4.3.2. Salt Marsh Classification

Figure 11 presented a closer look at the classification obtained in the salt marsh area. The three types of salt marsh distinguished—low, mid, and high marsh—appeared to be well identified.

The classification of the salt marsh channels was less correct: instead of wet sand and submerged sand, the classifier predicted rock and submerged rock in various areas.

#### 4.3.3. Seagrass Meadow Classification

In Figure 12, the focus was set on the seagrass meadow located in the north of the study site. Emerged and submerged parts of the rocky island were well mapped, even though submerged rock was detected in places where the seabed is sandy. The two types of underwater vegetation we attempted to map were very precisely defined: the patches of seagrass meadow and macroalgae were mapped with very low confusion with submerged sand or rock. However, the type of underwater vegetation (macroalgae or seagrass) was not correct in all places. There were classification errors in the seagrass meadow, where macroalgae was detected. Seagrasses were also found in macroalgae-covered areas.

Figure 12 also showed the precision of the classification of boats: boats mooring in the seagrass meadow were correctly labelled even though no training or test data were collected in that area for the boats class.

#### *4.4. Confusion Matrix Obtained with Green Waveform Features, Infrared Intensities and Elevations on the Test Dataset*

The confusion matrix obtained using green waveform features, IR intensities, and elevations on the test dataset is presented in Figure 13. It confirmed the observations made on the visual results. All classes were predicted with at least 70% of correctness. The most frequent confusions were between seagrasses and macroalgae, deciduous and evergreen trees, and submerged rock and submerged sand, which corroborated the observations made on the application of the model to the broader dataset.

The other confusion matrixes can be found in Appendix B (Figures A4–A6).

#### **5. Discussion**

We improved an approach initially designed to distinguish two seabed covers—fine sediment and seagrass—to map all land and sea covers present in our study scene in 3D. The findings showed lidar waveforms can be used to classify and map habitats of the coastal fringe and bridge the gap between marine and terrestrial surveys. All 21 selected classes were classified with at least 70% of accuracy in the best configuration obtained, which had an OA of 90%. Here, we discuss the results obtained regarding the classification predictors and the methodology employed. We also provide potential explanations for the performances of the algorithm.

#### *5.1. Green Waveform Features*

Our research partially aimed at exploring whether green lidar waveforms can be relevant for coastal habitat mapping. We defined 16 features to extract from the portions of the waveforms that correspond to layers of ground or seabed covers. These were efficiently retrieved both on land and underwater. However, our approach did not handle extremely shallow waters, where the surface component and the bottom return overlap in the waveforms. In these cases, the peak detection employed did not distinguish the seabed from the water surface and no features were retrieved. There was consequently a 24 m wide band without data in the surf zone on the sand beaches in our processed lidar dataset. We also noticed cases of confusion between seabed return and noise in the water column component of the waveform, which resulted in a mis-located detected seabed. These issues could be handled by improving the way the different waveform components are isolated: using waveform decomposition [31] or deconvolution [47,48] could produce better results on that aspect.

The features defined to describe the spectral signatures of coastal habitats seemed to be equally relevant for land and sea covers mapping. However, they did not provide a highly accurate classification (56% of OA). This can be explained by analyzing the green waveforms obtained with the HawkEye III on land. Since this sensor was particularly designed for bathymetry extraction, its green lasers are set to be powerful enough to reach the seabed up to several dozens of meters in coastal waters. Over land, the laser power is so high that most of the waveforms originating from highly reflective surfaces are saturated. The green wavelength alone might consequently not encompass a fine enough range of intensities over land to allow separation of similar environments such as plane habitats, different types of herbaceous vegetation, etc. The shapes of the saturated waveform returns are also affected: there is lacking information on the shape of the peak around its maximum. This can explain why there was a lot of confusion between topographic habitats when using green waveforms only.

Though green information alone may not be enough to distinguish the 21 habitats accurately, our findings suggested that a finer selection of the waveform attributes used for classification could enhance the green waveform feature predictions. The results presented in Table 3 and Figure 8 revealed negative interactions between some of the features chosen. Combining the full sets of statistical and peak shape features (defined in Section 3.7) resulted in lower accuracy than using them separately. Furthermore, the predictors' contribution assessment (Figure 8) showed that out of the 16 predictors, only nine contributed positively to the classification accuracy. This might be due to information redundancy between features relying on similar concepts such as mean and median intensity, for example. It could be due to the correction of attenuation performed on bathymetric waveforms. This exponential correction produced extremely high values of backscattered intensities under water, which made little sense physically. On the other hand, topographic waveforms were not corrected: their typical intensity order of magnitude was several times smaller, which might have disrupted the classifier. Fixing the issue of attenuation correction and using only a selected set of waveform predictors based on an assessment of their contribution would certainly result in better results when using the green wavelength alone.

The three different types of waveform features appeared to be complimentary: the nine predictors with a positive influence on the OA (Figure 8) represented each feature family defined in Section 3.7. This was consistent: the shape of the waveform return is characteristic of its nature, and the complexity, length, shape, maximum, and position of its maximum sum up the essential information differentiating one waveform from another.

#### *5.2. Infrared Data*

The addition of the IR wavelength increased the OA by 13%. The classification results certainly benefited from IR light's interaction with water and chlorophyll pigments, which provided essential information for the labelling of vegetation and other topographic classes such as wet sand. Considering that the green wavelength was less adapted to land cover classification, the performance increase obtained by using both lasers was expected.

Our research showed the added value of topobathymetric lidar: on top of providing quasi-continuity between land and water, both wavelengths provided complementary information for land covers' classification. The IR PC alone could not provide a coastal habitat map since it did not reach the seabed and riverbed; the OA obtained using only this wavelength (24%, presented in Table 3) confirmed that. They also showed that green lidar features alone do not provide a sufficient basis for classification either, reaching an OA of only 56%. Coupling both wavelengths improved the overall result significantly, bringing together the strengths of IR data on land covers, and the ability of green lidar to penetrate

the water surface. The matrixes presented in Figures A4 and A5 show that the addition of IR intensities to green waveform features resulted in an accuracy increase for all but two of the 21 classes. The gain ranged between 0% (submerged sand) and 39% (tar). The minimum accuracy observed over the 21 labels rose from 21% to 29%. Water covers classes such as seagrass and submerged rock also benefited from the addition of IR intensities, as their accuracy showed an improvement of 1% and 6%, respectively. As HawkEye III is tailored for bathymetry, its green laser's power was set to be high, which resulted in saturated intensities over land. The IR channel provided complimentary information in places when the green channel was weaker, which partially explains the algorithm's performances we observe. The classification accuracies obtained for land covers confirm that; for example, the classification of soil, wet sand, and lawn was significatively enhanced: +28%, +34%, and +23% of OA, respectively. Our future work will focus on exploiting both IR and green waveforms for habitat classification, to maximize the accuracy attainable with full-waveform topobathymetric lidar.
