrally similar tree species [39]. In fact, phenology was the highest-performing predictor across all classes for both sensors (Fig 5), especially for the Shrub class. However, winter red and green also contributed to ghost forest class separation for Landsat 8 (S3 Fig) and Sentinel-2 (S4 Fig), respectively. As reported by Ury et al. (2021) [9], distinguishing between shrub and ghost forest was more challenging for their classifier due to spectral similarities, which further highlights the importance of phenology and the CNN used in our study. This result further shows the strong overlap between ghost forest, shrub, and forest as revealed in the confusion matrix (Table 5). Additionally, despite the heavy reliance on phenology, Landsat 8’s overall test accuracy is slightly lower with phenology included than without it, which further could mean that Landsat 8’s 16-day revisit interval may carry higher estimation variance that is sparser and noisier compared to Sentinel-2’s 5-day revisit period, which produces phenology metrics with distinguishable signals. However, Landsat compensates heavily for missing phenology in the non-phenology model by relying on correlated signals remaining in the spectral and multispectral indices (see S5(a)).

Our results underscore that Sentinel-2 data enables better mapping of forests, especially the ghost forest class, as revealed in the classification accuracy (Fig 5). This offers an opportunity for deep learning-based harmonization of Landsat and Sentinel-2 [80] or for using the output features from Sentinel-2 models to fine-tune or pre-train models for Landsat 8, potentially improving its classification accuracy for ghost forest or other vegetation transition mapping in future studies. Taking a closer look at phenology, we observed high variability in the phenological metrics for each vegetation class across the two satellite sensors (Table 4), which could be attributed to the mix of deciduous and evergreen trees. For instance, the forest class includes some planted pine plantations within the study site, alongside deciduous and evergreen forests that exhibit varying phenology throughout the year. Additionally, a few valid observations were available when clouds were masked out, and these were slightly improved by the 3-year composite.

Furthermore, the total land area classified for each class varies slightly between the two sensors, which may be due to the mixed-pixel effect or differences in band placement. For instance, the water area estimated by Sentinel-2 is about 5.5% lower than Landsat estimates, which may be due to a mix of water and non-water features in some pixels, while Sentinel-2 captures finer details, thereby reducing overestimation.

Comparing our results to the work of Chen and Kirwan (2022) [32], who have initially employed phenology for coastal vegetation mapping, our overall accuracy is higher, but their producers’ accuracy value for the Ghost Forest class (producer’s accuracy = 90%, user’s accuracy = 91.2%) is slightly higher than what we observed in our analysis (producer’s accuracy = 83%, user’s accuracy = 94%). A possible explanation for the difference, aside from not using the same training dataset, is that they do not have a Shrub class. In our study and previous studies [9], the ghost forest class was often misclassified as shrub, likely due to similarities in spectral signatures between dead tree stands and early successional vegetation. Additionally, about 11% of ghost forest pixels were misclassified as forests, as also observed by previous studies [4,32].

4.2. Comparing existing products and previous studies

While our study employed both Landsat and Sentinel-2, previous studies around the APP had only used Landsat imagery, which will serve as our basis of comparison (Table 6). The survey conducted by Ury et al. (2021) [9] using a traditional classifier achieved an overall accuracy of 92.2% for the Landsat 8 classifier, which is lower than our classifier at 94.3%. Smart et al. (2020) [4], on the other hand, show a CV-kappa (cross-validated Cohen’s kappa (CV-kappa) measures agreement between predicted and observed classifications while accounting for chance agreement) value of 0.60, which is lower than our kappa value of 0.95. While Smart et al. (2020) [4] also employed the traditional classifier (random forest model), another possible explanation could also be that their images were from 2014, while ours used 2021 images, indicating a 7-year gap.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Comparing classification results and changed areas to existing studies along the Albemarle Peninsula and existing products.

https://doi.org/10.1371/journal.pone.0357346.t006

Furthermore, we compare our results to the annual national land cover dataset, as also done in previous studies [32]. We found that the yearly forest loss to marsh from the NLCD data is 165 ha/yr-1, which is about 38.6% lower than our forest-marsh annual loss (269 ha yr-1). It should be noted that we only considered the emergent wetland as a marsh in the NLCD data. Another reason for the large difference could also be due to differences in accuracy, especially given that the overall accuracy of the NLCD 2021 model is 82.82, and the 1985 model is 83.25, with their wetland class overlapping highly with forest [81]. In both datasets, the spatial distribution of coastal forest loss within the study area is primarily concentrated along the coast, especially the eastern side of the Alligator River National Wildlife Refuge, exhibiting more significant forest loss compared to the western side of the Refuge (see Fig 6).

Comparing our result to that of Martinez et al. (2024) [62], over the same spatial extent, their annual estimate of Forest-Ghost forest loss is 188 ha yr-1, which is lower than our estimated forest-ghost forest transitions. Furthermore, we compared our results of forest-ghost loss and forest-marsh loss to area loss estimates across all years from 1985 to 2019, as reported by Ury et al (2021) [9]. We found that over the same spatial extent, the annual forest loss to marsh is 259 ha yr-1, which is 4% lower than our study estimates. On the other hand, our forest-ghost yearly forest loss is 40% lower than [9] ‘s estimates. One possible explanation for this, aside from the 2-year difference between our results, is the comparison across all years rather than between two years. Our higher classification accuracy could also influence the area estimated for each class, ultimately resulting in a difference in total land area.

4.3. Drivers of forest change and ghost forest formation

Sea-level rise caused significant forest die-back along the coastal plains of North Carolina. Our time-series land cover maps for 1985, 2010, and 2021 reveal accelerating rates of coastal forest loss, concurrent with the observed rise in regional global sea level over the last 4 decades, as well as climate extremes [32,82]. For instance, 23,876 ha of forest were lost between 2010 and 2021 to marsh, ghost forest, and shrub, which is 1.4 times higher than the 16,968 ha lost between 1985 and 2010. Numerous extreme events marked this timeframe (i.e., 2010–2021). The various parts of the study area experienced disturbances to varying degrees, with some occurring alone and others alongside multiple events. Previous studies also [9,62] noted that this area experienced prolonged drought between 2007 and 2011, and Hurricane Irene (both in 2011). While some of these extreme events occurred several years before 2021, some areas never fully recovered, having been pushed beyond their boundary into a new ecosystem state, despite being highly protected.

Specifically, ghost forests had a net gain of 7561 ha (687 ha yr^-1) between 2010 and 2021 (82% from forest, the remaining 18% primarily from shrub), which is 2.5 times higher than the estimated net gain in ghost forest area between 1985 and 2010 (3087 ha). Additionally, forests are being converted to cultivated land, especially in the western part of our study area, indicating that resilient forest areas could also be subject to persistent anthropogenic land-use changes.

We found the greatest number of ghost forest transitions and expansions of wetlands into coastal forests along the estuary margins and coastlines. In line with Ury et al. (2021) [9], forest loss was concentrated about 1 km inland from the coast margin across our study site. The forest transitions, however, were more extensive along the Alligator River National Wildlife Refuge (Fig 6i), with a greater concentration on the eastern side, indicating a more significant loss of forest compared to the western side. Some forest loss was also observed in the interiors of the study area. One of the major drivers of coastal forest loss, resulting in both ghost forests and marsh, is the proximity to the channel [9], reflecting the vulnerability of coastal forests to inundation and saltwater intrusion. We found that forests lost to marshes, shrubs, and ghost forests were concentrated at an average distance of 1.2 km from major water bodies, though we identified areas of ghost forest formation that appeared up to 17 km away from the coastline and approximately 60 km away from the Pamlico Sound, which interconnected drainage networks or heavy storms that result in periods of inundation could have facilitated. This result complements earlier findings in coastal North Carolina [4,9,83].

Similarly, increasing salinity and higher RSLR were positively associated with change, supporting the role of ecological stressors and climate-driven processes such as ghost forest formation. These results are not unexpected, as a similar region along the eastern US coast is already recognized as a global hotspot for accelerated relative sea-level rise (RSLR) [6]. For instance, the average RSLRR across all stations in the coastal plain of North Carolina increased from 3.5 mm/year in 2010 to 5.1 mm/year in 2021, indicating a continuous increase in RSLRR from 1985 to 2021 compared to the period from 1985 to 2010 (see S6a Fig). Smart et al. (2020) also reported declines in aboveground biomass and carbon near shorelines and in areas with higher salinities. Our study also demonstrated that these factors are stronger drivers of coastal forest change than elevation and slope alone [4,47]. This region has experienced major hurricane events, including Hurricane Irene in 2011, classified as a category 3, which passed through APP as a category 1 [9,62], and Hurricane Julia in 2016, which could contribute to high salinity through inundation.

Interestingly, while elevation significantly affects forest-to-marsh loss, with low-lying relief areas experiencing the greatest forest change, our study reveals that this is not the case for ghost forest formation, indicating that elevation has a weak statistical influence but a larger magnitude on forest conversion to ghost forests. Slope, on the other hand, has a stronger impact in predicting forest-ghost forest compared to forest-marsh. We observed greater forest loss in the downslope area (0–20 degrees, Fig 7) than on the mid- or upper slope. These areas appear wetter and are highly prone to inundation, while forest loss observed on the upper slope might be attributed to dryness and lag effects of historic drought or fire. Therefore, our observed increased forest loss in downslope areas supports the idea that topography influences gradients in moisture and stress sensitivity [84,85], where increasing slopes tend to reduce the duration of tidal flooding, ultimately decreasing salinization and waterlogging conditions [1,86].

We also note that not only are forests transitioning, but also the conversion of farmland to wetland, which could be related to the abandonment of farmlands in response to salinization facilitated by drainage ditches and other legacy land management practices [45,87]. Some of the farmland on the coastal plain of North Carolina has since been transformed into forested wetland ecosystems [88,89]. These changes indicate increased flooding and probable salinization of these areas due to sea level rise [8,9,27]. We also observed forest regrowth from shrubs, as also reported by Ury et al (2021) [9]. Specifically, the patches around the Pocosin Lakes National Wildlife Refuge, which represent a shrub and ghost forests class shown on our 1985 classified map, represent areas that experienced a wildfire event in 1985 (Allan Road Fires, Poulter et al., 2006), which were subsequently replaced by forests in 2021. Some of the ghost forest areas also overlapped with historic wildfire events, specifically the 2011 Pains Bay Fire along the Southern portion of the Alligator Refuge shoreline [4,62]. This pattern suggests that multiple mechanisms are contributing to ghost forest formation across the landscape. Some of these mechanisms, i.e., wildfire, are cyclical and include vegetation recovery [62,90]. While the spatial extent is relatively small, this highlights the limitations of distinguishing ghost forest cover from other disturbance-related land types or from scanty trees in coarse 30-meter-resolution imagery, given the small training datasets, especially for the pre-2021 images. With our results showing that elevation is not a strong predictor of ghost forest presence and that ghost forest distribution varies widely across elevation gradients, one explanation is that some ditches may serve as pathways for saltwater intrusion, even at higher elevations [18,19]. It’s interesting to note that the entire forest loss is relatively concentrated in the low-relief zone (<5 meters, see Fig 7), so the definition of high elevation might have been exaggerated here because even areas considered as “higher elevation” are still low in an absolute sense, with elevation also increasing inland towards the western side of our study. Additionally, potential misclassification between ghost forest and marsh, particularly in low-lying transitional zones, may further obscure the relationship between elevation and ghost forest occurrence. We also observed that some of our forest losses occurred farther inland, indicating that loss is not only concentrated on the shoreline, a result in line with Ury et al (2021) [9]’s findings. Additionally, fire-driven ghost forests may occur more frequently at higher elevations, while saltwater-driven ghost forests occur at lower elevations. This could be explored further in future work by disaggregating ghost forests based on disturbance drivers. Although significant, elevation is not a major predictor of shrub presence, indicating that we observed an expansion of shrubs in high-relief areas of the coastal landscape. However, increased salinity favors shrub growth, consistent with a previous study [91]. Their work shows that moderate salinity favors M. cerifera growth and expansion, with shrubs advancing to areas that may later become ghost forests or marshes. We also observed that as marshes move inland, they are replacing shrubs (see Fig 6c), which could indicate the development of a shrub-scrub marsh. This could indicate that some of the transitions are happening in areas that are too wet to turn into swamps and too dry or shallow to develop into marshes [92].