Reconstructing Late Quaternary coastal landscapes by a machine-learning framework Authors: Gaia Mattei, Alessia Sorrentino, Gerardo Pappone, Angelo Ciaramella, Pietro P C Aucelli Published: 2025 ABSTRACT AbstractCoastal landforms, particularly sea cliffs and associated wave-cut platforms, preserve key evidence of past sea-level fluctuations, tectonic activity, and paleoclimate variability. In this study, we implement a supervised machine learning approach, trained on an original, expert-labeled geomorphological dataset, to detect and classify inherited and active coastal features - such as paleo-sea cliffs and polycyclic sea cliffs - along the south-Tyrrhenian. Using DTM and morphometric indicators, our model, based on a RandomForestClassifier trained on expert-based cartography and independently validated, accurately identifies the spatial signatures of Quaternary coastal evolution. These results are cross validated against independent geomorphological mapping and sea-level reconstruction datasets. The integration of geomorphological classification with sea level markers enables us to reconstruct coastal morphogenesis in relation to the last interglacial cycle. Our findings highlight the potential of machine learning to automate the identification of coastal paleo-landscapes, providing insight into the imprint of climatic forcing on their morphology. This approach offers a scalable framework for investigating past climate–landscape interactions and for supporting future coastal hazard assessments under changing climate conditions.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-31006-0.Keywords: Coastal landforms, Coastal change, Sea level change, Coastal paleo-landscape, Automated mapping, Mediterranean areaSubject terms: Climate sciences, Ecology, Ecology, Ocean sciences, Solid Earth sciences Introduction Coastal paleo-landscape reconstructions are crucial for understanding long-term morpho-evolution driven by millennial climatic changes1–3. In rocky coastal settings, assessing morphological responses to climate-driven forcings improves predictions of future submersion scenarios, aiding coastal protection and relocation planning4–7. Such reconstructions traditionally rely on extensive field surveys and expert mapping of landforms interpreted as relict landscapes6,8,9. Although indispensable, manual mapping is time-consuming and limits large-scale application10. Recent advances in geostatistics and Artificial Intelligence (AI) enable rapid, accurate identification of target coastal features11–14; This progress is facilitated by increasing access to high-resolution datasets from lidar and satellite imagery and supported by open data policies and enhanced computational power, which have expanded machine learning (ML) applications in geoscience15–18. In this context, ML-based paleo-landscape reconstructions deepen understanding of sea-level, climate, and tectonic influences over millennial timescales and provide robust baselines to model contemporary landscape responses to environmental change19–21. Here, we developed a ML model to reconstruct the Last Interglacial (LIG, between 130 and 115 ka BP) coastal landscape along the Tyrrhenian coast (with a RSL oscillating between 8 ± 1 m and +2.5 ± 0.5 m MSL) by integrating bibliographic analysis and extensive field surveys. We focused on assembling a robust training dataset, composed of 1,450 labeled records, including terrain derivatives, classified through supervised analysis as well as through fieldwork and literature (see section “On-site surveys and paleoenvironmental markers”) to obtain the training dataset for the ML model. Our goal is to combine ML with traditional mapping to create a predictive tool for reconstructing tectonically quasi-stable22,23 coastal sectors of the Cilento promontory, where LIG terraces and Holocene deposits indicate absence of appreciable vertical movements, and their climate-driven evolution during the LIG, aiming to forecast potential impacts of ongoing climate change. Study area The study area, encompassing approximately 102 km in length and 11 km in width (Figure 1), covers over 1000 km2 and corresponds to the entire coastal sector of the Cilento promontory, where the Last Interglacial (LIG) landscape was reconstructed. Geological sketch of the study area. Geological information is based on the CARG geological map of Italy, Foglio 520 “Sapri” (Monti et al., 2014). Topography from TINITALY DTM (https://tinitaly.pi.ingv.it/) and bathymetry from EMODnet (https://emodnet.ec.europa.eu/). The map was created in QGIS 3.38.3 (QGIS.org, https://qgis.org); final layout edited in Inkscape 1.3.2 (https://inkscape.org/release). Situated along the South-Tyrrhenian margin, the Cilento promontory forms the morpho-structural terminus of the southern Apennine thrust system, shaped by compressive tectonics from the Lower Miocene to the Lower Pleistocene24–28. Morphological indicators linked to Upper Pleistocene sea-level stands (LIGe and LIGc) suggest tectonic stability over the past 125 ka22,24,29,30. This is further supported by the elevation of Holocene beach deposits (LITHO-6)31 infilling fluvial incisions up to 2 km inland32. In this study, lithological units were derived from the ISPRA 1:50,000 geological mapping and reclassified into ten lithological classes (LITHO-1–10) to facilitate geomorphometric analyses and their use in the machine learning model. A general description of these classes is provided in Supplementary Materials 3, while the corresponding geological unit names are illustrated in Figure 1 and its caption. The NW coast is dominated by marine turbidite sequences belonging to allochthonous tectonic units (Cilento group; Figure 1, LITHO-7), overlain by Quaternary continental debris and marine-transitional deposits. The active sea cliffs of Punta Licosa expose siliciclastic arenaceous-pelitic turbidites of the “Pollica sandstone” dated between 5.97 and 13.82 Ma26. Southward, Oligocene pelitic-arenaceous turbidites of the Northern Calabrian Unit (Figure 1, LITHO-1) are well exposed between Acciaroli and Pisciotta. Underlying this unit, the tectonized Castelnuovo Cilento succession comprises varicoloured clays with turbiditic intercalations, including the Genesio clays, Torrente Trenico marls and calcarenites, and Pianelli sandstones, dated from Eocene to Lower Oligocene26,27. Further south, sea cliffs cut through carbonate sequences of limestones and dolomites (Bulgheria unit; Figure 1, LITHO-4 and 5) ranging from Upper Triassic to Lower Miocene, overlain by Miocene terrigenous deposits25,27. Results The results concern both the ML model validation process and the reconstruction of the LIG coastal landscape. Model accuracy and scores (model validation) Using Python analytical libraries, we developed scripts to optimize the model (RandomForestClassifier) parameters. GridSearchCV with 5-fold cross-validation and parallel processing (n_jobs=−1) tested combinations of n_estimators, max_depth, max_features, min_samples_split, and min_samples_leaf to balance accuracy and generalization. The best parameters identified through this grid search were: {‘max_depth’: 15, ‘max_features’: 2, ‘min_samples_leaf’: 1, ‘min_samples_split’: 2, ‘n_estimators’: 200} Model accuracy is very high, but accuracy alone can be misleading for unbalanced classes. For Class 0 (sea cliff=False), Precision is 0.98 and Recall 0.96, indicating excellent identification of negative cases. For Class 1 (sea cliff=True), Precision is 0.96 and Recall 0.98, showing strong positive detection. The F1-score for both classes is 0.97, reflecting balanced performance. Macro and weighted averages are also 0.97, confirming model robustness across classes. After numerically evaluating the performance of the training dataset, the ML model was tested in the Cala del Cefalo and Licosa Cape areas (Table 1 and Supplementary Material 5), selected since several paleo-environmental markers are present (see Supplementary Material 4). Such markers collected during specific onsite surveys and bibliographic research allowed the extensive mapping of the LIG paleo-sea cliffs in both areas by obtaining the thematic map used for the validation process. Table 1Performance metrics of the ML classification for the two validation areas.Test PointFPFNTPTNNum. pointsAccuracySensitivityPrecisionF1-scoreMoran’s IndexArea A (Cala Cefalo)1041516596612500.900.920.610.730.467Area B (Punta Licosa)483716475110000.920.820.770.790.375 Performance metrics of the ML classification for the two validation areas. The evaluation was obtained by comparing the prediction map of the ML model with the thematic map. The evaluation was performed by applying a classification threshold on the probabilistic outputs of the model: predictions with probabilities below 0.60 were assigned as class 0 (False), while those equal to or above 0.60 were assigned as class 1 (True). This conservative threshold was chosen because it provided the best agreement with expert-based validation, reducing spurious positives while preserving high sensitivity. At Cala del Cefalo (Test Point Area A), the analysis of 1250 test points shows an overall accuracy of 0.90, confirming that the majority of points are correctly classified. Sensitivity reaches 0.92, indicating a strong ability of the model to correctly detect positive cases. However, precision is lower (0.61), reflecting the presence of false positives. The high sensitivity (0.92) demonstrates that the model rarely misses true features, while the lower precision (0.61) suggests some over-prediction, i.e., features classified by the model but absent in the expert mapping. The resulting F1-score (0.73) indicates a reasonable trade-off between recall and precision. At Licosa Cape (Test Point Area B), the evaluation of 1000 test points shows a slightly higher overall accuracy (0.92) compared to Cala del Cefalo. In this case, sensitivity is lower (0.82), meaning that some true paleo-cliff features remain undetected. Precision, however, is higher (0.77), pointing to fewer false positives and a more reliable classification of detected features. This balance is reflected in the F1-score of 0.79, which exceeds that of Area A. Overall, the model performs consistently well in both test areas. At Cala del Cefalo, the higher sensitivity (0.92) ensures that nearly all true LIG sea cliffs are detected, but the lower precision (0.61) indicates a tendency to over-predict. At Licosa Cape, the model achieves higher precision (0.77) and a better balance between recall and precision (F1-score = 0.79), even though sensitivity is lower (0.82). These results highlight that the model is reliable across different geomorphological contexts, with future improvements ideally focused on reducing false positives while maintaining high sensitivity. Finally, we evaluated the spatial distribution of errors using Moran’s I33 in both validation areas, in order to investigate whether model errors were randomly distributed or exhibited significant clustering in space. The obtained Moran’s I value indicates that misclassifications are not randomly scattered, but rather spatially structured the constraints imposed during model training as described in Supplementary Material 5. In particular, the ML model predicts as true some sectors/clusters that satisfy the morphometric conditions but are considered false positives due to the lack of paleo-environmental indicators. Instead, the clusters of false negatives can be explained by the elevation threshold (<200 m) that was imposed during training to avoid misclassification of inland mountain slopes. While this reduced false positives inland, it introduced a systematic bias in Area B, where high-relief cliffs extend above 200 m. These false negatives therefore reflect the methodological constraint rather than a true limitation of the Random Forest model. LIG coastal landscape detection Due to the considerable extent of the study area, comprising approximately 18 million cells, we focused our result on the detailed description of the analysis of the three most representative sectors: Sector 1 – Punta Licosa, Sector 2 – Palinuro-Marina di Camerota, and Sector 3 – Acciaroli (see Figures 2,Figures 3,Figures 4). In the first two sectors, the predictions generated by the ML model were compared with 344 known markers, enabling the establishment of strong and consistent correlations and validation. a) Overview of Sector 2 showing the identified MIS5 sea cliffs, with paleo-environmental markers (e.g., inner margins of marine terraces and wave-cut platforms) overlaid on the map. b) Zoomed-in map of Sector 2, highlighting the relationship between the LIG sea cliffs and the deposits burying the base of the slope, demonstrating why the bathtub approach is ineffective in this area. Map created in QGIS 3.38.3 (QGIS.org, https://qgis.org). TINITALY DTM (https://tinitaly.pi.ingv.it/) and EMODnet Bathymetry (https://emodnet.ec.europa.eu/). Satellite image: EOX, Sentinel-2 Cloudless (2020)62, data source: ESA Copernicus. Final figure created in Inkscape 1.3.2 (https://inkscape.org/release). a) Overview of Sector 3 showing the identified LIG sea cliffs, without paleo-environmental markers, correlated to three distinct LIG sea-level stands. b) Zoomed-in map of the Acciaroli Harbour area, highlighting the relationship between the MIS5 sea cliffs and the surrounding geomorphology (beach, talus deposits, and marine terrace). Map created in QGIS 3.38.3 (QGIS.org, https://qgis.org). TINITALY DTM (https://tinitaly.pi.ingv.it/) and EMODnet Bathymetry (https://emodnet.ec.europa.eu/). Satellite image: EOX, Sentinel-2 Cloudless (2020) [62], data source: ESA Copernicus. Final figure created in Inkscape 1.3.2 (https://inkscape.org/release). Flowchart illustrating the methodological workflow adopted for the identification and analysis of paleo- and polycyclic sea cliffs: a) dataset creation and training data extraction; b) model testing and application to the full dataset; c) field photograph taken during ground-truth surveys. Figure created in Inkscape 1.3.2 (https://inkscape.org/release). Photo by AS. In each sector, we detected two main types of sea cliff: Type 1 (T1) - LIG paleo-sea cliffs, located more than 10 meters from the present coastline, actively shaped by the sea during LIG but no longer subject to active marine processes today; Type 1 (T1) - LIG paleo-sea cliffs, located more than 10 meters from the present coastline, actively shaped by the sea during LIG but no longer subject to active marine processes today; Type 2 (T2) - Polycyclic sea cliffs, still subject to ongoing marine processes, display continuous or recent activity, preserving distinct morphological features. Type 2 (T2) - Polycyclic sea cliffs, still subject to ongoing marine processes, display continuous or recent activity, preserving distinct morphological features. Sector 1 – licosa cape Sector 1 is located in the northern part of the study area and extends for approximately 15 km, from San Marco to the harbour of San Nicola a Mare (Figure 5a). This area is predominantly characterized by LITHO-10, particularly the Pollica Sandstones, dated to 15.97 ± 0.05 Ma and 13.82 ± 0.05 Ma26. a) Overview of Sector 1 showing the identified MIS5 sea cliffs, with paleo-environmental markers (e.g., inner margins of marine terraces and wave-cut platforms) overlaid on the map. b) Zoomed-in map of the Punta Licosa area, highlighting the relationship between the first two orders of paleo-sea cliffs and other geomorphological features in the area, mapped through expert judgement. Map created in QGIS 3.38.3 (QGIS.org, https://qgis.org). TINITALY DTM (https://tinitaly.pi.ingv.it/) and EMODnet Bathymetry (https://emodnet.ec.europa.eu/). Satellite image: EOX, Sentinel-2 Cloudless (2020)62, data source: ESA Copernicus. Final figure created in Inkscape 1.3.2 (https://inkscape.org/release). Within this geological framework, the application of the machine learning model enabled the automated identification of LIG sea cliffs at three different altimetric ranges (Figure 5b), also testified by paleo-environmental markers detected on site. The ML model successfully (Supplementary Material 5 and Figure 5) identified a well-developed higher LIG paleo-sea cliff (T1), extending approximately 7 km, with an average slope of about 30°. Despite the footslope is widely covered by talus deposits, the model accurately (Supplementary Material 5) recognised its key morphometric characteristics. Despite the sea cliffs being widely recognized along the whole Cilento promontory, no absolute dating has been published till now. However, recent research by De Santis et al., 202523 assigns this landform to a sea-level reoccupation during the LIG. The intermediate feature detected by the ML model includes T1 and 2 sea cliffs. In fact, they are mostly situated between +11 and 4.5 m MSL but locally reach present sea level. This landform corresponds to a well-documented wave-cut platform, both in the literature34 and through field observations, where an inner margin at around +5 m MSL. According to Isola et al. (2024)34, at Cala Infreschi (Sector 2, Palinuro Cape), a speleothem dated to MIS 6–8 covers several Lithophaga boreholes at +5.3 ± 0.18 m MSL consequently associated to the LIG sea level, in agreement with the model automated detection. The best-preserved remnant of this morphology, located in the central-southern sector, extends for approximately 500 m along the coastline, with an average slope of 15.6° and a roughness value of 6.26, values that fall within the morphometric ranges learned by the model during the training phase. Finally, the model also recognised a lower sea cliff consistent with bibliographic reports and our own field surveys, though strongly eroded and preserved only in discontinuous stretches. The predicted segments align morphologically with adjacent scarps, supporting their interpretation as remnants of this degraded coastal feature This feature develops behind the lower wave-cut platform and, when corresponding solely to the lowest stand, has a maximum elevation of +5.2 m MSL, with the inner margin ranging between +1.5 and +2 m MSL. According to Isola et al (2024)34, this can be ascribed to a minor sea-level oscillation during the Last Interglacial (LIG). In many areas, these landforms are partially covered by recent deposits, making their reconstruction through direct observation alone particularly challenging. The reconstructions provided by the machine learning model suggest that the identified paleo-sea cliffs were not plunging cliffs, but rather sea cliffs with platforms (T2), characterised by an active erosional front overlying a wave-cut platform. Sector 2 – Palinuro-Marina di Camerota Sector 2 extends for approximately 15 km. As shown in Figure 2a, the ML model has detected an almost continuous sea cliff system with a prediction value, evaluated with validation metrics, higher than 70% throughout the sector. The polycyclic sea cliff (T2) here detected can be divided into plunging cliffs in the northern sector (Palinuro Cape) and sea cliffs with platforms in the southern sector near Marina di Camerota (Figure 2a). Upon detailed analysis of the ML model detections, a cliff within LITHO-7 is identified to the north of the sector, near Ficocelle (Figure 2a). This sea cliff extends approximately 500 meters horizontally, with vertical elevations varying between +2.4 and +19 meters above sea level. At Palinuro Cape, a prominent sea cliff is identified in LITHO-4 (Monte Crivo limestone), which extends parallel to the coast for approximately 9 kilometres. This sea cliff, with a maximum height of +190 meters, steep slopes, and high roughness values (an average of about 81), holds more than 100 sea-level markers. In this case, the bathtub approach proves particularly effective, as it works well in the context of plunging sea cliffs with onsite-identified wave-cut platforms. Instead, where the alluvial deposits partially bury this morphology and its slope, such as at Molpa promontory, the model detected a continuous T1 sea cliff in the inland portion. From the mouth of Mingardo River to Cala d’Arconte, a 4.5 km-long polycyclic sea cliff (T2) is identified within LITHO-4 and LITHO-5, spanning a 5 km stretch. Here, a 3 km-long pocket beach of high naturalistic value borders a huge relic morphology, with a maximum elevation of +170 m MSL and 300 m from the present coastline, featuring an average slope of 40 degrees. A comparison with detailed geomorphological maps and our paleoenvironmental markers (26 coastal caves, 5 Lithophaga boreholes, and 17 tidal notches) demonstrates the high accuracy of the model, with prediction values exceeding 80%. During field surveys, the above-mentioned markers were observed at altitudes between 2 and 10 meters MSL, predominantly between 5 and 10 meters. The prediction suggests that during the LIG, the sector was characterized by a continuous sea cliff, which remained active during the LIG minor sea-level falls. When overlaying the contour corresponding to RSL elevation during the LIG peak, it is clear that the classic bathtub approach works effectively for plunging cliffs but fails where sedimentary deposits are present, particularly slope deposits (i.e. talus deposits, cones, and significant landslides). Moving toward the beaches of Porticello and Cala d’Arconte, the precision with which the promontories separating the bays have been mapped is evident. In the valley, the sea cliff continues inland, indicating a marine transgression now buried beneath deposits (Cala d’Arconte in Figure 2A) The sector from Cala d’Arconte to Marina di Camerota is dominated by LITHO-4. As the cliff extends toward the bay of Calanca, it borders wave-cut platforms23, at about +4 - 7 m MSL. A notable feature along the Camerota coastline is a T1 sea cliff (within the Monte Crivo limestones), which rises at +34 m MSL. Overlaps with geomorphological and lithological maps reveal a significant volume of sediment within the valley, contributing to an apparent elevation distortion. This suggests that the area may have functioned as a ria during LIG, now buried under sediments. To the east, the sea cliff forms a bay-shaped structure partially filled with recent deposits up to +20 m MSL. The system continues to identify both types of sea cliffs along a 4 km stretch (in the Upper Dolomite), alternating between plunging cliffs (T2) and those with platforms (T1) at elevations consistent with LIG, locally intruding into a valley, suggesting a dynamic interaction between coastal processes and the inland topography during the LIG period. Sector 3 - Acciaroli Analysing the results in Sector 3 was particularly insightful, as the ML model used was primarily calibrated for areas characterised by the presence of paleoenvironmental markers. Therefore, applying it to regions devoid of such indicators provided an effective validation of the model capabilities, representing a methodological challenge. Sector 3 was chosen for its lack of paleoenvironmental markers. Consequently, the results described below rely solely on the ML model predictions. This aligns with the research goal of extensively reconstructing the paleo-seascape of the Cilento coast, including areas where no paleoenvironmental markers were available. Results are reported only as probabilistic predictions, constrained by LIG-compatible elevation bands and litho-geomorphic continuity. This illustrates how the method can extend reconstructions into areas where independent markers are lacking, while calibration and validation remain anchored to indicator-rich sectors. Sector 3 extends from Capitello to Marina di Casal Velino, covering a length of approximately 13 km (Figure 3). By analysing the results obtained from the ML model in relation to the available literature and detailed satellite images, its effectiveness was assessed. The ML model identifies a relatively continuous LIG sea cliff along almost the entire sector, with a good prediction value ranging between 79% and 90%. The LIG sea cliff extends for about 11 km, with elevations ranging from +10 to +117 m MSL, and a maximum slope of 55°. The lithological information in the dataset indicates the presence, in the predicted data, of LITHO-1 (Turbidites of the Saraceno formation, dated to 33.9 and 23.04 Ma) and LITHO-7 (Trenico Marls and Calcarenites, dated to 23 and 20.45 Ma26,27). To the north of Acciaroli (Figure 3), the paleo-sea cliff is located approximately 300 m from the present coastline, with a slope of about 15°. It is bordered initially by a large beach (Spiaggia Grande di Acciaroli) and subsequently by a talus and a coastal anthropised area. The ML model accurately predicted the target geomorphological feature, identifying the base of the landform at around 25 meters, with an average distance of 300 meters from the coastline. The undetected slope of this morphology was likely buried by aeolianites and slope deposits (probably on a wave-cut platform) during the LIG. To the south of Acciaroli Harbour, the sea cliff was reactivated during the Late Holocene when a wave-cut platform (see location on Figure 3) was shaped and later covered by recent gravel and block deposits26. In this case, an analysis was also carried out on the cell values corresponding to the two types (LIG paleo-sea cliffs and polycyclic sea cliffs, see supplementary Material 1, Figures S1, Figures S2 and Figures S3), with no significant asymmetries emerging. Discussion According to our machine learning reconstructions, the LIG landscape of the Cilento coastal sector was predominantly characterised by very high plunging cliffs carved into resistant limestone formations (e.g., LITHO-4 and 5), interspersed with shorter sea cliffs occasionally bounded by platforms incised into more erodible lithologies such as sandstone and mudstone (LITHO-1, 7 and 10). This morphological contrast, along with the consequent morphometric distinctions between paleo-sea cliffs (T1) and polycyclic sea cliffs (T2), highlights the significant role of the underlying lithology in both the development and long-term preservation of coastal landforms shaped during LIG. In Figure 6, the LIG coastal landscape was reconstructed to underline that the classic bathtub approach with a shoreline at +5.3 ± 0.18 m MSL34 is effective only for T2 sea cliffs but fails where sedimentary deposits bury the footslope of T1 sea cliffs (i.e. talus deposits, cones, and significant landslides). Coastal paleo-landscape reconstruction during the Last Interglacial (LIG), showing the detected LIG sea cliffs with a prediction probability greater than 60%. Contour lines are displayed at 50 m intervals. Map created in QGIS 3.38.3 (QGIS.org, https://qgis.org). Topography from TINITALY DTM (https://tinitaly.pi.ingv.it/), bathymetry from EMODnet (https://emodnet.ec.europa.eu/). Final layout edited in Inkscape 1.3.2 (https://inkscape.org/release). Furthermore, our analysis allowed evaluating the climatic effect on the different lithologies underlining the morphological differences between T1 paleo-sea cliffs, active during the LIG and now inactive, and T2 polycyclic sea cliffs, currently subject to marine erosion and representing the active coastal front. Specifically, roughness exhibits higher values and variability in active T2 sea cliffs (e.g., median ~15.5 in turbidites, see Supplementary Material 3, Table S1) compared to stabilized T1 forms (~11.9), reflecting localized surface irregularities and enhanced erosion. Slope distributions similarly differ, with T2 sea cliffs showing steeper median slopes (~25.5°, see Supplementary Material 3, Table S1) than T1 (~22.7°, see Supplementary Material 3, Table S1), consistent with dynamic instability. Surface area ratio (SAR) decreases modestly from T2 (~111) to T1 (~108.4) in turbiditic cliffs, indicating a subtle flattening of surfaces at broader spatial scales (see Supplementary Material 3). Although SAR correlates moderately with roughness (Pearson’s r ~0.72) and slope (~0.79), it captures morphological complexity integrated over larger areas than the roughness metric, which is sensitive to microtopographic variation. In general, data highlight systematic differences between T1 and T2, consistent with their different degree of morphogenetic activity (see Supplementary Material 1 and Supplementary Material 2): T2 are more active, more fractured and more rough, typical of coastal slopes in active erosion. T2 are more active, more fractured and more rough, typical of coastal slopes in active erosion. T1 show more stabilized surfaces, with less morphological complexity, consistent with a relict and inactive form. T1 show more stabilized surfaces, with less morphological complexity, consistent with a relict and inactive form. In our model, roughness and slope emerge as the most influential predictors, with roughness showing the highest variable importance, consistent with its sensitivity to fine-scale heterogeneity in actively eroding cliffs. SAR, while less important, contributes unique information on meso-scale surface complexity, justifying its inclusion despite correlation with roughness and slope. The high correlation between roughness and slope (r = 0.91) suggests some redundancy; however, their combination improves classification robustness by jointly describing both texture and gradient controls on erosion. The lower correlation of SAR with these parameters reflects its ability to capture distinct spatial patterns of surface dissection, providing complementary morphometric information at a different scale. In summary, the combined use of slope, roughness, and SAR in the supervised classification framework effectively leverages their complementary scales and morphogenetic significance, enhancing discrimination between T1 paleo-sea cliffs and T2 active sea cliffs across diverse lithologies and erosional states. However, our analysis not only allowed assessing the effects of the climatic forcings of the last interglacial cycle on the two different types of paleo-sea cliffs but also discriminating which lithologies better support machine learning models for automatic classification. By considering differences between the two typologies for each lithology, we can further observe that active turbidites (T2) clearly show signs of marine erosion, with high morphometric variability. Notably, T1 is rougher and more complex. This may reflect lithological resistance, as these rocks are less susceptible to erosion. In the case of the dolomitic lithology, both typologies are morphometrically stable. This may represent an outlier in the classification. The Calcarenite, instead shows an inverted pattern, like the Sandstone. This suggests a higher resistance to erosion or a “morphological memory” in the T1. The Marl has minimal differences between T1 and T2, making it a lithology with little discriminating power for classification purposes. In summary, turbidite is the most effective lithology to distinguish T1 and T2 with significant differences on all parameters. On the other hand, Sandstone and Calcarenite show inverted patterns: paleo-sea cliffs are more complex than active sea cliffs, probably for geomechanical or historical reasons. Finally, Dolomite and Marl have very similar values, and although they may not contribute strongly to the classification, they corroborate the low climatic sensitivity of these lithologies over time. In terms of lithology, for the turbiditic sequences (LITHO-1, see Supplementary Material 3, Table S1) that dominate the northern area, T2 cliffs show moderate slopes (median 25.5°), elevated roughness (median ~15.5), and notable dissection (~0.49), indicating active morphological instability driven by marine erosion and frequent slope failures. Corresponding T1 paleo-sea cliffs exhibit gentler slopes (~22.7°), lower roughness (~11.9), and similar dissection (~0.46), consistent with surfaces smoothed by subaerial weathering and colluvial deposits. Surface area values decrease slightly from T2 (~111) to T1 (~108.4), marking a subtle but consistent flattening. For this lithology, the mean roughness considerably exceeds the median, especially in T2 (mean ~20.4 vs. median ~15.5), indicating localized surface irregularities and intensified erosion zones. The roughness distribution shows a long right tail, reflecting heterogeneous morphology. Slope distributions are similarly positively skewed with a right tail due to very steep sections, as confirmed by maximum slope values above 60°. Dissection has a slightly positive skewness in T2, with higher variability than T1, consistent with active surface fragmentation. Standard deviations reinforce the presence of significant variability in all parameters, more pronounced in T2, indicating a dynamically evolving landscape. Conversely, in well-cemented sandstone areas (LITHO-10, see Supplementary Material 3, Table S1), such as parts of central Cilento, slopes are relatively uniform (~24.9° for T1 and ~23.2° for T2). T1 paleo-sea cliffs present higher roughness (~14.5) and dissection (~0.44–0.46) compared to T2 (~12.3 and ~0.42), possibly due to recent smoothing from sediment mantling or differential retreat along bedding planes. Surface area remains stable (~110–111), indicating retention of relief complexity despite minor slope differences. The statistical distribution of these parameters tends to be more symmetric with means close to medians and moderate standard deviations (see Supplementary Material 3, Table S1), suggesting more uniform erosional dynamics and material behavior, supporting an interpretation of stable, steady-state erosional conditions. However, from a statistical viewpoint, carbonate sea cliffs show the highest positive skewness in roughness distributions among all lithologies, with means (45.3 for T1 and 56.4 for T2) significantly higher than medians due to extreme outliers (maxima over 1000). This suggests that while much of the sea cliff is rugged, some localized sectors have exceptional morphological complexity likely driven by structural controls such as joints. Slopes also show positive skewness, with maximum values approaching 78°, indicative of steep or even overhanging faces. Dissection distributions display moderate positive skewness and high variability, consistent with pronounced surface fragmentation. Standard deviations for roughness and slope are notably high, reinforcing the heterogeneous nature of these sea cliffs. Calcarenite sea cliffs (LITHO-6, see Supplementary Material 3, Table S1), exhibit moderate slopes (~19.1° in T2, ~20.1° in T1), modest roughness (~8.2 for T2 and ~9.0 for T1), and low to moderate dissection (~0.40 in T2, ~0.46 in T1). Median surface area (~106) suggests limited vertical complexity, consistent with subdued slopes and partial vegetation cover. These sea cliffs show relatively symmetrical distributions with means and medians close in slope and dissection parameters. Roughness presents mild positive skewness especially in T2 (mean 15.6 vs median 8.16) with some moderate outliers (max ~86), indicating that while most of the terrain is relatively smooth, some localized surface irregularities do occur. Standard deviations are moderate, supporting overall homogeneity with some patchy complexity. Finally, marly lithologies (LITHO-7, see Supplementary Material 3, Table S1), despite mechanical weakness, show unexpectedly similar metrics between T1 and T2 forms. Median slopes remain around 22.5–22.7°, roughness values between ~11.6 and 11.9, and dissection near 0.44–0.45, with comparable surface areas (~108.5). This indicates that post-LIG morphological evolution has likely reached relative equilibrium, with earlier degradation smoothing the paleo surfaces. Statistical analysis confirms this equilibrium: means and medians are close across all parameters, with limited skewness and only occasional outliers, mainly in roughness for T2. This suggests stable morphological conditions with balanced erosion and surface modification processes over time (see Supplementary Material 3). As far as concerns the spatial resolution of our dataset, accurate ML-based detection of these features is complicated by DTM resolution: coarser grids can overlook subtle morphological cues, while finer resolutions may amplify noise. In this study we relied on a 10 m DTM; however, we consider that employing a multi-resolution DEM framework, where available, could in the future enhance the robustness of landform classification, despite the usual scarcity of high-quality topographic datasets. In summary, the integration of morphological and lithological information (see Supplementary Material 3) in a ML model allowed us not only to reconstruct the LIG coastal landscape even when partially covered by more recent deposits or modified by subsequent processes, overcoming the limitations of the classic bathtub approach, but also to evaluate how climatic forcing since the LIG differently affected the various lithologies, conditioning their morphological response. The consistency of our result with the ground truth validation (see Supplementary Material 5) supported the extrapolation of this landscape reconstruction to those coastal areas without any detected LIG sea-level evidence. In turn, these detections provided detailed morphometric information on the successive climatic forcings that influenced the evolution of the coastal landscape during the Last Interglacial cycle. Beyond the geomorphological insights, this study also reflects on the methodological implications of applying machine learning to complex coastal landscapes35. A key advantage of Random Forest (RF) lies in its explainability and interpretability, offering clear insights into how predictors (see Figure 7) influence classification outcomes. a) Contribution of individual independent variables (predictors) to the final prediction of the dependent variable (target); b) Correlation matrix illustrating the relationships among the morphometric variables used as predictors in the machine learning classification; c) SHAP analysis plot. The Y-axis shows the model features ranked by impact, while the X-axis represents SHAP values, indicating each feature effect on the prediction; colours reflect feature values (blue = low, red = high). The distribution around zero suggests a balanced model; d) SHAP dependence plot for the roughness variable (X-axis), with SHAP values on the Y-axis indicating its influence on the prediction. Colours represent slope values (blue = low, red = high). Cells with high slope and roughness values tend to be positively classified as target; a plateau suggests a threshold beyond which additional roughness has minimal effect. Statistical plots generated in PyCharm 2022.2.5 (https://www.jetbrains.com/pycharm/download/other.html) using Pandas, Seaborn, Matplotlib, Pyplot, Shap, and NumPy libraries. Final composition edited in Inkscape 1.3.2. The RandomForestClassifier proved to be a robust and effective tool for the automatic detection of coastal cliffs and paleo-cliffs, achieving high overall accuracy in both training and validation areas. However, our analysis of error patterns highlighted some aspects related to the false-negative clustering due to an altimetric constraint imposed in the training dataset and the discrepancy between the “human” mapping approach and the morphometric calculations on a cell-by-cell basis. Our study shows that while machine learning offers reproducible, scalable, and highly accurate solutions for coastal geomorphological mapping, the interpretation of results must always account for both training design choices and the intrinsic differences between human and algorithmic perception of landscape structures. Future work should aim to integrate the two perspectives, coupling the objectivity of morphometric models with expert-driven contextual knowledge, in order to improve both accuracy and interpretability. In conclusion, this study proposes a new and effective methodology not only for regional-to-continental- scaled paleo-landscape reconstruction but also for the statistical evaluation of the climatic-driven effects on its millennial evolution. The application of such an approach at a global level could support the planning of adaptation strategies to the expected effects of the ongoing climate change. Methods The proposed methodology points out to detect the LIG paleo-sea cliffs in the whole study area through a Machine learning-based procedure applied to a 10x10 m geological-geomorphological dataset of more than 18.000.000 points obtained from the morphometric analysis of open-access sources (see section “LIG coastal landscape detection”). A RandomForestClassifier algorithm was used to create a Landform Classification system capable of identifying the paleo-sea cliff involved in the sea level oscillations that occurred during the LIG. Individuations and feature extraction The procedure involves two phases: the first phase (Figure 4a) is based on the creation of the morphometric dataset of the entire study area and the extraction/creation of the training dataset; the second phase (Figure 4b) focuses on the creation of the ML Model and related application. The workflow integrates morphometric analysis, machine learning classification, and a double-level validation process. Initial data preprocessing involved the extraction of topographic derivatives (slope, roughness, and others) from a 10 m resolution DTM. This dataset of about 18-million points was superimposed with a dataset deriving from the SEA-PROXY geodatabase. The resulting integration provided us a training dataset for the RandomForestClassifier, satisfying the design criterion in which both conditions (sea cliff yes/sea cliff no) are contemplated (Extraction phase in Figure 7a). After the model validation, the full dataset obtained from the DTM analysis was processed by the RandomForestClassifier to obtain a prediction over a broader area (Fig 7b). The final classification supported the reconstruction of LIG coastal landscapes, and the spatial delineation of geomorphic features related to sea-level high stands. The morphometric analysis was applied to a 10-meter resolution DTM36 by calculating a series of terrain derivatives (see Supplementary Material 2, DEM elevation, Slope, Roughness, Surface Area, Dissection, Lithology, Distance from coast), mostly thanks to Geomorphometry and Gradient metric toolbox37. The morphometric parameters were then coupled with lithological data obtained from the OneGeology portal38, an international initiative of the world geological surveys and other international and regional geoscience organisations (https://portal.onegeology.org/OnegeologyGlobal/), which in this area is based on the CARG geological map of Monti et al. (2014)25. This morphometric and lithological dataset was overlaid with expert-level information (i.e. independent datasets compiled from different sources used to constrain the presence and distribution of LIG features), including the SEA-PROXY39 geodatabase populated with data from bibliographic sources and onsite surveys (see Figure 4c), as well as the results of satellite images analysis. For each proxy point, we extracted the surrounding points matching the morphometric characteristics of a paleo-sea cliff, ensuring spatial alignment. The result of this integration allowed us to identify a dataset composed of 1450 features (Feature extraction) compatible with the target features (LIG sea cliffs). The records of this dataset, including the related terrain derivatives, were classified through a supervised analysis, according to ML approach, assigning prediction values (Labelling), both false (prediction value 0) and true (prediction value 1) in order to obtain raw training dataset for the machine learning model (Figure 4b). Finally, the so-trained ML model was then applied to the whole Real Dataset of about 18 million cells (data not known by the algorithm), to obtain the detection of LIG sea cliffs in the whole study area. Data The dataset is composed of three main sub-sets. The first sub-dataset, extracted from the DTM and enriched with terrain derivatives, as described above. The second sub-dataset includes sea-level indicators derived from geological, geomorphological, paleoenvironmental, and paleontological markers, primarily sourced from the SEA-PROXY geodatabase40,41, an open-access platform compiling sea-level proxies following recent international guidelines42,43. These data were supplemented by extensive bibliographic research23–25,29,32,34,44–51 and analysis of satellite imagery (1984–2023) from Google Earth52. The third sub-dataset comprises 186 newly acquired records from onsite coastal surveys aimed at identifying additional paleoenvironmental markers, incorporated into the expert judgement analysis. By overlaying these three sub-datasets, the final training dataset was built, comprising 1,450 records. The entire dataset was managed within a dedicated GIS project, enabling the spatial analyses detailed in Supplementary Material 2. The spatial distribution of the training dataset is random. Consequently, where spatial gaps exist—i.e., no indicator coverage (no expert-mapped points or sea-level indicators), we do not fill them; instead, after validation, we apply the trained model to predict target features in those uncovered areas (see Sector 3). Machine learning methodology The supervised ML method is characterised by training data with a known classification of the target variable, i.e. the geomorphological unit corresponding or not to the paleo-sea cliff typologies. The model development involved a training procedure by splitting the 1450-sample dataset into two parts: 70% for the training set and 30% for the testing one. For the development, training, and validation of the machine learning model, Python was employed with the Scikit-Learn RandomForestClassifier library, which provides a user-friendly interface to create and train an ML model. RandomForestClassifier is a supervised learning algorithm that combines multiple trees to create a single model. Each tree in the random forest was built on a different subset of data and performs its estimation independently. The final prediction was based on the average or weighted average of the predictions from all individual trees. The employed RandomForestClassifier algorithm offers two types of output: a binary classification mode (predict), where it predicts a discrete binary class (typically 0/1 or True/False), and a probabilistic prediction mode (predict_proba), where it estimates the probability of class membership as a percentage (between 0 and 1), representing the percentage that the cell belongs to the target variable. In this paper, we implemented both methodologies. Training dataset After extracting the training dataset using the methodology described above, we proceeded with the training, testing, and validation phases. During dataset composition, balancing features and assessing their impact on predictions is crucial. Therefore, we identified the most influential independent variables (predictors) driving the model outputs, while also detecting redundant or minimally impactful ones. This process streamlines the selection of dataset fields. This phase was developed through Python scripts in synergy with libraries such as NumPy, Pandas, Seaborn, Matplotlib, as well as numerous Scikit-learn modules. Figure 7a shows the Feature Importance graph, a bar histogram ranking predictors by their impact on prediction53. This analysis is key to balancing the feature set, highlighting the most influential predictors, and revealing redundant or low-impact variables. The graph clearly indicates that roughness, along with slope and surface area, carries the greatest weight in the machine learning process-consistent with the physical, orographic, and topographic nature of the target variable. After identifying the impact scale of the various predictors within the model, to obtain a clearer view of the interactions between features and the final prediction, a correlation matrix (using method=‘Pearson’) was created (Figure 7b). This matrix clearly highlights the strong correlation between the predictors slope and roughness (greater than 0.90), as well as between surface area, slope, and roughness (see Figure 7b). The correlations observed in the Correlation Matrix are further clarified through SHAP analysis (SHapley Additive exPlanations)54–57, which offers a detailed interpretation of feature contributions and their interactions with the prediction outcome, as discussed below. SHAP is a powerful interpretability technique grounded in game theory, assigning equitable contributions of each input variable to the final model prediction via Shapley values54. In geomorphological modelling, SHAP decomposes complex outputs - such as landscape change, coastal retreat, or vegetation degradation - into additive contributions from individual morphometric or environmental features. This quantifies how terrain attributes or climatic variables influence predictions, providing transparent, interpretable insights54,55. Explainability was a primary reason for choosing Random Forest as the ML method. A major advantage of SHAP is its model-agnostic nature, applicable across many algorithms, from decision trees to ensembles. This flexibility suits nonlinear processes and complex feature interactions typical of Earth system sciences. Combined with expert cartographic knowledge and paleoenvironmental markers, SHAP strengthens model interpretability, linking computational predictions with field-validated evidence54–57. The SHAP value (Figure 7a) indicates the influence of each variable on the prediction of the target variable, whether positive or negative. The analysis reveals that roughness and slope are the variables with the greatest impact on the model; cells with high slope and roughness values tend to positively classify the examined cell as the target. The variables surface area and dissection also have a significant, albeit smaller, impact on the final prediction. Finally, distance, DTM_elevat, and lithology exert a more limited influence. In the case of distance, very high values contribute negatively to the prediction, since the training dataset was configured to discourage the positive prediction of geomorphological forms that are very far from the coast. The distribution of points around zero indicates that the model is fairly balanced. Figure 7b displays a SHAP Dependence Plot, which compares two features using the SHAP technique. The features shown are slope and roughness, highlighting their strong correlation, as also evidenced in the Correlation Matrix of Figure 7b. The trend indicates that with low roughness values (between 0 and 10), the SHAP values are negative or near zero, especially for low slopes (dark blue). This underscores that low roughness values, combined with low slope values, contribute negatively to the probability of the target class (evaluating the examined cells as areas of little interest for the model target). This trend shifts at a relatively precise threshold; around roughness values of ≈10–15, the SHAP values become positive, particularly when slope values are high (colours trending toward red). The graph also reveals a plateau effect: after a roughness value of 30, the SHAP impact stabilises between 0.1 and 0.2. Even in the presence of higher slopes (red), the influence of roughness remains positive but does not increase further. In conclusion, the variable Roughness is a strong predictor, but only beyond a certain threshold (≈10). After identifying and analysing the impact of each feature on the outcome and the existing correlations among them, a numerical validation methodology of the ML model was applied. Performance and model validation The validation methodology was structured in two distinct blocks. The first block identifies the best possible modelling algorithm through a series of tests with mathematical/statistical evaluations, while the second block verifies the consistency of the predictions on a test point dataset based on real data from areas unknown to the model (Ground Truth). Statistical/Mathematical Evaluation Metrics - To evaluate the training model, metrics such as Accuracy, the Confusion Matrix, AUC, Precision, Recall, and F1-score were used. In particular, the F1-score (Equation 1) is a metric that combines Precision and Recall into a single value, which is useful for assessing classification models, especially when the data is unbalanced. Where: P is Precision, i.e. the proportion of samples classified as positive that are positive. R is Recall, i.e. the proportion of positive samples that were correctly identified. Quantifying model performance (Real dataset) - For the second block, a methodology was developed based on applying the ML model to real data unknown to the model itself. To this end, a test area (see Figure 2) was created at Cala del Cefalo (Test Point Area A – Cala del Cefalo), located to the east of the mouth of the Mingardo River, with a length of approximately 4,100 meters and a width of about 1,000 meters. It is characterised by a paleo-sea cliff with heights exceeding 100 meters, bordered mostly by a sandy beach with dunes that bury the footslope. In this area, the ML model was compared with an expert judgment by using five different grids (see Figure 6), each composed of 250 random points (using QGIS tools, Random Points Inside Polygon) with corresponding prediction results (True and False). Finally, we conducted a spatial correlation analysis to detect possible patterns. The statistical evaluation was performed using the Global Moran’s Index (Equation 2)33, which quantifies the degree of spatial autocorrelation between the location and the associated value of each test point. To assess whether the errors were spatially random or exhibited clustering, we employed a Python script using the test point dataset and the PySAL library, specifically the esda and libpysal modules (see Supplementary Material 5). On-site surveys and paleoenvironmental marker (Ground Truth) Geomorphological research that employs computational analysis tools such as machine learning requires, for robust validation, a series of field-based observations (see Figure 4c). This methodological step is commonly referred to as ground truthing. “Ground truth” may be understood as the conceptual framework concerning the verification of a given phenomenon through direct observation. In our case, we assessed the geomorphological features identified via ML models by comparing them with empirical field data. This comparison forms part of an iterative process, whereby feedback from the alignment (or misalignment) between model outputs and ground observations is used to refine and calibrate the ML algorithms. Extensive on-site field surveys were conducted across the study area, both direct and indirect methods were used to identify project-specific geomorphological features and paleoenvironmental markers. The topographic surveys were conducted using a high-precision RTK GNSS system (Emlid Reach M2), mounted on a 2-meter pole to ensure optimal satellite signal reception and positioning accuracy. Data were acquired in static mode, with the receiver remaining stationary for 2 minutes per point at least, allowing for increased measurement reliability under fixed RTK solution. In addition, the Autel Evo II RTK Series V2 drone was used during indirect surveys to obtain detailed information on the characteristics of Cala del Cefalo. The integration of new field survey data with pre-existing information enabled the development of high-resolution geomorphological maps of specific sectors in the study area. These served as both a comparative baseline and a validation dataset for the evaluation and refinement of the ML models employed in the study. The paleo-environmental markers used to populate our training dataset and to validate the model derive from both erosional and depositional sea-level markers (sensu Mattei et al., 202240.) altimetrically correlated to the LIG (i.e. +8 ± 1, +5.3 ± 0.18, +4 ± 0.5, +2.5 ± 0.5 m MSL23,29,34,58–60,). Specifically, the erosional features include Lithophaga boreholes, tidal notches, basal floors of coastal caves, marine terraces and related inner margins29,44–46,61. The depositional features include shoreface and backshore deposits with fossils shells32,58. The indicative meaning, reference water level, and classification of these markers are summarized below: Lithophaga borehole Lithophaga borehole • Marker Type: SLIP • Marker Type: SLIP • IR: MLW to MHW • IR: MLW to MHW • RWL: (MHW – MLW)/2 • RWL: (MHW – MLW)/2 Tidal notch Tidal notch • Marker Type: SLIP • Marker Type: SLIP • IR: MLW to MHW • IR: MLW to MHW • RWL: (MHW – MLW)/2 • RWL: (MHW – MLW)/2 Floor of coastal cave Floor of coastal cave • Marker Type: SLIP • Marker Type: SLIP • IR: MHW to MHHW • IR: MHW to MHHW • RWL: Midpoint of range • RWL: Midpoint of range Inner margin of marine terrace Inner margin of marine terrace • Marker Type: SLIP • Marker Type: SLIP • IR: MHHW to [context-dependent] • IR: MHHW to [context-dependent] • RWL: Stratigraphic or geomorphic assessment • RWL: Stratigraphic or geomorphic assessment Shoreface deposit Shoreface deposit • Marker Type: SLIP • Marker Type: SLIP • IR: MLW to MHW • IR: MLW to MHW • RWL: (MHW – MLW)/2 • RWL: (MHW – MLW)/2 Backshore deposit Backshore deposit • Marker Type: TLP • Marker Type: TLP • IR: Above MSL • IR: Above MSL • RWL: No marine reference • RWL: No marine reference Field-verified elevation markers were cross-referenced with ML outputs through a GIS-based overlay analysis, allowing for both spatial correlation and altimetric alignment. Supplementary Information Supplementary Information 1. Supplementary Information 2. Supplementary Information 3. Supplementary Information 4. Supplementary Information 5. Acknowledgements This work was developed within the scientific framework of the following projects, benefiting from discussions held during the meetings: •INQUA INR ONSEA – Evolution of Seascapes (INQUA CMP Project 2404); •context-AwaRe deCision-making for Autonomus unmmaneD vehicles in mArine environmental monitoring (ARCAD-IA) project (PE00000013_1 - CUP E63C22002150007) cascade call of the Future Artificial Intelligence Research (FAIR) project Spoke 3 - Resilient AI, within the National Recovery and Resilience Plan (PNRR) of the Italian Ministry of University and Research (MUR); •Working Group of the International Association of Geomorphology in Coastal Geoarchaeology •European Union - Next-GenerationEU - National Recovery and Resilience Plan (NRRP) – MISSION 4 COMPONENT 2, INVESTIMENT N. 1.1, CALL PRIN 2022 D.D. 104 02- 02-2022 – PRIN_2022ZSMRXJ “GAIA project- Geomorphological and hydrogeological vulnerability of Italian coastal areas in response to sea level rise and marine extreme events CUP I53D23002130006. We are grateful to the Soprintendenza Archeologia, Belle Arti e Paesaggio for the provinces of Salerno and Avellino for their collaboration and for authorizing fieldwork activities in the study area. We sincerely thank Luigi De Luca for his valuable support in drone data acquisition and processing, and Francesco Peluso for his indispensable support and insightful discussions throughout the research. This study made use of the following softwares: ArcMap 10.4 (Esri, https://www.esri.com), QGIS version 3.38.3 (QGIS Development Team, https://qgis.org), and Inkscape version 1.4.2 (Inkscape Project, https://inkscape.org). Digital Terrain Model (DTM) data were obtained from TINITALY DEM (https://tinitaly.pi.ingv.it/), DBM from EMODnet (https://emodnet.ec.europa.eu/). Original geological data were based on OneGeology (https://portal.onegeology.org/OnegeologyGlobal/) and the geological sketch by Monti et al. (2014): Carta Geologica d’Italia Sapri-Foglio 520 della carta 1:50.000 dell’IGM. All figures and illustrations were created by the authors. We sincerely thank the Editor and the Reviewers for their constructive comments and insightful suggestions, which significantly improved the clarity and quality of this manuscript. Author contributions GM contributed to the conception, acquisition, analysis, interpretation of data, creation of new software used in the work, and writing of the draft and revised versions. AS contributed to the conception, acquisition, analysis, interpretation of data, creation of new software used in the work, and writing of the draft and revised versions. GP contributed to the conception and supervision. AC contributed to the conception and creation of new software used in the work. PPCA contributed to the conception, analysis, interpretation of data, and supervision. All authors reviewed the manuscript. Data availability The data that support the findings of this study are contained in the manuscript and supplementary materials. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Information 1. Supplementary Information 2. Supplementary Information 3. Supplementary Information 4. Supplementary Information 5. Data Availability Statement The data that support the findings of this study are contained in the manuscript and supplementary materials.