Together with mangrove and salt marsh, seagrass has been evaluated as an effective coastal ecosystem for blue carbon storage [1
]. However, ongoing degradation of seagrass meadows [4
] is leading to a requirement for accurate mapping and monitoring methods to facilitate the MRV (Monitoring, Reporting, and Verification) approach necessary for broad scale evaluation of their contribution to blue carbon reservoirs [5
]. In the last decade, satellite imagery has been used extensively in developing seagrass mapping techniques by employing various classification algorithms with or without parallel traditional field surveys [6
]. Among them, Sentinel-2 imagery is becoming more popular for seagrass mapping. Operated by the European Space Agency since 2015, this sensor supports a high quality image at spatial resolutions between 10 and 60 m [7
]. Sentinel-2 data have been distributed free-of-charge at the top-of-atmosphere corrected level (level 1C) for blue, green, red, and near infrared (NIR) bands at 10 m resolution, and provides a very good resource for intertidal and subtidal ecosystem mapping. Using these data to derive ecosystem spatial properties requires classification algorithms and overfitting, and the inaccurate edge detection of different substrata remains a limitation of traditional classification methods [6
]. For seagrass mapping, the problems of misclassification usually relate to the impact of deep water on pixel values or the mixture of substrata within a seagrass meadow [11
]. To overcome this problem, very high resolution (VHR) imagery and a variety of classification approaches can be considered [12
]. Most frequently, probability-theory based models such as the maximum likelihood classifier (MLC) have been applied for seagrass classification [6
]. This approach, however, requires conditions that are difficult to satisfy in the marine environment including a normal distribution of probabilities, equal co-variance, and large amounts of validation input data [14
]. In addition, the utilization of the linear or quadratic discrimination functions of a MLC may not work when the boundaries of classes are not well defined [15
In recent years, machine learning (ML) has emerged as a novel approach for seagrass mapping and monitoring [6
]. Machine learning has the benefits of rapid learning, accommodation of non-linearity [16
], and the availability of an increasing number of new, open source algorithms [17
]. In the field of seagrass mapping and monitoring, however, the application of machine learning is still in its infancy [6
]. Examples used to date include weighted majority voting using Quickbird images [18
]; logistic model trees (LMT), AdaBoost, random forest (RF), and artificial neural networks (ANN) using digital images [19
]; support vector machine (SVM) using Sentinel-2 images [13
]; and decision trees (DTs) using aerial photographs [21
]. In these examples, when used with high spatial resolution images (<1 m), machine learning models achieved an accuracy of 92–100%. Decision tree models using aerial photographs, however, achieved a lower accuracy of 66% for seagrass meadows when the plant cover was below 60% [21
]. These mixed results support the exploration of novel machine learning approaches, particularly for improving low coverage seagrass mapping.
Among the various DT ensemble machine learning algorithms, rotation forest (RoF) and canonical correlation forest (CCF) algorithms are now emerging as reliable techniques for land cover mapping [22
], landslide mapping using multi-spectral [23
] or hyper-spectral [24
] imagery, and rapid building mapping using multi-source data [25
]. Using bootstrap sampling, combining multiple independent base classifiers, and applying statistical analysis (principal component analysis in the RoF model) [26
], these learning algorithms are well-known for reducing the variance and overfitting of the classification results, resulting in a better detection of multi-class boundaries [28
]. In addition, the CCF model does not require the optimization of hyperparameters [31
], which makes this model simpler to apply for mapping tasks. To our knowledge, these techniques have not been used for seagrass mapping, however, they potentially offer benefits in the classification of low coverage through enhanced recognition of edge boundaries. Therefore, our goal in this study was to compare the use of three ML algorithms, RF, RoF, and CCF, to the more traditional MLC approach for mapping the aboveground distribution of seagrass communities at low and high coverage using Sentinel-2 data.
Our target was Tauranga Harbor, New Zealand, for which ground truth data were available, and which offers a mosaic of dense, sparse, and zero seagrass coverage. We discuss here the difference in performance of the selected models for seagrass detection at two densities. Our results are expected to contribute alternative solutions for the mapping and monitoring of seagrass at various regions in the world, and assist in the conservation of this important blue carbon ecosystem.
As far as we are aware, this research is the first attempt to compare the performance of the RF, RoF, CCF, and MLC methods for seagrass mapping with a full radiometric correction of the image. Desirable characteristics for seagrass mapping are both high precision and recall. High precision means that the classifier is able to detect the seagrass pixels precisely, whilst high recall means that the classifier is able to find all possible pixels of seagrass. To give a final coherence score and harmonize the values of precision and recall, the F1 score is usually preferred to evaluate a model’s performance. The research presented here suggests that ML models detect dense seagrass meadows well, and outperform the traditional MLC approach.
Of the machine learning ensemble approaches used, the CCF and RF models performed less well than the RoF model, contradicting a superior performance of CCF in other studies [31
]. CCF produced lower recall whilst RF created a lower precision for both the dense and sparse seagrass classes. For sparse seagrass meadows, CCF detected more precisely than the RF model. In addition, MLC produced very high recall, but low precision scores for both seagrass classes, leading to a lower F1 score and accuracy than the ensemble based models. Overall, our results show that RF, RoF, and CCF are good performers with a balance of high precision and recall scores, whilst very low precision scores of dense and sparse seagrass classes ranging from 0.34–0.63 were found for MLC. These results confirm the robustness and consistency of machine learning ensemble based methods in comparison to the MLC. We hypothesize that the poor performance of MLC may be because of the need for input data to satisfy the built-in assumptions, described above, which are difficult to sustain in a spatially heterogeneous marine environment.
Of the methods tested here, only the RF technique has previously been applied to seagrass mapping using very high spatial resolution imagery. In that case, high precision (0.947) and recall (0.968) values were determined mapping Posidonia oceanica
from digital airborne images, though no comparison to other methods was attempted [19
]. In another seagrass study, the overall accuracy only reached 82% using the RF algorithm applied to RapidEye imagery [67
]. Considering the size of the seagrass meadows and the mix of substrate in Tauranga Harbor, the measured scores in our results were reliable for both dense and sparse seagrass mapping using medium spatial resolution of Sentinel-2 data (10 m pixel size). In other studies that allowed for a comparison of RF, RoF, and CCF models, CCF slightly outperformed the RF and RoF models for land cover mapping [22
], RoF outperformed RF and CCF for mangrove mapping [68
] whilst a similar performance of RoF and CCF models was noted for landslide mapping [23
In addition to the performance advantages, RF and RoF are easy to execute in Python™, whilst CCF is confined to the MATLAB environment. The open source operating environment and diverse Python™ libraries provide multiple solutions for seagrass mapping and monitoring, and enhance the capacity to develop novel algorithms for various tasks in marine science [69
]. Several libraries in the Python™ environment support a built-in framework for classification problems with a long list of state-of-the-art machine learning algorithms [70
]. This ease-of-use approach allows a person with minimum programming skills to make a classification more reliable with machine learning. Recently, cloud computing broadens the executable environments for the big Earth observation data, especially in coastal resource mapping. A cloud computation system is able to deal with massive amounts of remote sensing datasets, parallel processing of satellite image using multiple data centers, and can respond to real time monitoring on country and global scales [71
]. The use of open source machine learning algorithms in PythonTM
and a cloud system, therefore, promises large scale and more reliable mapping in the future.
Our results have validated the performance of the RF, RoF, and CCF models for seagrass mapping and suggest that the RoF technique is a promising novel approach to further seagrass monitoring at various sites around the world. However, the current study does have some limitations. The mismatch between the number of ground truth points (GTPs) and Sentinel-2 pixel size (10 × 10 m) may raise a degree of uncertainty in classification, particularly for sparse seagrass. However, we considered that despite this mismatch, a sufficient and representative number of GTPs for each class was collected during the field survey (as presented in Section 2.2
) to have confidence in the classification. Related issues that might explain the low values of precision and recall for sparse seagrass meadows are issues of mixed pixels, whereby small seagrass patches or dispersed clumps within a pixel challenge the classification process. To our knowledge, these effects are not easy to compensate for in the case of low to very low coverage seagrass using Sentinel-2 imagery. Thus, the use of very high spatial resolution sensors such as WorldView (~0.3–0.4 meters) [57
] or Pleiades-1 (0.5 meters) is currently being investigated for future studies for seagrass mapping. Moreover, with the development of computer vision and pattern recognition, deep learning approaches using a variety of algorithms such as convolutional neural networks (CNNs or recurrent neural networks (RNNs) for semantic segmented imagery applied sub-pixel techniques should be encouraged for future studies [6
We tested the performance of ML ensemble-based and MLC methods for seagrass mapping from Sentinel-2 data. Using Tauranga Harbor as a validation site, our comparison indicated that all ML-based approaches significantly outperformed MLC. MLC failed to detect sparse seagrass meadows, with a low F1 score of 0.50. We noted a better performance of RoF compared to the RF and CCF models with the highest F1 scores of 0.91 and 0.75 for dense and sparse seagrass classes, respectively.
Our results attest to the reliable application of the RoF model for the mapping and monitoring of seagrass in shallow water using Sentinel-2 imagery. Despite a lower accuracy for sparse than dense seagrass meadow classification, the CCF model shows potential for the mapping of seagrass and merits further testing at various scales and in various case studies. Regarding MLC, this model is still an applicable candidate for dense seagrass meadows, however, it may not be applicable for the mapping of sparse to very sparse seagrass meadows.