1. Introduction
Small island developing states (SIDS), including those in the Pacific, are disproportionally threatened by climate change [
1]. Pacific Island countries and territories, including the Federated States of Micronesia (FSM) and the Solomon Islands, depicted in
Figure 1, are often described as being on the “frontline”, with increases in global average temperature predicted to not only increase the intensity and frequency of disaster events [
2] but to lead to sea level rise. This, in turn, will exacerbate coastal erosion [
3], impacting valuable ecosystems, including wetlands. Recent developments in both Earth observation methods, including the satellite and geodetic data obtained via remote sensing (RS), and artificial intelligence (AI) have increased opportunities for not only rapid, accurate assessment of sea level, coastlines and vegetation but also for researchers to make predictions about how they will change, allowing communities and policymakers to better respond to climate threats.
Globally, sea level rise (SLR) is caused by a combination of the thermal expansion of ocean waters and freshwater input from melting glaciers and ice sheets, increasing the total volume of water in the sea [
4]. At a regional level, ocean currents, wind and pressure changes can contribute to SLR [
5]. While international climate negotiations often focus on air temperature thresholds [
6], even a 1.5 °C target is predicted to result in significant SLR.
Between 1901 and 1990, the global mean SLR was 1.35 mm/year, faster than any century in at least 3000 years [
7]; between 1993 and 2018, this rate accelerated to 3.25 mm/year, and it is projected to rise further. Yet, SLR varies greatly based on geography. Church et al. [
8] highlight the variability in sea levels between 1993 and 2001, with large rises across the western Pacific and eastern Indian Oceans and drops in the eastern Pacific and western Indian Oceans. In the western Pacific nations of Micronesia and the Solomon Islands, Klein [
9] reports that SLR of up to 12 mm/year, more than triple the global average, has occurred since the early 1990s. As a result, small island nations are predicted to suffer wetland degradation, increased flooding and saltwater intrusion as a result of SLR [
10]. Mangrove wetlands are known as both a “bioshield” protecting coastlines and reefs and one of the Pacific’s most vital “blue carbon” sinks.
While some research has been undertaken in both the FSM and Solomon Islands, there has been relatively little scholarly attention paid to these islands in comparison with areas of Melanesia, such as Fiji [
11]. The need for additional study is evident given the unclear relationships between complex climate variables and the conflicting accounts and predictions of ecosystem loss. Projections for the Solomon Islands’ mangroves, for example, range from no change [
12] to a loss of 68% [
10]. Remote sensing technologies, including satellite and Tide Gauge (TG) observations, with the application of new AI approaches to predict sea levels in these areas, provide an important opportunity to accurately assess and predict the interlinked challenges facing not only Pacific Island nations but the entire world. In response to these needs, this study employs new data-driven hybrid AI model(s) to provide predictions of the sea level in the FSM and the Solomon Islands using geodetic data. Hence, a new hybrid CNN-BiLSTM deep learning model is developed with a Multivariate Variational Mode Decomposition (MVMD) technique for the prediction of sea levels. Furthermore, it will utilise Landsat satellite imagery to detect coastal wetland changes in the FSM and the Solomon Islands.
Convolutional Neural Networks (CNNs) have been successfully combined with TG data to predict sea levels in other South Pacific islands [
13], as well as worldwide sea level surges [
14]. Bidirectional Long Short-Term Memory (BiLSTM) deep learning methods have also shown high levels of accuracy in predicting the sea level in Kiribati and Tuvalu [
15] and wave height in Australia [
16], with the BiLSTM model’s wave predictions outperforming all other models, including BiLSTM, EEMD-SVR and SVR alone. Hybrid CNN-LSTM methods have been employed to predict phenomena as diverse as flooding [
17] and air quality [
18]. As Sharma et al. note, CNNs employ efficient multistage architecture via convolution, yet additional improvements can be achieved using a secondary LSTM-based deep learning architecture, the LSTM helping to vanish gradient issues and explore sequential data using unique gates. As a result, the amalgamation of CNNs with LSTM-based architecture is an active area of research.
2. Materials and Methods
The PSLGM includes a network of geodetic monitoring stations implemented and maintained by Geoscience Australia, providing Global Navigation Satellite System (GNSS) measurements, which permit the absolute determination of the vertical height of the gauges measuring the sea level. In addition to sea level data (meters above TG zero), the PSLGM also provides measurements of climate variables, including water and air temperatures (in degrees Celsius), barometric pressure (in hPa), residual and adjusted residual sea levels (in meters), wind direction (in Degrees True), as well as wind gust and speed (in m/s).
2.1. Data Preprocessing, Partitioning and Normalisation
Missing or erroneous values in the PSLGM dataset are set to a value of −9999, which must either be removed or replaced via interpretation before analysis can take place. Any month with more than 1000 missing values (i.e., >15% of the total data points) was excluded from the analysis. Then, the remaining data were checked to ensure no more than a maximum of 72 missing consecutive hourly values (i.e., three days) in any given column. Finally, interpolation was conducted using the popular Pandas Linear method in Python.
Lags are an important aspect of time-series data modelling [
19]. To determine significant lags, the Auto-Correlation Function (ACF) and Partial Auto-Correlation Function (PACF) were computed, which take into account seasonal and cyclic trends and residuals to find correlations [
20].
Figure 2 shows the results of this analysis for the FSM.
Signal decomposition is the extraction and separation of signal components into their intrinsic mode functions (IMFs) [
21]. Multivariate Variational Mode Decomposition (MVMD) was used, which has the ability to simultaneously capture the non-stationary and non-linearity of a multichannel signal to overcome mode-mixing issues [
22]. This procedure helps extract hidden features from the sea level time-series signal and facilitate effective AI model learning from these input variables for accuracy in forecasting.
Figure 3 shows the sea level signal decomposition using MVMD for the FSM.
The corrected oceanic dataset containing the input and target variables was then partitioned into training, validation and testing, as shown in
Table 2.
Periods for which data were missing are omitted from the partitions, as described above. Note that the TG record for the Solomon Islands location is longer (starting in 1994) than for the FSM facility, as the TG at this location was only installed in 2001.
Figure 4 shows the correlation of the Solomons Islands’ sea level with its predictor inputs. Similar analysis was undertaken for the FSM.
The final step of data preprocessing is normalisation using the equation:
This scaling process helps reduce the time taken in the learning stage of data modelling by avoiding the computation of large values [
23]. Once modelling is complete, the values are converted back into their original form using the equation below:
where
is the input data value,
is the overall minimum and
is the overall maximum value.
2.2. Sea Level Prediction
Machine Learning (ML), a subset of AI in which computer systems learn automatically with experience rather than being explicitly programmed, is commonly used in the analysis of ocean data, where traditional methods have many shortcomings [
24]. Deep learning (DL) may be defined as a subset of ML and refers to techniques which layer algorithms and computing units or “neurons” into artificial networks designed to mimic the brain. Researchers have predominantly used Artificial Neural Networks (ANNs), followed by Support Vector Machines (SVMs), to predict flooding and the use of mangroves for risk mitigation [
25], particularly in conjunction with satellite data [
26].
In the present study, four models were tested.
2.2.1. Multilinear Regression
Multiple linear regression or multilinear regression (MLR) can estimate the relationship between two or more explanatory variables and one response variable. An MLR model is a supervised learning algorithm which can be used, for example, to predict sea level given multiple input variables.
The formula for MLR is:
where
is the predicted or expected value of the response or dependent variable,
is the
y-intercept (i.e., the value of
y when all other parameters are set to 0),
through
are the regression coefficients of the explanatory or independent variables
through
(i.e., the effect that increasing the value of the independent variable has on the predicted
y value) and
is the model error (i.e., how much variation exists in the estimate of
y).
2.2.2. Random Forest Model
A Random Forest (RF) model utilises a combination of tree predictors, where each tree depends on the values of a random vector sampled independently with the same distribution [
27]. The algorithm applies the bootstrapping aggregation to tree-based learners [
28]. These bootstrap samples of the training sets are selected repeatedly, and Gini impurity fits
trees in these samples. Then, the equation below is used to calculate the predicted values for unseen complexes:
where
is the number of times the bootstrapping aggregation or “bagging” is performed, and
is the input variable.
2.2.3. Multi-Layer Perceptron
A Multi-Layer Perceptron (MLP) is a supervised ML algorithm [
29], one of the simplest ANNs, in which the node’s connections do not form a loop, i.e., the flow of information is unidirectional [
30]. MLPs are some of the most widely used ML algorithms, capable of robust and efficient flood prediction [
25]. While a single-layer perceptron consists of a single-layer output node directly connected to the input by a series of weights, a multi-layer perceptron is an interconnected network with multiple hidden layers [
31]. In the hidden layers, the input data undergo a series of weighted sums, and after calculating the weighted summation of each hidden neuron, the result is applied to an activation function,
f, and the result of this function is again weighted and summed to obtain the output [
24]:
where
is the input vector from the previous layer, and
is the weight vector, generating the scalar product
[
32].
2.2.4. Bidirectional Long Short-Term Memory
A Bidirectional Long Short-Term Memory (BiLSTM) model architecture consists of two long short-term memory (LSTM) networks. Long short-term memory (LSTM) neural networks are another type of ANN [
30], which selectively memorise input [
24].
LSTM networks comprise three layers: input, one or more hidden layer(s) and an output layer (similar to an MLP), with the neuron number in the input/output layer equivalent to the amount of feature space [
33]. The memory cell(s) within the hidden layer have three gates: forget, input and output, and at every time step
t, each gate is presented with the input
and the output of the memory cells at the previous time step,
.
At each time step, the cell state
and output
are calculated. The gates act as filters, with the forget gate deciding which particular is detached from the cell state, the input gate specifying which information supplements the cell state and the output gate deciding which data from the cell state are used as output [
33]. Finally, the sigmoid function scales all values from 0 (forget completely) to 1 (remember completely):
The second step is determined by the LSTM layer, which adds information to the network’s cell states, by computing candidate values for
and activation values
The third step involves the design of new cell states
based on the results of the previous steps, with ◦ representing the Hadamard product:
The final step is the calculation of the output,
, using the following two equations:
LSTM neural networks have a strong learning and predictive ability for time-series data such as sea surface temperature [
24] and saltwater intrusion [
34].
2.3. Model Evaluation
Following Raj [
15], five statistical metrics were used to evaluate the performance of the models described above.
The first three equations are efficiency metrics, used to measure the accuracy of the models. The correlation coefficient r determines the relationship between two variables, indicating the strength of association, e.g., between the observed and predicted SLR. Willmott’s Index of Agreement d indicates the ratio of the mean of square error and potential error, detecting proportional differences between the observed and predicted values. Legates and McCabe’s Index LM is a more advanced index utilising the adjustment of comparisons in the evaluation of Willmott’s Index.
- 2.
Willmott’s Index of Agreement (d)
- 3.
Legates and McCabe’s Index (LM)
The error metrics used for evaluation are the root mean square error (
RMSE) and mean absolute error (
MAE). The
RMSE is the square root of the mean square error and measures the average difference in error between the predicted and observed values [
35]. The
MAE is the mean of the absolute errors between the predicted and observed values [
36].
- 4.
Root mean square error (RMSE)
- 5.
Mean absolute error (MAE)
2.4. Wetland and Mangrove Detection
A Random Forest model using satellite data was also used for the detection of wetlands, with the Landsat data for supervised classification obtained from Surface Reflectance Tier 1 of the USGS Landsat 7 and 8 image classification dataset, at a 30 m digital elevation and with images selected based on the lowest prevalence of cloud cover.
Spectral indices were computed for the wetland mapping, including the Normalised Difference Mangrove Index (NDMI) [
37], Modified Normalised Water Index (MNWI) [
38], Simple Ratio Vegetation Index [
39], Green Chlorophyll Vegetation Index (GCVI) [
40] and normalised difference vegetation index (NDVI).
The NDVI is a simple graphical indicator of whether an area under observation contains live green vegetation [
41]. It is often used as a proxy for vegetation productivity, and therefore overall health [
42] and growth [
43], and is calculated as follows:
where NIR represents near-infrared and RED represents red wavelengths [
44]. These bands contrast the absorption of chlorophyll pigment at the red end against the reflectance of mesophyll at the NIR end [
41]. Healthy vegetation tends to absorb most of the light at the red end of the spectrum and reflect a large portion of the NIR light, while unhealthy or sparse vegetation will reflect more red light and less NIR light [
45].
NDVI values can range between −1 and +1, with positive values representing vegetation of varying health and negative values representing other land use/land cover (LULC) classes [
46]. A total of 330 samples were taken for the Solomon Islands and 340 for the FSM. These samples were then split into training (80%) and testing (20%) sets. The model used 200 trees and 5 randomly selected predictors per split. Each stratified point was checked with ground-based images, as shown in
Figure 5, to evaluate the label for correct classification of the coastal wetland.
Following evaluation, the class accuracy plugin in QGIS was used to compute the accuracy and the Kappa value for both study areas, using Equation (11) below.
Figure 6 shows a close-up view of the stratified points:
5. Conclusions
This study utilised data-driven AI models to predict the sea level using a range of climate input variables. It successfully used TG data to predict the sea level in the Solomon Islands and FSM from 1994 to 2022. The geodetic variables of water, air temperature, barometric pressure, wind direction, gust and speed were used to train the AI models. The hybrid CNN-BiLSTM model provided the most accurate sea level prediction for both study sites. The mean sea level was used to provide an annual trend analysis, which revealed a projected increase for both locations above the global trends. The mean sea levels were found to have risen 6.0 mm/year over the last two decades in the Solomon Islands (a slight decrease compared to the 6.3 ± 2.1 mm/year rate of rise since 1994) and 7.2 ± 2.2 mm/year in the FSM. The SLR in both locations appears significantly higher than the global averages.
The identification of wetlands in general in small island nations such as the FSM and Solomon Islands can be difficult, with challenges including the availability of cloud-free scenes and the detection of small, yet ecologically significant, patches of vegetation. The detection of coastal vegetation changes was carried out using data from Landsat satellite images and an RF model to examine the coastal wetlands on Guadalcanal and Pohnpei and data from GMW to examine mangrove trends in the Solomon Islands and FSM more broadly. A high level of accuracy (>0.98) was achieved using the RF model at both study sites. The coastal wetlands in general were found to have decreased in extent. Analysis of vegetation health is of particular importance in the Solomon Islands given the detrimental effects of logging observed by Minter and van der Ploeg [
49]. The mean NDVI in the Pohnpei region in the FSM (ranging between 0.4 and 0.8) was found to be higher than in the Guadalcanal region in the Solomon Islands (ranging between 0.3 and 0.7), consistent with the other results of the present study. Apparent losses in wetlands were recorded in both study areas using the Landsat satellite imagery analysis, while such decreases may be associated with logging and other climate change effects, including a rising sea level.
The findings of the present study demonstrate the importance of long-term monitoring and the importance of taking the length of records into account. Shorter records are more susceptible to extremes and may mask the true effects of SLR. As PSGLM records continue to lengthen, future research will be able to make more accurate observations (and hence, predictions) of the MSL. While the wetland detection using satellite imagery and RF classification utilised in the present study achieved high levels of accuracy, comparable with or exceeding previous studies, more research is required in this area, including fieldwork, which can provide ground truth data.