*3.2. Morphometric Derivatives*

Four main morphometric by–products were calculated:


(**a**) (**b**)

**Figure 9.** Morphometric variables derived form the best Pleiades–1 DSM#2–3\_1: slope in degrees (**a**); aspect (**b**); topographic position index, (TPI) (**c**); TPI–based landform classification (**d**).

#### *3.3. Pixel–Based Classification*

3.3.1. Overall Accuracy at the Landscape Scale

At the landscape scale, each band combination was assessed. To test individual contribution, the RGB spectral composite was used as a basis, achieving an OA of 76.37% (Figure 10).

**Figure 10.** Overall accuracy of the eight classifications performed with the maximum likelihood classifier.

The NIR band was also tested and obtained a score of 84.34% combined with the RGB. Thus, in addition to the RGB, the DSM significantly enhanced the OA by +12.51%, but RGB + DSM + slope decreased the score (−11.23%) compared to the RGB + DSM combination (Figure 10).

The morphometric predictor combinations provided a strong contribution to the classification, with +9.05%, +11.45%, and +12.52% for the RGB + DSM + TPI and RGB + DSM + morphometric predictors and for RGB + DSM + aspect, respectively.

Finally, the best combination was the TPILC predictor combined with RGB + DSM, with a classification score of 89.37%, namely an augmentation of +13%.

#### 3.3.2. Evaluation at the Class Level

The analysis of the confusion matrix of each morphometric predictor added to the basic RGB highlights a heterogeneity between the classes (Table 3, Figures 11 and 12).

**Table 3.** Producer accuracy (in %) from confusion matrix using the maximum likelihood classifier for the salt marsh, dune, rock, urban, field, forest, beach, road, and seawater classes.


**Figure 11.** Bar plot of the producer accuracy of the morphometric predictor on the basis RGB at the class level (salt marsh, dune, rock, field, forest, beach, road, and seawater).

**Figure 12.** *Cont*.

(**c**) (**d**)

**Figure 12.** Coastal mapping classification at the class level computed with the maximum likelihood classifier: Basis RGB (**a**); RGB + NIR (**b**); RGB + DSM (**c**); RGB + DSM + slope (**d**); RGB + DSM + aspect (**e**); RGB + DSM + TPI (**f**); RGB + DSM + TPILC (**g**); RGB + DSM + morphometric predictors (**h**).

Thus, the urban class obtained the worst results with the RGB and increased in terms of classification performance when another predictor was added. The slope predictor increased by 2.46. The addition of a morphometric predictor allowed the 70.13% of classification performance to increase until the threshold of 84.13% as reached with the combination RGB + DSM + morphometric predictor. The NIR predictor holds up reasonably, achieving a score of with 71% (Table 3, Figures 11 and 12).

The trend for the forest class seems to be the same as the trend observed for the urban class. The slope provides a modest contribution of 2.47%, followed by the NIR predictor with a contribution of 16.67%, and then by the DSM contribution with a contribution of 32.27%. When added to the RGB + DSM, morphometric predictors aspect, TPILC, TPI, and the entire combination obtain values of +30.93%, +30.97%, +37.6%, and +40.2%, respectively.

The salt marsh, beach, and dune classes increased in classification accuracy when a morphometric variable was added to the reference RGB: +10.73%, +17.4%, and +36%, respectively with the RGB + DSM combination until the addition of TPI variable at 98.6% and 99.07% for the salt marsh and dune classes. The beach class obtained the best result

with RGB + NIR, achieving results of 99.4%. In contrast to the salt marsh and beach classes, the rock class performed worst, with 13.27% and 26.33% with the RGB + TPI combinations and the complete combination of morphometric variables added to the RGB, respectively.

The road class obtained linear classification scores between 91.47% for RGB + DSM + TPI and 94.2% for the RGB + DSM + morphometric predictors.

The field class follows a slightly different pattern than the other classes since the worst result, which was still very acceptable, is achieved by the combination RGB + DSM + slope, with a score of 86.47%. The RGB combination increased the classification performance by only +0.2% and +0.93% for the RGB + DSM + aspect and +1.13% for the RGB + DSM combination. The RGB + DSM + TPILC, RGB + NIR, and RGB + DSM + morphometric predictors increased the classification performance by +2.2%, 2.8%, and +5.46%. The best combination for the field class was obtained by RGB + DSM + TPI, with a score of 93.6%.

The seawater class achieved high scores of almost 100% PA, regardless of the predictor (Table 3, Figures 11 and 12).

#### **4. Discussion**

#### *4.1. Pleiades–1 Digital Surface Model*

4.1.1. The Intersection Angle as a Key Determinant

At the global scale, nine DSM were computed via RSP from three Pleiades–1 satellite images captured at three different angles of incidence: 16.41◦ for image #1, 15.35◦ for image #2, and 16.05◦ for image #3 (Figure 2a and Table 1). DSM#2–3\_1, which was derived from the stereo reconstruction of images #2 and #3 outperformed, the other DSMs. According to the results of the satellite photogrammetry, reconstructions with the closest intersection angles (5.13◦) produced better point–measurement accuracy compared to the the 2018– LiDAR altimetric reference. Another answer can also be provided by focusing on the solar angle [21]. Moreover, close or near–similar intersection angles increase the risk of "hidden sides" because of their proximity [22,23].

However, many studies have shown interest in using tri–stereo satellite images to benefit, when possible, from a nadir view [24]. The benefit of such a tri–stereo enables a reduction in the shadows that are created by trees or buildings. This approach is highly valued by urban planners, as it limits the risk of shadows and hidden areas [25].

As for photogrammetric reconstructions from UAV images, a deficient RMSE could be explained by reconstruction artifacts related to the algorithm or to the images themselves [26].
