Scalable Mangrove Monitoring with Limited Field Data: Integrating MREDT and DACN-M

Abstract

Mangroves play a crucial ecological and economic role but face significant threats, particularly on Hainan Island, which has the highest mangrove species diversity in China. Remote sensing and AI techniques offer potential solutions for monitoring these ecosystems, but challenges persist due to difficult access for field sampling. To address these issues, we propose a novel model combining a Mangrove Rough Extraction Decision Tree (MREDT) and a Dynamic Attention Convolutional Network (DACN-M). Initially, we used drones and field surveys to conduct multiple observations in Dongzhaigang Nature Reserve, identifying the boundaries of the mangroves. Based on these features, we constructed the MREDT model to mitigate model failure caused by light instability, simplifying transfer to other study areas without requiring annotated samples or extensive field surveys. Next, we developed the DACN-M model, which refines the rough extraction features from MREDT and incorporates contextual information for more accurate detection. Experimental results demonstrate that our proposed method effectively differentiates mangroves from other vegetation, achieving F1 Scores above 75% and IoU values greater than 60% across six study areas. In conclusion, our proposed method not only accurately identifies and monitors mangrove distribution but also offers the significant advantage of being transferable to other study areas without the need for annotated samples or field surveys. This provides a robust and scalable solution for protecting and preserving critical mangrove ecosystems and supports effective conservation efforts in various regions.

1. Introducción

Mangroves are significant socio-economic entities that play a crucial role in coastal protection, storm and soil erosion resistance, flood and water flow regulation, carbon sequestration, and providing fertile habitats for marine species and animals [1,2,3,4]. Despite the importance of mangroves, there has been a substantial decline in their area over the past few decades [5,6]. This decline is attributed to agriculture and aquaculture, urban development, and logging, leading to significant degradation of wetland ecological functions [7,8]. Consequently, mangroves are facing an imminent threat of rapid disappearance [9,10].

Hainan Island, a typical mountainous tropical island, boasts diverse and extensive coastal wetlands and is the region with the highest diversity of mangrove species in China [11]. The use of remote sensing technology, such as satellite imagery or drones, can acquire ground image data. These data are crucial for identifying and monitoring the distribution and changes of mangroves [12,13,14].

Using vegetation indices to extract mangroves is an efficient method that can help researchers identify and monitor the distribution and health status of mangroves [14,15]. Some vegetation indices are widely used to distinguish mangroves from other vegetation or non-vegetation areas. Yang et al. [16] proposed the Enhanced Mangrove Vegetation Index (EMVI), which is used for mangrove mapping based on hyperspectral imagery. Tested in six typical areas, EMVI performed better in distinguishing mangroves from other vegetation compared to other indices. Baloloy et al. [17] proposed a Mangrove Vegetation Index (MVI) for rapid and accurate mangrove mapping using Sentinel-2 satellite data. In tests conducted in the Philippines and Japan, MVI achieved an accuracy of 92%. These indices, by analyzing the spectral characteristics of mangroves, can achieve rapid localization and dynamic monitoring of large areas of mangroves, providing scientific basis for ecological protection and management.

In the past, many researchers have utilized multidimensional spectral information in conjunction with machine learning techniques for pixel-level classification of mangroves [18,19]. These methods are well-suited for small sample training and exhibit high classification accuracy for small patches of mangroves and mangrove boundaries [20,21]. Jiang et al. [22] compared the application of high-spectral drones, WorldView-2 multispectral satellite, and their fused data in the classification of mangrove species. They employed Recursive Feature Elimination-Random Forest (RFE-RF) to select important spectral and texture features and used random forest (RF) and Support Vector Machine (SVM) algorithms for classification. The RF algorithm demonstrated higher accuracy and efficiency, and it was found that data fusion improved classification accuracy. Huang et al. [23] combined the Continuous Change Detection and Classification (CCDC) algorithm with the random forest model to track the changes in the mangroves located in the Beibu Gulf, Guangxi, from 1990 to 2022. They discovered that 814.57 hectares of mangroves changed, with Spartina alterniflora invasions and human activities leading to mangrove degradation. Woltz et al. [24] combined the KNN model with the random forest model to estimate above-ground carbon storage in forest structures. They drew a map depicting the predicted dominant species and determined changes (losses and gains) in mangrove coverage over the past few decades. The study concluded that the KNN model outperforms the RF model in characterizing species dominance.

The development of deep learning has driven a transformation in remote sensing analysis methods [25,26], providing robust analysis techniques for Remote Sensing image information [27,28,29]. This has also propelled advancements in the field of mangrove ecological monitoring and analysis. Compared to general machine learning, deep neural networks often possess higher accuracy [30]. Sun et al. [31] proposed a new deep learning U-net mapping model that integrates space, time, and method, using multi-sensor data to accurately draw maps with a resolution of 2 m. Applied in three locations in the Beibu Gulf of Guangxi, the average accuracy was 92.54%. Gabriel et al. [32] explored the integration effects of ResNet-101, VGG16, and Efficient-net-B7 within the U-net framework. Experimental results showed that the U-net model based on Efficient-net-B7 had superior performance, with an overall accuracy of 97.35% and a recall rate of 84.96%. This confirmed the practicality of this model in continuous observation and mapping reproduction of mangrove habitats.

However, the development of mangrove monitoring still faces the following challenges:

(1) Deep learning relies heavily on a large amount of manual annotation, which is labor-intensive and lacks sample labeling data [33]. The quality of existing public labeling data is poor, with low accuracy, which is not conducive to the training of deep learning networks [34,35]. In addition, the deep learning methods based on image convolution have a bottleneck in the recognition of small mangrove patches, and cannot be accurately distinguished in lower spatial resolution [36,37,38].

(2) Due to the dynamic nature of species composition and the spectral similarity between species, it is difficult to distinguish mangrove crowns from other terrestrial vegetation [31,39,40].

(3) Due to different research locations, species categories, image products, and classifiers, the comparability of past research is poor, and there is no clear best method for mangrove species classification [41,42]. These factors limit the effectiveness of these models to specific research fields, and in other regions, the models need to be retrained and fine-tuned according to factors such as image illumination conditions and time [16,43].

(4) Transfer learning and weak supervision methods have reduced the reliance on labeled data to some extent [44]. However, the quality of unlabeled data remains crucial for these approaches. Uneven data distribution can lead to the model learning inaccurate or biased features. Additionally, these methods exhibit poor generalization across different regions [45]. This is particularly problematic when there are significant differences in environment, species, or spectral characteristics; a model trained in one region may not be directly applicable to another. When dealing with complex terrain and spectral backgrounds, these methods are prone to false positives and false negatives, affecting classification reliability [46].

In response to the existing challenges, this paper proposes a mangrove detection method by integrating the spectral characteristics of mangroves. Firstly, based on multilinear regression analysis of mangrove spectral properties, a Mangrove Rough Extraction Decision Tree (MREDT) is constructed. MREDT replaces the need for a large amount of manual annotation and is used for the automatic collection of mangrove samples. On this basis, we designed a Dynamic Attention Convolutional Network Model (DACN-M) for mining mangrove spectral correlation information. By using the superimposed residual method, it effectively distinguishes mangroves from other vegetation, thus enabling fine extraction of mangrove areas. To verify the reliability of the proposed method, we collected real samples from multiple mangrove scenes and compared them with the model’s predictions using indicators such as confusion matrix and Intersection over Union (IoU). The results show that compared with traditional methods such as KNN, RF, and SVM, the combination of MREDT and DACN-M significantly improves the classification accuracy of mangrove crowns, indicating that the method in this paper can accurately identify and monitor the distribution of mangroves.

The main contributions of this method are as follows:

(1) Reduction of Manual Annotation Requirements: Our method, using the Mangrove Rough Extraction Decision Tree (MREDT), eliminates the need for extensive manual annotation, enabling automatic collection of mangrove samples and simplifying the transfer to other study areas without requiring annotated samples or extensive field surveys.

(2) Enhanced Classification Accuracy: By integrating the MREDT with the Dynamic Attention Convolutional Network Model (DACN-M), our approach effectively distinguishes mangroves from other vegetation. This combination significantly improves the classification accuracy of mangrove crowns compared to traditional methods such as K-Nearest Neighbor (KNN), Random Forest (RF), and Support Vector Machines (SVM).

(3) Monitoring Growth and Health Status: Our method’s predictive capabilities extend beyond mere classification; it also provides insights into the growth density and health status of mangroves. This makes it particularly useful for monitoring the growth and changes in newly established mangrove areas, such as those in Dongzhaigang Nature Reserve on Hainan Island.

2. Material

2.1. Study Area

Hainan Island in China has the highest diversity of mangrove species. Among the 27 true mangrove species in China, 26 are found on Hainan Island [40]. This study focused on six selected mangrove areas on Hainan Island, as shown in Figure 1.

The first and second regions are located in the Dongzhaigang National Nature Reserve in Hainan (110°32′–110°37′ E, 19°51′–20°1′ N). The reserve covers a total area of 3337.6 hectares, making it the largest coastal mangrove forest in China. It comprises 35 species across 18 families, including 24 true mangrove species from 10 families and 11 semi-mangrove species from 8 families. It is the first national mangrove nature reserve in Hainan Province [47]. The third research area is the Sibiwang Mangrove Wetland within the Dongfang black-faced Spoonbill Reserve (08°37′24″~108°40′15″, 19°11′33″~19°13′20″), covering an area of 3750 mu (approximately 250 hectares). The fourth research area is the Xinyingwan Mangrove Nature Reserve in Danzhou City, with a total area of 115.4 hectares, dominated by the Rhizophora stylosa species. The fifth research area is Xinying Mangrove National Wetland Park, with a total area of 507.05 hectares, including a wetland area of 310.59 hectares (with a natural forest mangrove area of 126.9 hectares). It is home to 18 species of mangroves, including 13 species from seven families of true mangroves (including one introduced species) and 5 species from five families of semi-mangroves. The sixth research area is the Qinglan Port Nature Reserve (19°34′ N, 110°45′ E), which boasts the highest number of mangrove species in China. It has 24 species of true mangroves and 10 species of semi-mangroves, accounting for 88.89% and 100% of the total number of true and semi-mangrove species in China, respectively. The total mangrove area in the Qinglan Port Nature Reserve is approximately 835.5 hectares, with a research area of 386.32 hectares, including a mangrove area of 209.99 hectares. The research area map is shown in Figure 1.

2.2. Data

Sentinel-2 imagery, launched by the European Space Agency in 2015, serves as a pivotal tool in environmental monitoring. This study utilized remote sensing data from the Sentinel-2 MSI (Multi-Spectral Instrument) sensor on board the Sentinel-2 satellite as the primary data source. The collected Level-1C data underwent atmospheric and glint correction using the Sen2Cor toolkit, converting it to Level-2A surface reflectance data to provide more accurate surface property information [48]. Sentinel-2 features 13 spectral bands: Bands 2, 3, 4, and 8 at 10 m resolution; Bands 5, 6, 7, 8A, 11, and 12 at 20 m resolution; and Bands 1, 9, and 10 at 60 m resolution, with a revisit time of 5 days. The satellite data for different study areas were acquired as follows: on 21 June 2023, for the Dongfang Black-faced Spoonbill Reserve; on 6 July 2023, for Xinying Mangrove National Wetland Park and Xinyingwan Mangrove Nature Reserve; and on 7 April 2022, for Dongzhaigang Nature Reserve and Wenchang Qinglan Port (accessed on 10 June 2024, at https://browser.dataspace.copernicus.eu). During data collection, surface illumination was sufficient, and cloud cover was less than 10% over the study areas. Classification samples were generated through visual interpretation. High spatial resolution imagery was accessible via Google Earth and ArcMap. Additionally, the Global Mangrove Watch (GMW) [49] provided significant references.

3. Methods

3.1. Support Vector Machine (SVM)

Support Vector Machine (SVM) is a supervised learning algorithm used for classification tasks [50]. The core idea of SVM is to find an optimal hyperplane that separates data points of different classes, thereby achieving classification. The objective of SVM is to maximize the distance between the nearest data points (support vectors) on either side of the hyperplane, known as the margin, to improve classification accuracy and generalization ability. For linearly separable data, SVM directly finds a linear hyperplane to separate the classes. For non-linear data, SVM uses kernel functions (such as polynomial kernels, radial basis function kernels, etc.) to map the data into a higher-dimensional space. In this high-dimensional space, a linear hyperplane is found to perform complex classification tasks.

3.2. Random Forests (RF)

Random forest (RF) is an ensemble learning algorithm that was first proposed in 2001 [51]. Since then, it has gained widespread application in mangrove classification research [22]. RF combines the predictions of multiple decision trees to classify and recognize spectral data. The algorithm randomly samples a subset of the original dataset with replacement to create a training set. Simultaneously, a random subset of features is selected as input. Multiple decision trees are then constructed using these training samples and features, and the final classification result is determined through a voting mechanism.

3.3. K-Nearest Neighbors (KNN)

The K-Nearest Neighbors (KNN) algorithm is a commonly used non-parametric supervised learning algorithm that classifies based on the similarity between samples [52]. The basic idea is to assign a new spectral sample to the class of its K nearest neighbors. In the spectral recognition of mangroves, KNN can classify based on the spectral similarity between samples without assuming the probability distribution of the data in advance. It is suitable for nonlinear and complex classification problems.

3.4. Linear Model

Linear regression is a commonly used supervised learning algorithm that is also applied in the spectral recognition of mangroves using remote sensing data [53]. The linear regression algorithm aims to establish a linear model to classify and predict based on the linear relationship between the features of the spectral data and the reflectance. In the spectral recognition of mangroves, the linear regression algorithm can predict the vegetation type of mangroves and other related attributes based on the features of the spectral data, enabling classification and prediction.

In this study, the expression for spectral regression discrimination is as follows:

$${\mathrm{H}=\mathrm{a}}_1\mathrm{L}\left({\mathrm{\lambda }}_1\right){+\mathrm{a}}_2\mathrm{L}\left({\mathrm{\lambda }}_2\right){+\mathrm{L}+\mathrm{a}}_8\mathrm{L}\left({\mathrm{\lambda }}_8\right){+\mathrm{a}}_9$$

(1)

Here, H represents the linear prediction result, a₁ to a₉ represent the fitting parameters, and λ₁ to λ₈ represent the spectral data.

3.5. Vegetation Index

Vegetation indices are essential tools in remote sensing for assessing various aspects of plant health and distribution. The Normalized Difference Vegetation Index (NDVI) is one of the most widely used indices [54,55], leveraging the contrast between red and near-infrared reflectance to provide a measure of vegetation density and health. The Difference Vegetation Index (DVI) offers a simpler alternative, directly subtracting red reflectance from near-infrared reflectance [56,57]. This index is particularly useful in areas with dense vegetation cover. The Modified Normalized Difference Vegetation Index (MNDVI) builds upon the NDVI concept but incorporates additional spectral bands to enhance its sensitivity to variations in canopy structure and chlorophyll content [58,59]. These indices, among others, play crucial roles in monitoring vegetation dynamics, assessing biomass, and tracking changes in plant health over time, providing valuable insights for ecological research, agricultural management, and environmental monitoring.

$$\mathrm{NDVI}={\displaystyle \frac{\left(\mathrm{NIR}-\mathrm{R}\right)}{\left(\mathrm{NIR}+\mathrm{R}\right)}}$$

(2)

$$\mathrm{DVI}=\mathrm{NIR}-\mathrm{R}$$

(3)

$$\mathrm{MNDWI}={\displaystyle \frac{\left(\mathrm{G}-\mathrm{SWIR}\right)}{\left(\mathrm{G}+\mathrm{SWIR}\right)}}$$

(4)

where R represents the red band, NIR represents the near-infrared band, and SWIR represents the short-wave infrared band.

3.6. Proposed Method

The key steps of the experimental scheme proposed in this study are shown in Figure 2. The experiment mainly includes two key parts: the construction of the MREDT general mangrove feature extraction model, and the application of DACN-M in different scenarios.

3.6.1. Spectral Characteristics Analysis of Mangroves, Vegetation, and Water Bodies

Figure 3 examines the spectral characteristics of mangroves, other terrestrial vegetation, and water bodies in Sentinel-2 collected data. The data used are from the Dongzhai Port National Nature Reserve. Considering that mangroves grow in the intertidal zone, mixed pixels at the growth boundaries and the spectra of water bodies and tidal flats can affect the analysis. Therefore, we randomly selected 5600 mangrove sampling points and 3600 other vegetation training sampling points from the identified mangrove areas. Scatter plots were drawn with different bands as the main viewing angles for observation. The mean of all points was calculated to plot the spectral curves, as shown in Figure 3.

The y-axis represents the collected remote sensing spectral reflectance (Rrs), while the x-axis denotes the different mangrove samples. It can be seen that bands B02, B03, B04, and B05 have similar distribution characteristics, bands B06, B07, and B08A have similar distribution characteristics, and bands B11 and B12 have similar distribution characteristics. Based on observations, we categorized the bands into four groups: B2 with B4; B3 with B5; B6 with B7 with B8; and B11 with B12. This grouping is based on the trends shown in scatter plots (a)–(i) that have similar change characteristics.

Mangroves are a unique wetland ecosystem with distinct vegetation composition and spectral characteristics. As a densely grown vegetation, mangrove crowns exhibit high absorption in the visible light range (400 nm–700 nm), including the blue, green, red, and red edge bands provided by the Sentinel-2 satellite [60]. This reflection characteristic provides an indicator for assessing the vegetation coverage and type of mangroves [35]. In the near-infrared range, the reflectance of mangroves increases sharply, reaching a peak of around 30%–40% in the multispectral band B8A (865 nm) of the Sentinel-2 satellite. Additionally, mangroves grow in coastal areas and have close contact with water bodies, so their band reflectance is often influenced by the radiometric characteristics of water [17]. Compared to other terrestrial vegetation, mangroves have higher absorption of blue light and lower absorption of green and red light [61,62]. Furthermore, as the wavelength increases, water bodies exhibit higher absorption in the short-wave infrared range, and the water content has a significant impact on the reflectance in this band [63]. Mangroves show significant differences in reflectance values in the SWIR region compared to other vegetation types. Previous research has demonstrated that the SWIR1 band can effectively express the unique greenness and moisture information characteristic of mangroves [64,65].

Based on observational data, vegetation exhibits an absorption peak in the B8A band (Figure 3j). The differences between Band B8A and both lower wavelengths (B02, B03, B04, and B05) and higher wavelengths (B11 and B12) contribute to distinguishing mangroves from other targets.

3.6.2. Mangrove Rough Extraction Decision Tree Model (MREDT)

Many studies have shown that combining various spectral indices helps to reduce misclassification between similar types of land cover [66,67]. The green light, red-edge, and water absorption regions are the optimal wavelengths for species identification [68]. Similarly, in our previous observations and analyses, bands B3, B8A, B11, and B12 were proven to be crucial in distinguishing mangroves from other vegetation types. To obtain a richer range of variation options, we additionally added a perspective feature in the B2, B4 grouping. Based on these features, we jointly constructed the MREDT model. We conducted our experiments following the process shown in Figure 4.

As shown in Figure 4, based on the selected bands, we conducted numerous combination experiments, including operations such as addition, subtraction, multiplication, and division. B2 and B4 were each combined with other bands for experimentation. This combination method includes multiple repetitions to achieve better combination effects. The obtained combination results reached hundreds of types. Then, screening was performed, and different results were integrated using linear regression methods to obtain corresponding parameters. Due to the large amount of computation, this step requires computer assistance. Finally, through heatmaps from different perspectives, we intuitively observed the effects of different combinations on mangroves. Equation (5) provides an effective method for mangrove feature analysis, as shown in Equation (5).

$$\mathrm{f}\left(\mathrm{x}\right)=396.48\times \left({\displaystyle \frac{{\mathrm{\lambda }}_3}{{\mathrm{\lambda }}_2{+\mathrm{\lambda }}_5}}+{\displaystyle \frac{{\mathrm{\lambda }}_3}{{\mathrm{\lambda }}_1{+\mathrm{\lambda }}_2}}\right)-{2\mathrm{\lambda }}_4+486.28$$

(5)

The symbol ρ_λ represents the band center reflectance, while λ₁, λ₂, λ_3, λ₄, and λ₅ denote the central wavelengths at 490 nm, 560 nm, 865 nm, 1610 nm, and 2190 nm, respectively, corresponding to the electromagnetic wavelengths captured by the B2, B3, B8A, B11, and B12 bands. The ratio of B8A to (B3 + B12) captures the differences between mangrove canopies and other vegetation across lower and higher wavelengths. Additionally, the ratio of B8A to (B2 + B3) further enhances the distinction of mangrove canopies from other vegetation at lower wavelengths. Moreover, the significant differences observed between mangroves and other vegetation in the B11 band are also incorporated as factors in Equation (5). This method is more sensitive to the spectral characteristics of mangroves compared to the general vegetation index.

To eliminate the interference of water bodies on the experimental results, the Normalized Difference Water Index (NDWI) [69] is introduced to filter out water body information.

$$\mathrm{NDWI}={\displaystyle \frac{\mathrm{G}-\mathrm{NIR}}{\mathrm{G}+\mathrm{NIR}}}$$

(6)

G represents the green band, and NIR represents the near-infrared band. The f(x) operator and NDWI together constitute the decision scheme of MREDT. Subsequently, the extracted regions require further boundary ablation processing (Figure 3d). In the field of computer image processing, morphological erosion is used to handle structural elements in digital images [70]. Morphological erosion aids in the separation and differentiation of adjacent objects or elements within an image, thus refining the target features in remote sensing imagery by ablating boundary information. Defined through set theory, where A represents a set in a binarized image and B represents a structural element (also a set), morphological erosion can be defined as the result of A being eroded by B, denoted as A⊖B (Equation (7)).

$$\mathrm{A}\ominus \mathrm{B}=\left\{\mathrm{z}\in \mathrm{E}|{\left(\mathrm{B}\right)}_{\mathrm{z}}\subseteq \mathrm{A}\right\}$$

(7)

In this equation, E represents Euclidean space, z is an unknown pixel in E, (B)z is the translation of structural element B relative to pixel position z, and ⊆ denotes a subset. When structural element B is entirely contained within set A, pixel position z will be included in the erosion result. This part of the processing approach is illustrated in Figure 3d.

3.6.3. Dynamic Attention Convolutional Network for Mangrove Feature Analysis (DACN-M)

DACN-M is a fully convolutional neural network. The main architecture of the network is divided into two core parts: the feature dimensionality increases section and the feature dimensionality reduction section (Figure 3d).

Assuming the input feature matrix is X∈ℝ^H×W and the convolution kernel is K∈ℝ^1×n; the convolution operation is performed by sliding on the two-dimensional data. The mathematical expression is as follows:

$${\left(\mathrm{X}\times \mathrm{K}\right)}_{\mathrm{i},\mathrm{j}}={\displaystyle {\sum }_{\mathrm{m}=0}^{\mathrm{n}}{\mathrm{K}}_{1,\mathrm{m}+1}\cdot }{\mathrm{X}}_{\mathrm{i},\mathrm{j}+\mathrm{m}-1}$$

(8)

Here, X × K represents the convolution operation, and (i, j) is the position of the output feature matrix. It is assumed that we have applied appropriate padding to the input feature X.

The section on feature dimensionality increases employs convolution operations to facilitate interactions among input channels, thereby excavating more expression patterns of channel features. This module undergoes four convolution groups with a scale of 1 × 3 and a channel attention module. Morphological erosion DACN-M retains the results during the computation process from low-feature stages to high-feature stages and superimposes them onto the features of the same size in the feature dimensionality reduction section. The channel attention mechanism is a key component of the feature amplification module, which includes global pooling and multilayer perceptron (Figure 3e). This module is applied in high-feature stages, emphasizing or reducing the importance of different input channel features, thereby enhancing the model’s understanding of key features and improving the model’s stability and accuracy (Equation (9)).

$$\mathrm{X}\prime =\mathrm{X}\odot \left(\mathrm{\sigma }{\left({\mathrm{W}}_2\cdot \left(\mathrm{ReLU}\left({\mathrm{W}}_1\cdot \left(\mathrm{Maxpool}\left(\mathrm{X}\right)\right){+\mathrm{b}}_1\right)\right){+\mathrm{b}}_2\right)}^{\mathrm{T}}\right)$$

(9)

Here, X∈ℝ^H×W represents the input feature matrix, W∈ℝ^F×W denotes the weight matrix of the fully connected layer, and b∈ℝ^F signifies the bias. Maxpool stands for the global maximum pooling operation, and σ represents the Sigmoid activation function (Equation (9)). The attention module learns from the input features X and transforms it into a weighted feature matrix $\mathrm{X}\prime$.

$$\mathrm{\sigma }={\displaystyle \frac{1}{1+{\mathrm{e}}^{-\mathrm{x}}}}$$

(10)

The feature dimensionality reduction section undergoes processing through three convolution groups with a scale of 1 × 3, gradually reducing the dimensionality of the features and obtaining the final regression results. During the process from high-feature stages to low-feature stages, the feature dimensionality increase section provides the semantics of the original features, ensuring the stability of feature regression in the final stage of DACN-M regression. The computation results at different stages of DACN-M all apply the activation of the Rectified Linear Unit (RELU) nonlinear mapping method. It has a simple calculation method, but it has gradient stability, enhancing the nonlinear fitting effect of the model (Equation (11)).

$$\mathrm{Relu}\left(\mathrm{x}\right)=\left\{\begin{array}{cc}\mathrm{x}& \mathrm{x}\ge 0\\ 0& \mathrm{x}<0\end{array}\right.$$

(11)

The strategy of overlaying residuals is key to the stability and strong generalization capability of DACN-M. Typically, model fitting requires a large number of positive and negative samples to enhance generalization (e.g., semi-mangrove windbreaks that are easily confused with mangrove plants). However, this task is enormous, and MREDT can only provide a limited number of positive samples for DACN-M weight training. Therefore, DACN-M does not use conventional classification labels (usually with positive samples labeled as 1 and negative samples as 0). Let the samples collected by MREDT be defined as set E ⊇ [λ₁, λ₂, ⋯, λ_n], where λ represents reflectance data at different wavelengths. In each training session, DACN-M selects a particular band λ_i from E as the fitting variable Y, with the remaining band information serving as the training data X. The model is trained using mean squared error (MSE) as the loss function. In this manner, for each channel, we obtain a set of model weights. These weights are defined as the predicted correlation values of other mangrove bands to λ_i under DACN-M observation, with a residual ε between the predicted and true values. Furthermore, if the residual ε is smaller, the likelihood that the pixel information represents mangroves, as observed from band λ_i, is higher. Consequently, overlaying the residual results integrates the observations of DACN-M across different bands.

The “threshold segmentation of overlay residual results” method effectively addresses the issue of model generalization, allowing DACN-M to work with MREDT to achieve accurate segmentation without relying heavily on observational samples.

3.7. Experimental Setup and Evaluation Methods

To compare the effectiveness of mangrove monitoring methods, an equal-distance sampling method based on image analysis was adopted. For each study area, a validation point set was created by extracting a sample point every three pixels. Through visual interpretation, the areas of mangroves were manually labeled. The visual interpretation was based on Google phase images with high spatial resolution and combined with the research findings and field surveys of related studies for attribute labeling. Figure 5 shows more detailed results of the experiment’s output.

The red point set represents non-mangrove canopies, while the white point set represents mangrove canopies. This sampling method effectively ensures that the class distribution of the validation points is consistent with the original image. It also guarantees the randomness and independence of the validation points, which is beneficial for evaluating the accuracy of different monitoring methods in mangrove identification.

In this study, the confusion matrix is used to evaluate the accuracy of the classification results, and the specific evaluation indexes include Intersection over Union (IoU) and F1-score. The specific formulas are as follows:

$$\mathrm{MIoU}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}}$$

(12)

$$\mathrm{F}1 \mathrm{Score}={\displaystyle \frac{2\times \mathrm{TP}}{2\times \mathrm{TP}+\mathrm{FP}+\mathrm{FN}}}$$

(13)

where TP indicates the number of true positive samples, FP indicates the number of false positive samples, and FN indicates the number of false negative samples.

4. Result

4.1. MREDT Extraction Results

The intermediate output results of the original images transformed by the MREDT module are shown in Figure 6. The first row displays the visual effect image of true color band composition (used for observation). The remaining rows show heatmaps of calculation results for different indices.

As shown in Figure 6, we adjusted the mapping color parameters to make each index tend towards green in areas detected as vegetation, blue in areas with high water content, and yellow transitioning closer to land. This is performed to obtain a more suitable visual effect for comparison. The calculation results for MRDET and DVI indices are reduced by ten thousand times to achieve appropriate color mapping.

NDVI and DVI have been extensively validated and can be used to detect vegetation greenness and health. However, they have a critical problem in distinguishing mangroves, as they cannot differentiate between other vegetation and mangrove features. MNDVI has been commonly used in the past to detect submerged vegetation, which provides a unique advantage for mangroves. However, its results are mixed with tidal flats and some land areas. This is particularly severe in Qinglan Port and Dongzhaigang Nature Reserve Region 2. The results in the heatmap indicate that under the influence of MREDT, the mangrove canopy results are concentrated in the 0–0.1 numerical range, significantly different from non-mangrove areas (compared to the marked samples in Figure 6). Although this transformation result is easily confused with water bodies, NDWI effectively eliminates this interference, ensuring the reliable stability of the mangrove extraction results.

4.2. Compare the Optimal Parameters of the Model

KNN, RF, and SVM rely on different parameters. To compare the best predictive performance of different algorithms using DACN-M, 500 positive samples and 500 negative samples were extracted from the marked data in each research area from Figure 5 for experimentation. The experiment ensures that KNN, RF, and SVM use the same data during different parameter experiments, and the sampling points are spatially distributed randomly and uniformly.

The SVM model mainly relies on the key parameters’ “kernel” and “C”. The line graph comparing results under different parameter settings is shown in Figure 7:

Among the multiple options for kernel, “rbf” shows the best performance. The gamma parameter, acting on the basis of rbf, adopts a dynamic strategy, which is marked as “scale”. However, based on rbf, the regularization parameter (C) does not show a clear trend of change. In the 0.5–1 interval, the effect in the Qinglan Port area showed a significant decrease, while losses also occurred in Dongzhaigang area 2. However, in contrast, Dongfang Black-faced and Dongzhaigang area 1 showed an upward trend in the 0.5–2 interval. In a comprehensive comparison, C = 1 is chosen as the experimental parameter, as it performs more stably.

The KNN model mainly relies on the “n_neighbors” parameter. The line graph comparing results under different parameter settings is shown in Figure 8:

The results show that a larger n_neighbors is not always better, as oscillations occur. However, it can be observed that when n_neighbors > 5, the effect curve becomes more stable. Therefore, n_neighbors = 8 is chosen as the best practice. At this point, the best results are achieved in Qinglan Port and Dongzhaigang area 2, and it also produces significant effects in other areas.

In the RF model, dynamic strategies were developed for the n_estimators, max_features, and max_depth parameters, which are superior to manual tuning. Therefore, the “n_estimators” parameter, which the RF model mainly relies on, is chosen for comparison to verify its effectiveness. The line graph comparing results under different parameter settings is shown in Figure 9.

As shown in the figure, the parameter setting at 0 represents dynamic n_estimators. The RF model shows very stable results under different parameter settings. It can be seen that smaller n_estimators lead to minor losses in the model’s performance. However, this change has almost no effect when n_estimators > 100. The dynamic strategy also ensures this effect.

4.3. DACN-M Extraction Results

The detection performance of the SVM, RF, KNN, and DACN-M models was evaluated using labeled validation samples, with the results presented in

Table 1. Classification accuracy assessment.

	Mangrove (pcs)	Others (pcs)	Model	Precision	Recall	F1 Score	IoU
Dongzhaigang Nature Reserve Region 1	1933	18422	SVM	93.90%	74.26%	82.93%	70.84%
			RF	60.27%	83.04%	69.84%	53.66%
			KNN	90.95%	75.13%	82.28%	69.90%
			DACN-M	86.45%	86.09%	86.27%	75.85%
Dongzhaigang Nature Reserve Region 2	1825	7731	SVM	54.23%	76.09%	63.33%	46.34%
			RF	29.82%	81.74%	43.70%	27.96%
			KNN	56.80%	68.78%	62.22%	45.16%
			DACN-M	76.68%	77.10%	76.89%	62.64%
Dongfang black-faced Spoonbill Nature Reserve	216	5004	SVM	86.57%	82.02%	84.23%	72.76%
			RF	68.52%	89.70%	77.69%	63.52%
			KNN	87.04%	75.20%	80.69%	67.63%
			DACN-M	89.35%	90.61%	89.98%	81.78%
Xinying Bay Nature Reserve	941	25469	SVM	72.90%	75.38%	74.12%	58.88%
			RF	75.13%	72.51%	73.80%	58.48%
			KNN	74.60%	76.81%	75.69%	60.88%
			DACN-M	88.42%	94.55%	91.38%	84.13%
Danzhou Xinying	575	7957	SVM	57.57%	68.11%	62.39%	45.34%
			RF	37.74%	82.20%	51.73%	34.89%
			KNN	61.04%	67.63%	64.17%	47.24%
			DACN-M	97.57%	87.93%	92.50%	86.04%
Wenchang Qinglan Port	1247	5253	SVM	67.07%	89.45%	76.66%	62.15%
			RF	44.12%	94.36%	60.13%	42.99%
			KNN	66.27%	86.53%	75.06%	60.07%
			DACN-M	95.44%	88.77%	91.99%	85.16%

.

Among these, the bolded data represents the optimal indicator data. The results shown in Table 1 demonstrate that DACN-M, through multi-level feature extraction and the fusion of contextual information, can uncover the semantic differences between mangrove canopies and non-mangrove canopies. In terms of Precision, Recall, F1 Score, and IoU evaluation metrics, DACN-M exhibits an absolute advantage and achieves excellent results. The Precision values for all categories are greater than 85%, and the IoU exceeds 70%, indicating a high level of consistency between the annotated ground truth and the detection results obtained by DACN-M.

In Figure 10, we present a visualization of the segmentation results obtained from various algorithms. The column labeled “RGB band composite image” displays the composite images of the RGB bands for each study area. The “Labels” column shows the markers used for segmentation evaluation, with white indicating mangrove crowns. The remaining columns display the prediction performance of the DACN-M, KNN, RF, and SVM models on the collected dataset.

The portions of the study area identified as mangroves by the model will be marked in white (background color). When the “Labels” are marked as mangrove crowns and the model detects them as such, they are displayed in green (correct detection). When the “Labels” are marked as mangrove crowns, but the model detects them as non-mangrove crowns, they are displayed in blue (missed detection). When the “Labels” are marked as non-mangrove crowns, but the model detects them as mangrove crowns, they are displayed in red (false detection). When the “Labels” are marked as non-mangrove crowns and the model also detects them as such, they are displayed in black. In Figure 10, we provide an intuitive observation of the segmentation results from various models. Superior models will have a higher number of green markers (correct detection) and the least number of red (false detection) and blue markers (missed detection).

The “Dongzhaigang Nature Reserve”, “Xinying Bay Nature Reserve”, and “Qinglan Port Nature Reserve” contain a significant amount of other terrestrial vegetation, resulting in a complex species composition. This complexity, to some extent, affects the prediction results of the models, with all models achieving lower accuracy. The “Dongfang black-faced Spoonbill Nature Reserve” and “Xinying Bay Nature Reserve” have a larger distribution of water bodies and contain numerous lands and building areas. The features of the mangrove crowns are distinct, leading to a higher overall accuracy in the detection by all models.

In comparison, RF performs the worst in differentiating between vegetation types. In the case of Haikou Yanfeng Town, the IoU value for RF detection is as low as 36.5%. This is mainly due to overestimating the mangrove area and misclassifying many non-mangrove canopies as mangroves. KNN performs better than RF but slightly worse than SVM. Its main disadvantage lies in estimating relevance solely based on feature distances, leading to misclassification of semi-mangrove or mangrove-like vegetation spectra, such as Casuarina equisetifolia Fors and Talipariti tiliaceum. This discrepancy is particularly evident in the case of the Dongfang black-faced Spoonbill Reserve, where the difference in segmentation IoU results between KNN and SVM is within 2% in most cases.

SVM benefits from its kernel function structure and performs better than KNN and RF. However, this model excessively relies on the quantity and quality of samples. In the results of this experiment, the SVM algorithm achieves IoU values exceeding 47.85% in all scenes. DACN-M addresses these limitations and improves upon them.

In various scenes, DACN-M demonstrates improvements in F-score and IoU values. Moreover, the difference between Precision and Recall decreases, indicating more stable detection results. DACN-M effectively compensates for the shortcomings of SVM and provides better performance in terms of segmentation accuracy and stability.

5. Discussion

5.1. Prediction of MREDT and DACN-M at Different Times

To evaluate the robustness and effectiveness of the proposed MREDT and DACN-M models, we assessed their performance in extracting mangrove information from satellite images captured at different time points. The study was conducted in the Dongzhai Port Nature Reserve using satellite images from 18 May 2019; 7 May 2020; 21 June 2021; 7 April 2022; and 17 May 2023. All images were cloud-free to ensure optimal analysis conditions [71]. The MREDT model was applied to these images to obtain coarse extractions of mangrove boundaries. The DACN-M model was then used to further improve the accuracy of mangrove detection by integrating contextual information. The extraction results are shown in Figure 11.

As shown in the heatmap results in Figure 11, the models demonstrated stable and reliable performance across all examined time points, accurately distinguishing mangroves from other vegetation types. The consistent performance of the MREDT and DACN-M models over different periods highlights their ability to adapt to varying environmental conditions, vegetation dynamics, and other factors affecting the spectral and spatial characteristics of mangrove ecosystems.

Figure 12 shows the detection results of the method presented in this paper on 23 March 2023; 17 May 2023; 1 July 2023; 4 October 2023; and 8 December 2023. These results demonstrate the response across different seasons.

As shown in the heatmap results in Figure 12, the MREDT and DACN-M models demonstrate stable and reliable performance at all examined time points. This indicates that they exhibit relatively stable capabilities even under different seasonal change factors. Furthermore, when using this model, ground surface data can reduce time constraints.

This stability is a significant advantage as it allows the proposed framework to be directly applied for long-term monitoring and assessment of mangrove habitats without extensive retraining or model adjustment. Furthermore, the successful application of the MREDT and DACN-M models to satellite images from 2019 to 2023 demonstrates their potential for providing continuous and comprehensive monitoring of mangrove distribution. This capability is particularly valuable for understanding the temporal dynamics of mangrove ecosystems, detecting changes, and effectively supporting the formulation of conservation and management strategies. However, since the experiments conducted in this study are located in the tropics, the validity of experiments in more latitudinal dimensions needs further validation in subsequent research.

5.2. The Advantage of DACN-M in Residual Strategy

Although the residual strategy provides effective positive feedback for DACN-M predictions, this strategy is not applicable to general regression tasks of other models [72]. To validate this conclusion, we applied the residual strategy to RF, SVM, KNN, and linear models and compared their regression results with a control group. To ensure fairness, we shared the coarse extraction results of MREDT. (In practice, obtaining such a large number of accurate positive samples would require a significant amount of time and may not be suitable for data collected from different devices at different times. This is where the advantage of MREDT lies.)

The segmentation evaluation of DACN-M and other comparative models in the mangrove growth area in northern Hainan are shown in

Table 2. Comparison of prediction results of different algorithms.

	Model	Residual Threshold	Precision	Recall	F1 Score	IoU
Dongzhaigang Nature Reserve Region 1	SVM	3200	99.78%	18.98%	31.9%	18.98%
	Linear	40	56.54%	13.82%	22.22%	12.5%
	RF	250	89.81%	18.19%	30.25%	17.82%
	KNN	610	98.76%	15.16%	26.28%	15.13%
	DACN-M	2650	86.45%	86.09%	86.27%	75.85%
Dongzhaigang Nature Reserve Region 2	SVM	1300	93.53%	19.64%	32.46%	19.37%
	Linear	80	72.15%	35.25%	47.36%	31.03%
	RF	2420	99.84%	18.99%	31.91%	18.98%
	KNN	1410	99.89%	18.99%	31.91%	18.99%
	DACN-M	810	76.68%	77.10%	76.89%	62.46%
Dongfang black-faced Spoonbill Nature Reserve	SVM	300	49.07%	15.43%	23.48%	13.3%
	Linear	90	68.52%	58.5%	63.11%	46.11%
	RF	50	62.96%	35.05%	45.03%	29.06%
	KNN	80	67.13%	33.72%	44.89%	28.94%
	DACN-M	1840	89.35%	90.61%	89.98%	81.78%
Xinying Bay Nature Reserve	SVM	400	68.54%	13.11%	22.01%	12.36%
	Linear	40	48.99%	12.87%	20.38%	11.35%
	RF	40	51.43%	20.09%	28.9%	16.89%
	KNN	90	67.38%	22.25%	33.45%	20.08%
	DACN-M	1000	88.42%	94.65%	91.43%	84.21%
Danzhou Xinying	SVM	300	56.52%	28.48%	37.88%	23.36%
	Linear	60	61.39%	26.66%	37.18%	22.83%
	RF	60	56.70%	22.64%	32.36%	19.3%
	KNN	100	69.04%	21.39%	32.66%	19.52%
	DACN-M	1000	97.57%	87.93%	92.50%	86.04%
Wenchang Qinglan Port	SVM	1400	91.53%	37.58%	53.28%	36.31%
	Linear	90	76.74%	43.62%	55.62%	38.52%
	RF	150	84.17%	42.34%	56.34%	39.22%
	KNN	190	88.89%	43.08%	58.04%	40.88%
	DACN-M	1000	95.44%	88.77%	91.99%	85.16%

:

Among these, the bolded data represents the optimal indicator data. As shown in Table 2, the Residual Threshold is the optimal result among the residual segmentation thresholds calculated by each model (determined by IoU). The values range from 0 to 5000. Based on the evaluation results of IoU and F1 Score, DACN-M performs excellently in various study areas, with IoU values greater than 60% and F1 Scores higher than 75% for crown prediction. Especially in study areas with simpler topography (lower proportions of other vegetation), such as Dongfang black-faced Spoonbill Reserve, Danzhou Xinying Bay, and Danzhou Xinying, the accuracy of the DACN-M model can exceed 80%.

In contrast, the RF, SVM, KNN, and Linear models have IoU values below 40% and F1 Scores below 60%. This indicates that these models have poor performance in practical segmentation tasks and cannot be used for accurate segmentation. Looking at the Precision and Recall values, RF, SVM, and KNN have extremely high Precision values in the Haikou Dongzhai Port and Haikou Yanfeng Town areas, but unusually low Recall values. This suggests that a large number of non-mangrove areas are being falsely detected as mangrove areas within these study areas. Linear and KNN show relatively stable results, with no significant difference between Precision and Recall, but overall, their performance remains unsatisfactory.

Further analysis was conducted by computing the residual heatmaps of RF, SVM, KNN, and Linear in different study areas.

Figure 13 illustrates the overlay of residual values as heatmaps for different models, with values ranging from 0 to 500. DACN-M has a larger optimal segmentation threshold, as depicted in Figure 13, showing the distribution of overlayed residual values between 0 and 2500. Smaller overlayed residual values indicate that the model considers the pixel to have a smaller difference from the mangrove crown’s characteristics, with values closer to red. Conversely, larger overlayed residual values indicate that the model deems the pixel less likely to belong to the mangrove crown. Combining this analysis with Table 2, the main factor contributing to the deterioration of model performance is the confusion between land vegetation crowns and mangrove crowns. RF, SVM, and KNN achieve differentiation between water bodies and buildings, with overlayed residual values exceeding the set upper limit (500). In the case of Linear, which is based on linear relationships, turbid water bodies interfere with model detection, while distinguishing land vegetation also poses a challenge. However, the model strategies of Linear and KNN have less dependence on the dataset, resulting in more stable regression results across different scenarios without a greedy impact on IoU values. SVM benefits from the concept of kernel functions and performs better than other models when dealing with land vegetation and mangrove vegetation. However, considering IoU, F1 Score, and residual maps, it still does not exhibit satisfactory detection performance.

Therefore, although the residual overlay strategy achieves good results in DACN-M, it is not easily transferable to all models. The calculated residual values rely on relatively accurate regression results, which are not stable in general models. Perhaps overlaying regression values of certain bands in other models could yield good results, but further experiments and discussions are needed to confirm this.

5.3. Hierarchical Analysis of Residual Superposition Results

The DACN-M employs a residual strategy for regression prediction, generating residual values alongside classification results. During field surveys in Yanfeng Town, we observed that these residual values could partially reflect the canopy density and geographical characteristics of the area—something traditional classification and regression models cannot achieve. This effect is sensitive to both artificially planted and naturally regenerated young mangroves, showing promise for detecting new forest growth. Figure 14 displays real scene images, remote sensing composite images, and residual overlay heatmaps for four observation points in Yanfeng Town.

In Figure 14a, the area represents an artificially planted mangrove zone, primarily consisting of mangrove seedlings. Due to soil and water quality, the growth conditions of the cultivated mangroves vary. On the right side of the river (b) and (c), the mangroves are often submerged, and field surveys revealed only sparse mangrove survival. This is not clearly reflected in the spectral images. On the left side of the river, the cultivated plants show better growth and denser planting, which can be observed in the spectral images. In the residual heatmap, the DACN-M shows a good response to the differences on both sides of the river. The lower residual values in the sparse areas on the right side indicate lower mangrove density and poor growth. Figure 14d–f show the dense distribution of mangrove canopies coexisting with the river. In the DACN-M residual results, there are significant high residual value sequences indicating the distribution of the river. Near Figure 14g–i, there is a frequently active human-made port, which somewhat contributes to the reduced canopy density of the mangroves. This result is also reflected in the residual heatmap. Figure 14j–l depict areas of newly grown mangroves. However, their growth shows significant variation. Figure 14k shows better growth due to lower terrain near the river, while some mangroves in Figure 14l exhibit poorer growth.

Overall, the residual values obtained by DACN-M show, to some extent, a correlation with mangrove growth. During our field survey in Yanfeng Town, we acquired some on-site landscape images for comparison. This suggests that the results of MREDT and DACN-M have the potential to be applied in the dynamic monitoring of mangrove growth.

5.4. Limitations and Challenges

Based on the MREDT and DACN-M models, we conducted monitoring and analysis using mangrove data from Hainan Island. The results indicate that this method significantly outperforms traditional methods in terms of classification accuracy. However, there are still some limitations in its application.

Firstly, although the DACN-M model excels in classification accuracy compared to traditional methods, its reliance on the overlay residual strategy requires substantial computational resources. As the number of bands increases, the training and inference processes become significantly more time-consuming [73]. This high computational demand may limit its broader application in practical scenarios.

Secondly, the data used in this study primarily come from the mangroves of Hainan Island. The applicability and generalization ability of the model to other regions still need further validation. Due to differences in ecological environments and species composition among mangroves in different regions, the model’s performance in other areas cannot be guaranteed to be consistent.

Thirdly, although MREDT reduces the need for manual labeling, the model may still require a certain amount of manual intervention to ensure result accuracy in some cases. When the surface terrain structure is complex or there are significant differences in surface information, false detections may occur.

Fourth, the feasibility of using the calculation results from MREDT and DACN-M in dynamic monitoring of mangroves needs further research in the future.

The MREDT and DACN-M models currently run on high-performance computing platforms, which may be inconvenient for users. In the future, their development on more user-friendly web platforms could be considered to facilitate easier access. Additionally, while the current models primarily utilize multispectral data, future research could explore the integration of multi-band and multi-scale data to further enhance model performance.

6. Conclusions

This study proposes a practical method for delineating mangroves using Sentinel-2 satellite imagery. The method introduces a mangrove rough extraction model, MREDT, and a deep learning regression model, DACN-M, for predicting mangrove crowns. Additionally, we found that different classification models have their own strengths and limitations when it comes to species classification in mangroves. Our method successfully depicts species in mangroves while limiting the influence of confounding factors. It outperforms segmentation frameworks like KNN and SVM. The experiments demonstrate that this method can be applied to satellite images collected at different times and locations, reducing the workload of field surveys and exhibiting broad potential. It is important to note that although the remote sensing techniques used in this study show good performance, there are still challenges and limitations. For example, the resolution and coverage of remote sensing images may be limited, which could affect the acquisition and extraction of mangrove information. Additionally, the cost and availability of remote sensing technologies can also be limiting factors. Therefore, in practical applications, it is necessary to choose appropriate remote sensing techniques and methods based on specific circumstances and conduct thorough accuracy assessments and validations.