Automated Dating of Recent Landslides Using Sentinel-2 and Sentinel-1 on Google Earth Engine

Highlights

What are the main findings?

• Development of an algorithm for automatically dating recent landslides in vegetated areas.
• Introduction of a reproducible and fully automated methodology, applicable to multiple study areas.

What is the implication of the main finding?

• Objective approach for landslide dating analysis.
• Support for defining trigger thresholds and developing predictive models for landslide risk management.

Abstract

Landslides are complex phenomena controlled by natural and anthropogenic factors. In recent years, the need to understand their dynamics has driven the development of methodologies for improving risk monitoring and mitigation. In this context, landslide occurrence dating helps identify triggering causes and critical thresholds. This study introduces a fully automated and objective methodology, implemented on the Google Earth Engine platform, which allows access to and processing of large volumes of satellite data online, speeding up analyses and facilitating method sharing. The procedure exploits the complementarity between changes in vegetation cover detected through vegetation indices and changes in radar backscattering, intending to narrow the time window in which the landslide occurred. In 45 out of 46 cases analyzed, the time interval of landslide occurrence could be correctly identified, with a mean temporal window of approximately 8 days (range—3–12 days), confirming the robustness of the approach across different geomorphological settings and landslide types. The complete automation of the workflow is among the most innovative aspects of the methodology, as it allows the script to be directly and consistently applied to a wide range of recent and vegetated landslides with sizes larger than about 10 Sentinel-2 pixels without requiring additional manual procedures.

1. Introduction

Landslides are movements of masses of rock, debris or soil along a slope under the influence of gravity [1], characterized by high spatial, temporal and velocity variability [2]. Their occurrence depends on natural factors such as ground conditions (e.g., topography and lithology), seismic events, and meteorological and hydrological conditions [3]; anthropogenic factors, such as deforestation and changes in land use, also increase the likelihood of landslides [4]. Climate change has significantly influenced the frequency and intensity of landslide events [5]: rising global temperatures contribute to changing precipitation regimes and intensify extreme weather events [6], which saturate the soil and reduce its stability [7].

The socio-economic consequences of such phenomena are significant [8]: the direct impacts on populations and infrastructure [9,10], combined with the high costs of emergency management and reconstruction, result in economic costs that weigh heavily on the sustainable development of the affected areas [11]. For these reasons, a thorough understanding of landslide mechanisms is crucial for risk mitigation. Identifying susceptible areas [12,13] and implementing accurate predictive models allow for preventive measures and timely interventions [14], safeguarding populations and reducing the economic impact of events [15].

Landslide prediction models are mainly based on historical data from inventories of past landslides [16], which attempt to correlate the location of events with predictive variables that reflect geological, geomorphological, and climatic characteristics of the area. The localization of landslide events can be carried out through field surveys [17]; however, these traditional methods are often time- and resource-consuming, especially in large or difficult-to-access areas. Consequently, the integration of alternative techniques such as remote sensing image analysis [18] is becoming increasingly central to the efficiency of monitoring and forecasting processes [19,20]. Thanks to the availability of spatially and temporally high-resolution data, which can be accessed freely or at low cost [2,21], remote sensing is assuming an increasingly central role in the production of up-to-date and detailed inventories [22]. In addition to allowing the detection and mapping of landslides [23,24], satellite imagery also makes it possible to estimate the date of occurrence of the landslide event [25], information that is essential but not always available to correctly attribute the triggering cause.

Knowing the initiation time represents a fundamental factor in defining critical thresholds [26]: for example, knowing exactly when the event occurred makes it possible to more accurately determine the intensity and duration of the precipitation that led to the slope failure. In the literature, the comparison of pre- and post-event satellite images [27] has proven extremely useful for identifying the exact chronology of events and for performing an in-depth analysis of soil changes [28]. Traditionally, these comparisons often rely on semi-automated technologies, with QGIS, Python, or R, which require manual intervention for data interpretation. Such a comparison makes it possible to identify changes in soil structure and vegetation cover [29], highlighting areas that have undergone significant alterations because of the landslide displacement [30].

Several studies have addressed the problem of determining the date of landslide occurrence [25]. Automatic or semi-automatic dating of landslide events is indeed a crucial aspect for improving the quality of temporal inventories and supporting susceptibility, hazard, and risk assessments. Recent approaches propose the integration of optical and radar data, such as Sentinel-2 and Sentinel-1 [22,31], to detect variations in spectral and backscattering signals, useful for identifying the period of landslide activation. These methods make use of multiple platforms, such as QGIS, Python, R, MATLAB, and Google Earth Engine, and include an interpretation phase of the results based on the observation of graphs or images [19,29], which may introduce some variability in the results.

In this context, the development of a fully automated workflow implemented within a single web platform could represent a desirable evolution; earlier semi-automated approaches still require manual intervention for tasks such as threshold selection and the interpretation of graphs or images, introducing subjective decisions that could affect the results. The proposed methodology has a limitation related to the temporal resolution of satellite data, which cannot guarantee the same accuracy as ground-based monitoring systems, such as seismic signals or direct field observations. However, landslides often occur in remote and difficult-to-access areas, where it is not always possible to carry out immediate surveys, and in many cases, the event is only recognized after the fact rather than in real time. In this context, the use of satellite images makes it possible to identify time intervals of landslide occurrence, obtaining results in complex terrains where ground monitoring is often difficult to perform.

The novelty of our workflow lies in the systematic integration of optical and radar satellite data into a fully automated, objective, and replicable process developed within a single working environment. Unlike other methods that require multiple environments and subjective choices, our approach allows us to narrow down the time interval of landslide occurrence autonomously, ensuring consistent results in both large and small areas. This combination of automation, objectivity, and accessibility represents the main scientific contribution of the methodology and highlights its originality compared to traditional, well-established approaches, making the workflow a fast and operational tool for landslide monitoring. The overall objective is to efficiently identify the initiation time of recent landslides, located in vegetated areas, with a minimum size of approximately 1 ha (~10 Sentinel-2 pixels), using exclusively open-source satellite data and a workflow entirely developed on Google Earth Engine [32,33]. The method integrates Sentinel-2 optical and Sentinel-1 radar data, exploiting the complementarity between vegetation indices and radar backscattering. The Normalized Difference Vegetation Index (NDVI), obtained from the normalization of the surface reflectance in the NIR and Red bands, is used to identify a time interval in which a significant change in vegetation cover is observed, allowing for the identification of pre- and post-event dates.

The temporal variability of vegetation cover, and thus of vegetation indices, influences the temporal evolution of backscattering [34]. When relying solely on optical data, cloud cover represents a limiting factor for the proposed methodology. Soil backscattering is available even under overcast conditions, provided that the atmospheric systems are not heavily laden with precipitation. However, given that backscattering also varies with vegetation phenology and other environmental variables, such as soil moisture content, to refine the monitoring interval, Sentinel-1 data are analyzed by comparing backscattering values between the landslide area and a surrounding buffer zone. This refinement process further allows for the identification of signals consistent with landslide reactivation. The specific objective of the study is therefore twofold:

(1)

To provide an objective method able to identify the date of recent landslides in vegetated areas by enhancing objective timing compared to other approaches.

(2)

To offer a freely available, fully operational tool implemented on GEE.

2. Materials and Methods

In our approach, data processing is conducted automatically via a JavaScript script on GEE v1.6.3. GEE is a cloud-based platform developed by Google LLC that allows users to create, save, and share scripts in JavaScript or Python; GEE offers an up-to-date catalog of satellite data, including Sentinel and Landsat, and climate, hydrological, geophysical, and socioeconomic datasets. Users are afforded the capability to navigate and select datasets directly through the interface, thereby obviating the need for manual acquisition of extensive image collections. This ensures shorter processing times for both wide spatial coverage and longtime intervals, such as extended satellite time series. Thanks to the full shareability of scripts in GEE, it is possible to access the original code and execute the same analytical sequence.

The methodology is based on the integration of optical (Sentinel-2) and Synthetic Aperture Radar (SAR; Sentinel-1) satellite data, according to a sequential and conditional logic, as represented in the flow chart in Figure 1.

The process begins with the analysis of optical images using the NDVI to identify the PI. This temporal interval is defined by the two temporal extremes at which a change in vegetation cover is observed within the landslide area: the first extreme corresponds to the last date when vegetation cover was present, while the second marks the first date when vegetation was reduced or is absent.

Once the PI has been identified, the next step depends on the availability of Sentinel-1 radar images within this interval. When radar acquisitions are unavailable, the PI is directly adopted as the Estimated Interval (EI), as an estimate of the interval in which the landslide is assumed to have occurred. This choice is justified by the fact that the PI is sufficiently narrow (so much so that no radar acquisitions are available within it) and therefore represents the narrowest time frame available to estimate the date of the landslide.

If one or more Sentinel-1 radar images are present within the PI a backscattering analysis is performed. This step involves evaluating the ground radar response to detect signal variations that may be associated with landslide-related ground morphological changes. If the analysis detects no intersection, i.e., no dates where backscattering shows anomalies consistent with a landslide event, then the PI is considered to be the EI. The intersection refers to the comparison of the time series of VV and VH polarizations, both in ascending and descending geometry: the absence of consistent variation between these four signals indicates that there is insufficient radar evidence to narrow the time interval further.

Conversely, if the backscatter analysis shows an intersection between at least two time intervals, meaning two dates that delimit an interval of signal variation consistent with a potential landslide event, a new EI is defined based on the radar data.

2.1. Landslide Inventory

This study presents a global database of 46 landslides, selected for vegetated locations, occurrence between 2018 and 2024 to ensure satellite data availability on GEE, and a minimum size adequate for correct detection (~10 Sentinel-2 pixels). The database was compiled through a comprehensive review of scientific literature and publicly accessible online resources (

Table 1. Inventory of the analyzed landslide events. The columns show the ID, the geographic coordinates (latitude and longitude in decimal degrees, referring to the landslide centroid, WGS84 datum), the place name, the country, the landslide size in hectares, the date or estimated initiation period of the landslide, and the source of information (online reports or literature).

ID	Coordinates (Lat, Lon)	Place Name	Country	Size (ha)	Date	Source
A01	22.22842, 104.52953	Ba Thau	Vietnam	5.81	13 September 2024	[35]
A02	6.97045, 125.18599	Barangay Buhay	Philippines	20.13	31 October 2019	[36]
A03	18.38351, −73.93987	Beaumont	Haiti	6.13	14 August 2021	[37]
A04	43.48752, 17.80872	Bijela	Bosnia and Herzegovina	3.76	6 April 2024	[38]
A05	24.60476, 121.51879	Cilan	Taiwan	5.31	16 October 2022	[39]
A06	27.25203, 88.45953	Dipu Data	India	1.40	20 August 2024	[40]
A07	−33.88261, 19.09017	Franschhoek	South Africa	6.63	24 September 2023	[41]
A08	37.46992, 127.20293	Geombok	South Korea	1.12	9 August 2022	[42]
A09	35.19685, 127.13540	Gokseong	South Korea	3.02	7 August 2020	[43]
A10	64.78622, −22.10795	Hitardalur	Iceland	208.04	7 July 2018	[31]
A11	31.06474, 110.74863	Jiajiadian	China	5.09	17 July 2024	[44]
A12	26.26093, 104.67038	Jichang	China	27.45	23 July 2019	[45]
A13	−5.38042, 143.36475	Kaokalam	Papua New Guinea	11.80	24 May 2024	[46]
A14	11.41080, 76.23644	Kavalappara	India	19.66	8 August 2019	[47]
A15	11.48171, 76.14920	Kerala	India	66.67	30 July 2024	[48]
A16	22.12913, 104.53233	Lang Nu	Vietnam	70.12	10 September 2024	[49]
A17	14.64298, −90.52547	Las Calaveras	Guatemala	0.64	22 November 2023	[50]
A18	14.16362, −88.88318	Los Cabros	El Salvador	0.54	2 September 2022	[51]
A19	29.46262, 102.20946	Luding	China	3.82	5 September 2022	[52]
A20	−5.32677, 119.65777	Mangempa	Indonesia	1.39	23 January 2019	[53]
A21	−5.32654, 119.65342	Mangempa	Indonesia	0.50	23 January 2019	[53]
A22	−5.32629, 119.65509	Mangempa	Indonesia	0.59	23 January 2019	[53]
A23	−5.32432, 119.65599	Mangempa	Indonesia	0.39	23 January 2019	[53]
A24	18.41162, −73.75071	Maniche	Haiti	1.15	14 August 2021	[37]
A25	18.36872, −73.71781	Maniche	Haiti	0.77	14 August 2021	[37]
A26	39.93528, 15.74938	Maratea	Italy	1.18	30 November 2022	[54]
A27	7.39182, 126.02645	Masara	Philippines	8.55	6 February 2024	[55]
A28	7.82527, 98.31090	Nakkerd Hill	Thailand	1.24	23 August 2024	[56]
A29	10.16821, 77.01167	Pettimudi	India	2.91	6 August 2020	[57]
A30	16.43207, 107.30960	Phong	China	3.06	13 October 2020	[58]
A31	6.19571, −75.65686	San Antonio de Prado	Colombia	1.61	13 July 2022	[59]
A32	−23.75711, −45.71738	Sao Sebastiao	Brazil	0.45	19 February 2023	[60]
A33	−23.74848, −45.71918	Sao Sebastiao	Brazil	0.79	19 February 2023	[60]
A34	−23.75040, −45.73085	Sao Sebastiao	Brazil	0.76	19 February 2023	[60]
A35	−23.75768, −45.74419	Sao Sebastiao	Brazil	0.78	19 February 2023	[60]
A36	−23.76158, −45.69111	Sao Sebastiao	Brazil	2.21	19 February 2023	[60]
A37	−23.76068, −45.69233	Sao Sebastiao	Brazil	1.28	19 February 2023	[60]
A38	−23.75821, −45.69736	Sao Sebastiao	Brazil	1.70	19 February 2023	[60]
A39	−23.75382, −45.69472	Sao Sebastiao	Brazil	0.76	19 February 2023	[60]
A40	−23.75229, −45.69203	Sao Sebastiao	Brazil	0.18	19 February 2023	[60]
A41	−23.75283, −45.69326	Sao Sebastiao	Brazil	1.75	19 February 2023	[60]
A42	30.36681, 109.30083	Shaziba	China	22.00	21 July 2020	[61]
A43	42.70315, 43.67387	Shovi	Georgia	26.47	4 August 2023	[62]
A44	−39.56080, 175.26846	TeOreOre	New Zealand	2.65	3 October 2019	[63]
A45	35.80859, 110.65383	Zaoling	China	1.34	15 March 2019	[64]
A46	29.34340, 102.69556	Zhonghaicun	China	7.09	21 August 2020	[65]

and Figure 2).

Landslide boundaries were delineated by hand-tracing vector polygons in Google Earth Engine and then converted into a raster at 10 m resolution, defining a consistent reference system for all processing. The effectiveness of the method depends strictly on the choice of buffer size: buffers that are too small may not adequately represent the undisturbed reference area, while buffers that are too large risk including surrounding areas affected by other secondary processes or adjacent instability. To explore this sensitivity, the use of variable buffers (10, 20, 30, 40, and 50 m) was tested on a sample of landslides representative of the minimum, maximum, median, first, and third quartile sizes.

Table 2. Sensitivity test of the methodology as a function of buffer size (10–50 m) on five representative landslides: A40 (minimum), A10 (maximum), A36 (median), A08 (first quartile), and A07 (third quartile). The table reports the performance of Δσ⁰ (VH and VV polarizations) in correctly detecting the time interval containing the landslide (“yes”/”no”).

	A40		A10				A36		A08		A07
	VH-D	VV-D	VH-A	VH-D	VV-A	VV-D	VH-D	VV-D	VH-A	VV-A	VH-A	VV-A
10 m	no	yes	yes	yes	yes	yes	yes	no	Yes	yes	no	yes
20 m	no	no	yes	yes	yes	yes	yes	no	Yes	yes	no	yes
30 m	yes	yes	yes	yes	yes	yes	yes	yes	Yes	yes	yes	yes
40 m	yes	yes	yes	yes	yes	yes	yes	yes	Yes	yes	yes	yes
50 m	yes	yes	yes	yes	yes	yes	yes	yes	Yes	yes	yes	yes

summarizes the results obtained, indicating for each buffer and for each landslide whether the Δσ⁰ variation in VH or VV correctly detected time intervals containing the instability (‘yes’) or not (‘no’).

For very small buffers (10–20 m), some incorrect detections are observed in landslides of minimum or intermediate size, indicating that such small buffers may not adequately represent the undisturbed context. With buffers ≥30 m, the landslides in the sample are correctly detected. Therefore, the smallest buffer of 30 m was chosen, as it allows comparison with undisturbed areas not affected by the trigger but with characteristics similar to those of the landslides. A 30 m buffer was generated around each landslide, also rasterized to the same resolution to maintain data uniformity. For some landslide events, particularly those characterized by rapid vegetation regrowth or recent occurrences not yet represented in the available GEE imagery, it was necessary to supplement the research with an additional methodological step. In these cases, detailed Sentinel-2 images in true color composition (RGB) were used to examine the area. The polygons drawn with this methodology were then compared with Planetscope images, delivered at a spatial resolution of 3 m. This comparison enabled the refinement of the landslide boundaries, ensuring an accurate spatial representation even for events that were difficult to delineate, with an estimated uncertainty of ±10 m.

2.2. Processing of Sentinel-2 Data

Sentinel-2 is a constellation of satellites, Sentinel-2A and Sentinel-2B, launched by the European Space Agency (ESA), equipped with a multispectral imager (MSI) with decametric resolution: four bands at 10 m (band 2, blue; band 3, green; band 4, red; band 8, near-IR), six bands at 20 m (band 5, red-edge 1; band 6, red-edge 2; band 7, red-edge; band 8a, narrow near-IR; band 11, SWIR 1; band 12, SWIR 2) and three bands at 60 m (band 1, coastal aerosol; band 9, water vapor; band 10, cirrus). Sentinel-2A has been operational since June 2015, followed by Sentinel-2B in March 2017, allowing the constellation to capture images of the Earth’s surface every 5 days.

The time window was set between 2018 and 2024 to analyze vegetation changes associated with landslide events. This period was chosen because Sentinel-2 images were available on Google Earth Engine for all the selected landslides. The Sentinel-2 image collection was imported into the GEE platform, applying a preliminary filter based on the percentage of cloud cover, to include only acquisitions characterized by favorable atmospheric conditions. The cloud cover threshold was set at 50%; however, in the case of events of greater spatial extent, the high volume of data generated errors that exceeded GEE’s computational capacity. In such circumstances (6 cases out of 46), the threshold was progressively reduced until the maximum value compatible with the platform’s processing limits was identified.

A mask based on the Cloud Score Plus product was then applied, using the quality band “cs_cdf”, which gives each pixel a score between 0 (presence of cloud cover) and 1 (absence of cloud cover). To exclude data affected by atmospheric disturbances, a threshold of 0.60 was adopted, as suggested in the example script provided by GEE. Pixels with a score below this value were removed from the analysis, allowing only those portions of the image acquired under favorable atmospheric conditions to be retained. This procedure ensured an automated and objective selection of observations.

On each scene, exclusively for the pixels remaining after the application of the Cloud Score Plus mask, the NDVI was calculated for the landslide polygon and the respective buffer, using Equation (1):

$$\mathrm{NDVI}={\displaystyle \frac{{\mathrm{\rho }}_{\mathrm{NIR}}-{\mathrm{\rho }}_{\mathrm{RED}}}{{\mathrm{\rho }}_{\mathrm{NIR}}+{\mathrm{\rho }}_{\mathrm{RED}}}}$$

(1)

defined as the normalized ratio between the reflectance of the bands in the near infrared (Band 8) and red (Band 4), commonly used to quantify the state of vegetation. A landslide phenomenon over vegetated areas would result in a significant change in NDVI.

For each acquisition date, the mean NDVI values were initially computed. Subsequently, the difference between the average NDVI within the landslide polygon ($\overline{{\mathrm{NDVI}}_{\mathrm{L}}}$) and that of the surrounding buffer area buffer ($\overline{{\mathrm{NDVI}}_{\mathrm{B}}}$), referred to as ΔNDVI, was determined in accordance with Equation (2):

$$\Delta \mathrm{NDVI}=\overline{{\mathrm{NDVI}}_{\mathrm{L}}}-\overline{{\mathrm{NDVI}}_{\mathrm{B}}}$$

(2)

This configuration makes it possible to distinguish NDVI variations attributable to the event from those due to seasonal fluctuations, soil moisture dynamics, or other generalized environmental factors. Employing a buffer area as a methodological reference enables the definition of a control zone located near the landslide site, sharing comparable geological features, soil types, vegetation cover, and climatic conditions, yet remaining unaffected by the landslide itself. This configuration facilitates the discrimination of NDVI changes specifically linked to the landslide event from those caused by seasonal variability, soil moisture fluctuations, or other broad environmental influences.

Without this differential approach, relying solely on individually computed NDVI mean values for the landslide area could have led to misinterpretations, as such readings might not accurately reflect true vegetation dynamics due to atmospheric interference. In the present work, Sentinel-2 L2A (Bottom-of-Atmosphere, BOA) reflectance was used.

Next, an algorithm within the script was iterated over all values of ∆NDVI to calculate the absolute difference between each pair of consecutive dates. The adopted criterion defines the likely landslide occurrence interval as the period showing the greatest change due to the vegetation index drop from sudden cover loss. The two dates associated with the maximum deviation are assigned as estimates of the moments before (pre-event) and after (post-event) the movement, respectively. In this way, the procedure ensures an automatic, repeatable, and subjectivity-free identification of a PI (Figure 3).

2.3. Processing of Sentinel-1 Data

The Sentinel-1 acquisitions in Interferometric Wide (IW) mode are imported in both ascending orbit and descending orbit, relative to the PI. In cases where the acquisition is not performed on both orbits, the analysis is conducted on the only available orbit.

To avoid excluding portions of the relevant time window, an additional step was implemented to automatically include the earliest available acquisition preceding, as well as the one following, each of the previously selected dates, across both orbital paths, where applicable. This procedure ensures full radar coverage of both the pre- and post-failure phases.

Sentinel-1’s synthetic aperture radar (SAR) sensor, operating in C-band, ensures almost uninterrupted observations of the Earth’s surface regardless of weather conditions and sunlight. Thanks to its cloud-penetrating capability and the emission of radar waves both day and night, Sentinel-1 is an important resource for the continuous monitoring of areas with persistent cloud cover and for the study of surface dynamics in regions with poor optical visibility [66]. The Sentinel-1 collection available in GEE includes all Ground Range Detected (GRD) products from the Sentinel-1A and Sentinel-1B missions, obtained both in ascending and descending orbit.

Backscattering is a measure of the amount of radar signal (in the microwave spectrum) that, after being transmitted by the satellite toward the Earth’s surface, is reflected and detected by the sensor. This measurement, expressed as the backscattering coefficient (σ⁰), depends on surface properties such as roughness, moisture, and vegetation cover [34].

For each Sentinel-1 acquisition, σ⁰ is calculated for the landslide area and the buffer area, in the two polarization configurations: VV, corresponding to transmission and reception in vertical polarization (co-polarized), and VH, in which the signal is transmitted vertically and received horizontally (cross-polarized). The average backscattering values within the landslide perimeter ($\overline{{\mathrm{\sigma }}_{\mathrm{L}}^0}$) and in the surrounding buffer area ($\overline{{\mathrm{\sigma }}_{\mathrm{B}}^0}$) expressed in dB are therefore calculated.

For each of the four (or two) polarization–orbit combinations (ascending and descending VV and VH), the difference in average backscattering values between the landslide and the buffer for the same acquisition dates is calculated using Equation (3):

$$\Delta {\mathrm{\sigma }}^0=\overline{{\mathrm{\sigma }}_{\mathrm{L}}^0}-\overline{{\mathrm{\sigma }}_{\mathrm{B}}^0}$$

(3)

These differences make it possible to highlight landslide-induced anomalies; in particular, the triggering of a landslide changes the terrain surface roughness, causing sudden changes in the differential backscattering values. The difference in the average backscattering value of the buffer from that of the landslide reduces the background noise, increasing the robustness of the temporal dating of the landslide event and reducing the risk of false positives induced by variations unrelated to the phenomenon. Furthermore, this operation produces clear graphical representations in which the variations between mean values are visually sharp and interpretable compared to the analysis of raw signals.

Similarly to what is done for the NDVI, the script iterates over the time series of backscatter differences (∆σ⁰), calculating for each pair of consecutive dates the absolute variation of the mean values in dB. Among all the variations obtained, the one with the greatest amplitude is identified (one for each of the two or four datasets). The interval corresponding to the most pronounced change in the radar signal for each dataset of images (Temporal Interval VH polarization Ascending orbit, VH-A; Temporal Interval VV polarization Ascending orbit, VV-A; Temporal Interval VH polarization Descending orbit, VH-D; Temporal Interval VV polarization Descending orbit, VV-D) is then assumed as the interval of landslide collapse initiation, the Estimated Interval (EI) (Figure 4).

2.4. Temporal Delimitation

An automatic intersection of the obtained time intervals is performed by the developed script, which applies hierarchical selection criteria to estimate the triggering time span of the landslide event. Specifically, the method first searches for common dates across all four backscattering intervals; if none are found, it proceeds to search for common dates among three intervals, and if still none are found, it considers common dates between two intervals. This approach ensures an objective and reproducible procedure for the temporal identification of landslide events.

Since backscattering days are treated as boundaries of time intervals (consecutive intervals defined by start and end dates), it may happen that the intersection is calculated on dates coinciding with the edges of these intervals, producing many intersections on these very days. To avoid possible bias resulting from this overlap, the script systematically excludes the first day of each interval when calculating the intersection. Figure 5 illustrates the six possible scenarios of intersection between the time intervals obtained from the radar signals (VH and VV in ascending and descending orbit), in relation to the PI identified via NDVI.

Figure 5a shows the intersection of all four radar intervals, together with the PI. In this scenario, only dates common to all intervals are considered, allowing a reduction in the landslide triggering window. Figure 5b shows a case where the intersection involves three out of four radar intervals. The resulting interval includes only the dates shared by these three, completely excluding the non-overlapping interval. This configuration still allows for a narrower triggering window, albeit with a lower level of agreement among the intervals than in the previous case. Figure 5c illustrates a situation in which the intersection occurs between only two radar intervals. The temporal output thus reflects the minimum acceptable overlap for defining the event, leading to a less precise delimitation of the triggering period. Figure 5d illustrates a case where two distinct pairs of radar intervals show overlap, but there is no intersection between the two pairs; furthermore, the overlaps partially fall outside the PI. In this configuration, since no reliable intersection can be identified within the bounds of the PI, the entire PI is retained as the estimated trigger interval. Figure 5e shows a similar case, where two distinct pairs of radar intervals overlap. However, these overlaps contribute to narrowing the Preliminary Interval: the resulting interval is defined from the beginning of the first overlap to the end of the second overlap, resulting in a narrower time window than the PI. Finally, Figure 5f represents the case where no intersection between the available radar intervals is observed. In the absence of shared dates between the radar datasets, the entire PI is adopted as an estimate of the triggering interval, recognizing the limitations arising from the lack of convergence in radar signals.

2.5. Validation Strategy

To assess the effectiveness of the methodology developed for dating landslide events, a validation procedure is adopted.

The first step of validation is to check if the real date of the landslide is included in the EI. This step allows for an assessment of the method’s ability to include the triggering date of the event within the identified time window. Subsequently, the duration of the preliminary interval, expressed in number of days, is calculated. This parameter provides an initial estimate of the period when the landslide event might have occurred, based on changes in vegetation cover detected through Sentinel-2 data.

Then, for each landslide event, the number of Sentinel-1 images in the preliminary interval was counted, both in ascending and descending geometry, to estimate the effective temporal coverage of the radar dataset. Having identified the intervals corresponding to the maximum absolute difference in backscattering between consecutive dates, it was assessed how many of the four (or two) intervals overlapped, i.e., the presence of common dates between the different signals.

The intersection of the common dates made it possible to identify the EI, also calculated in days, which represents the restricted period obtained from the application of the proposed methodology. The percentage reduction between the PI and EI makes it possible to quantify the effectiveness of the methodology in refining the initial time interval.

Finally, the validation phase includes the calculation of the Hit Rate (HR), i.e., the success rate of the proposed methodology. This indicator is defined as the ratio between the number of landslide events for which the actual date of occurrence falls within the estimated time interval and the total number of events analyzed. The calculation of HR allows the predictive performance of the method to be assessed in terms of predictive ability. A high HR value indicates that the methodology can correctly identify the time window of the event.

3. Results

3.1. Application of the Dating Procedure to an Example Landslide

This section presents the results of a single landslide event selected as a representative case study, i.e., the Nakkerd Hill landslide (ID = A28; Table 1), which occurred on 23 August 2024 in Thailand, triggered by an extreme rainfall event [56]. The landslide, which can be classified as a debris flow, is characterized by an evident contrast in vegetation cover between the landslide area and the surrounding area.

Before 23 August 2024, the distribution of mean NDVI values shows a sinusoidal trend, except for a few outliers, reflecting the normal evolution of the vegetation cover during the year (Figure 6a).

The outliers, probably associated with residual cloud cover, can also be recognized in the NDVI trend observed in the buffer area (Figure 6b). Therefore, the presence of such outliers does not influence the dating of the landslide event, which is instead based on a clear difference in the NDVI value between the landslide polygon and its buffer.

Between the image of 16 August 2024 and the subsequent image with less than 50% cloud cover, dated 27 September 2024, there is a clear decrease in the average NDVI values in the landslide area, from 0.777 to 0.261. This decrease highlights a deterioration of the vegetation cover, presumably attributable to the landslide event and the consequent surface soil disturbances. In contrast, in the adjacent buffer area, the average NDVI values remain stable, suggesting the absence of direct impacts on vegetation in this area. This result confirms that the buffer area was not affected by the landslide event.

Figure 7a shows the overlap of the time series of mean NDVI values for the landslide area and the buffer zone.

The joint representation highlights two relevant aspects: (1) the presence of consistent outliers between the two series, suggesting that these deviations are related to cloud cover conditions, rather than to distinct local phenomena; (2) the sharp decline in NDVI values in the landslide area from 23 August 2024 onwards, in contrast to the trend without significant anomalies recorded in the buffer zone.

To further investigate the changes in vegetation cover and isolate the signal specifically associated with the landslide event, ΔNDVI was calculated for each available date (Figure 7b). This indicator is defined as the difference, for each observation date, between the average NDVI value detected within the landslide area and the average value obtained for the adjacent buffer area. The use of ΔNDVI serves to mitigate the influence of external factors that may affect both areas similarly (such as residual atmospheric variability), highlighting vegetation anomalies directly linked to the landslide event.

The GEE script developed for this article allows the proposed procedure for dating landslide events to be applied automatically. In particular, the script systematically analyses all available consecutive acquisitions and calculates, for each date, the variation of ΔNDVI between the landslide area and its buffer. The time interval showing the maximum absolute variation of ΔNDVI between two consecutive images is selected as the PI, with the first date identified as pre-event and the following as post-event. In the example landslide that occurred in Thailand, the script identifies the preliminary interval between 16 August 2024 (ΔNDVI = 0.009) and 27 September 2024 (ΔNDVI = −0.492). The reduction in ΔNDVI observed between these two dates shows a vegetative change in the landslide area compared to the buffer zone, consistent with the impact of a large landslide event.

To further investigate the dynamics of the event and refine the temporal delimitation, a radar backscattering analysis was subsequently conducted using Sentinel-1 data. All Sentinel-1 acquisition dates within the PI, i.e., between 16 August and 27 September 2024, were identified, including one prior and one subsequent acquisition to ensure full coverage of the period. For each selected date, the four average backscattering values (σ⁰) (ascending and descending VH, ascending and descending VV) were calculated for both the landslide area and the buffer zone. Subsequently, the difference between the mean values of the two areas was then computed to identify any variations consistent with changes in the surface structure caused by the landslide (

Table 3. Δσ⁰ values, defined as the difference between the mean σ⁰ of the landslide area and the buffer.

Date	$\mathbf{\Delta }{\sigma }_{VH\mathbf{-}A}^{\mathbf{0}}$	$\mathbf{\Delta }{\sigma }_{VV\mathbf{-}A}^{\mathbf{0}}$
11 August 2024	−1.21	−1.16
23 August 2024	−3.13	−3.93
4 September 2024	−3.33	−2.91
16 September 2024	−2.43	−2.93
28 September 2024	−3.73	−3.41
date	$\mathbf{\Delta }{\sigma }_{VH\mathbf{-}D}^{\mathbf{0}}$	$\mathbf{\Delta }{\sigma }_{VV\mathbf{-}D}^{\mathbf{0}}$
18 June 2024	1.24	1.81
17 August 2024	0.67	1.01
29 August 2024	−2.27	−0.44
10 September 2024	−1.75	−0.61
22 September 2024	−2.20	−1.38
4 October 2024	−1.97	−0.91

) (see Supplementary Materials, Table S1).

In line with the approach adopted for the analysis of ΔNDVI, automated processing is conducted on Δσ⁰ to identify the maximum absolute variation between consecutive acquisitions. This procedure enables the objective identification of the time interval characterized by the most pronounced discontinuity in the radar response of the ground, which can be associated with the occurrence of the landslide. The analysis was performed separately for each of the four combinations of polarization and orbit (ascending VV, ascending VH, descending VV, descending VH). In the case of the landslide selected as an example, the greatest variation in Δσ⁰ for both VV and VH polarizations was detected in the ascending orbit data between 11 and 23 August 2024, while in the descending orbit, the interval with the most significant discontinuity was between 17 and 29 August 2024.

The four intervals, for which the maximum backscattering variation was observed, are compared with each other and with the interval derived from the ΔNDVI (PI) analysis to determine the days on which the highest number of overlapping time ranges occurs.

Given that the PI is 16 August–27 September 2024, the two backscattering intervals in ascending orbit are 11–23 August 2024, and the two intervals in descending orbit are 17–29 August 2024; there is an overlap of all 5 intervals in the period of 18–23 August 2024. To avoid inaccuracies related to the inclusion of boundary dates in the backscattering intervals, as described in the Materials and Methods Section, the first day of each interval was excluded from the overlap calculation.

3.2. Assessment of Temporal Accuracy Across the Landslide Inventory

Table 4. Summary of analytical results for 46 landslide events.

ID	Contained (Yes/No)	NDVI Pre-Event	NDVI Post-Event	PI Amplitude	Asc Acquisitions	Desc Acquisitions	BS Overlap Count	EI Amplitude	Interval Reduction (%)	Hit Rate
A01	YES	0.799	0.283	11	0	0		11	0.00	0.98
A02	YES	0.841	0.209	26	2	2	4/4	8	69.23
A03	YES	0.726	0.158	6	0	0		6	0.00
A04	YES	0.268	0.132	13	1	1	3/4	7	46.15
A05	YES	0.841	0.151	41	3	4	4/4	12	70.73
A06	YES	0.603	0.114	16	0	2	2/2	12	25.00
A07	YES	0.484	0.176	21	2	0	2/2	6	71.43
A08	YES	0.818	0.639	81	12	0	2/2	12	85.19
A09	NO	0.797	0.257	31	2	3	4/4	6	80.65
A10	YES	0.478	0.101	28	8	5	4/4	3	89.29
A11	YES	0.686	0.375	21	1	0	2/2	9	57.14
A12	YES	0.554	0.173	46	4	4	3/4	10	78.26
A13	YES	0.661	0.216	6	0	0		6	0.00
A14	YES	0.841	0.175	111	0	22	2/2	7	93.69
A15	YES	0.717	0.197	311	0	49	2/2	7	97.75
A16	YES	0.748	0.178	11	0	0		11	0.00
A17	YES	0.789	0.285	31	3	2	4/4	4	87.10
A18	YES	0.822	0.424	19	1	0	2/2	8	57.89
A19	YES	0.779	0.159	41	4	6	4/4	5	87.80
A20	YES	0.825	0.158	51	9	4	4/4	6	88.24
A21	YES	0.839	0.378	225	27	14	3/4	6	97.33
A22	YES	0.822	0.505	51	9	4	4/4	6	88.24
A23	YES	0.789	0.333	51	9	4	4/4	6	88.24
A24	YES	0.723	0.300	21	3	3	2/4	9	57.14
A25	YES	0.650	0.178	6	0	0		6	0.00
A26	YES	0.543	0.186	8	1	0	4/4	4	50.00
A27	YES	0.683	0.247	31	0	2	2/2	8	74.19
A28	YES	0.777	0.261	43	3	4	4/4	6	86.05
A29	YES	0.807	0.241	86	0	7	2/2	12	86.05
A30	YES	0.823	0.227	81	17	9	3/4	3	96.30
A31	YES	0.678	0.161	16	0	2	2/2	12	25.00
A32	YES	0.867	0.435	11	0	1	2/2	10	9.09
A33	YES	0.817	0.450	11	0	1	2/2	10	9.09
A34	YES	0.765	0.419	31	0	3	2/2	12	61.29
A35	YES	0.805	0.343	11	0	1	2/2	10	9.09
A36	YES	0.885	0.319	11	0	1	2/2	10	9.09
A37	YES	0.885	0.432	11	0	1	2/2	10	9.09
A38	YES	0.877	0.343	11	0	1	2/2	10	9.09
A39	YES	0.848	0.451	11	0	1	2/2	10	9.09
A40	YES	0.869	0.478	11	0	1	2/2	10	9.09
A41	YES	0.862	0.261	11	0	1	2/2	10	9.09
A42	YES	0.772	0.100	41	2	0	2/2	12	70.73
A43	YES	0.735	0.108	21	4	4	4/4	6	71.43
A44	YES	0.731	0.287	13	3	1	4/4	7	46.15
A45	YES	0.428	0.127	3	0	0		3	0.00
A46	YES	0.761	0.232	56	8	5	3/4	5	91.07
		0.742	0.269	39.24	3.00	3.80		8.02	51.23	Mean
		0.784	0.244	21	1	1.5		8	59.59	Median
		0.268	0.100	3	0	0		3	0.00	Min
		0.885	0.639	311	27	49		12	97.75	Max

presents the results obtained for the 46 landslides selected for the application of our methodology across diverse geographical and environmental contexts.

The “Contained” column indicates whether the EI includes the trigger date of the event: the value is YES when the EI contains the actual date, or NO when the EI does not contain the trigger date.

The columns “NDVI pre-event” and “NDVI post-event” report, respectively, the mean NDVI values calculated over the landslide area in the acquisition immediately before and immediately after the triggering. These values result from considering only the images available after filtering by cloud cover percentage and applying the CloudScore+ mask. Analyzing all cases, NDVI pre-event values show a mean of 0.742 and a median of 0.784, with a minimum of 0.268 and a maximum of 0.885. For the NDVI post-event values, the mean decreases to 0.269 and the median to 0.244, with a minimum of 0.100 and a maximum of 0.639. The analysis of NDVI values associated with the landslides showed variations between the post-event and pre-event periods, with a median of 0.46 and an interquartile range of 0.17, indicating a marked change in vegetation parameters across the cases considered. This quantitatively confirms the loss of vegetation associated with the landslides triggering.

The “PI Amplitude” column shows the duration, expressed in days, of the PI, which represents the period when the greatest variation in ΔNDVI is observed between consecutive acquisitions. Values range from a minimum of 3 days to a maximum of 311 days, with a mean duration of approximately 39 days and a median of 21 days, reflecting the heterogeneity of spectral responses among the different landslide areas analyzed. The “Asc Acquisitions” and “Desc Acquisitions” columns indicate the number of Sentinel-1 acquisitions in ascending and descending orbit available within the PI. The number of acquisitions is proportional to the PI duration: longer intervals contain more images, while shorter intervals may include few or none. Indeed, in some cases, no acquisitions are recorded, either because of the absence of satellite passes over certain orbits at some landslides, or because the PI is too short to include acquisition dates. The “BS Overlap Count” column indicates the number of backscattering intervals between which temporal overlap is observed, i.e., the presence of common dates. In some cases, such as landslides A01, A03, A13, A16, A25 and A45, no overlap between backscattering intervals could be calculated because the PI is too short to include Sentinel-1 acquisitions. Consequently, for these landslides, the EI coincides with the PI. Regarding other landslides, the number of overlapping intervals varies for landslide A24, and there is an overlap between 2 out of 4 intervals, while for landslides A04, A12, A21, A30, and A46, there is an overlap between 3 out of 4 intervals. In 85% of the landslides with overlapping intervals, the overlap is observed across all analyzed intervals: 4 out of 4 when both ascending and descending orbit acquisitions are available, or 2 out of 2 when only a single orbit can be analyzed. The “EI Amplitude” column reports the amplitude, in days, of the EI, calculated by considering only the common dates resulting from the overlap between the different available backscattering intervals. “EI Amplitude” values range from a minimum of 3 to a maximum of 12 days, with a mean and median duration of 8 days. The “Interval Reduction (%)” column, calculated using Equation (4),

$$\mathrm{IR}\left(\%\right)=100-{\displaystyle \frac{{\mathrm{EI}}_{\mathrm{Amplitude}}\times 100}{{\mathrm{PI}}_{\mathrm{Amplitude}}}}$$

(4)

shows the percentage reduction in the duration of the PI compared to the EI, with a mean reduction of 51.23%, a median of 59.59%, a minimum of 0%, and a maximum of nearly 98%. High percentage values indicate that the EI has been significantly narrowed compared to the PI; small values suggest that the EI has a similar or equal duration to the PI.

Finally, the Hit Rate, defined as the ratio of the number of landslides for which the actual trigger date falls within the EI to the total number of landslides analyzed, was calculated. Out of 46 events considered, for 45 landslides, the trigger date falls within the EI identified by the proposed methodology (see Supplementary Materials, Table S2). The Hit Rate of 0.98 indicates the effectiveness of the methodology in accurately circumscribing the period of landslide triggering. The high Hit Rate confirms the robustness and adaptability of the method under varying conditions of data availability and landslide characteristics, reliably identifying a time window consistent with the observed event.

4. Discussion

This section discusses the results obtained and the limitations encountered from the application of the proposed methodology. The results were also compared with those in the literature, highlighting methodological differences from other studies that have used approaches similar to ours. Finally, some possible future developments are outlined, aiming to expand the application potential of the proposed method and apply it to other geomorphological contexts.

4.1. Evaluating the Results and Limitations

Analysis of the data in Table 3 shows that “PI Amplitude” ranges from a low of 3 to a high of 311 days. Low values indicate continuity in Sentinel-2 image acquisitions, with minimal or no cloud cover that does not affect observation. In contrast, high values reflect persistent cloud cover, which limits the availability of Sentinel-2 images. Cloud cover caused the PI to exceed 100 days in three cases (A14, A15, and A21). For A14, although the cloud cover threshold was set at 50%, this still resulted in the exclusion of many acquisitions, extending the PI; for A15, the threshold was reduced to 5% because of the large spatial extent of the landslide; and for A21, a 30% threshold was necessary due to the high volume of data causing errors in GEE. In these cases, the lack of optical data is compensated by more Sentinel-1 images. This allows us to observe that, for landslides A01, A03, A13, A16, A25 (Figure A1 in the Appendix A), Sentinel-2 images alone are sufficient to define a narrow time interval in which the landslide event occurs. Under favorable cloud cover conditions, the use of optical imagery is prioritized for identifying the date of triggering of landslide events, while under adverse weather conditions, integration with radar data becomes indispensable to obtain a time interval with precision of a few days. The overlap analysis of the time intervals derived from the maximum variation of $\Delta {\mathrm{\sigma }}^0$ shows some critical issues related to landslide area delineation. For landslide A04 (Figure A1 in the Appendix A), portions of soil devoid of vegetation before landslide initiation pose a problem in manually delineating the landslide area. The lack of vegetation makes it difficult to discriminate the bare soil caused by the landslide initiation from that present beforehand. Possible errors in delimiting the landslide area affect backscattering values, returning common dates only among 3 out of 4 backscattering intervals. In the case of A12, A21 and A30 (Figure A1 in the Appendix A), vegetative growth is observed after the landslide event (Figure 8).

Even in these cases, it is challenging to manually delineate the affected areas, despite the aid of RGB images acquired close to the estimated date of the events.

Specifically, we conducted a preliminary test on three landslides (A04, A21, and A30), selected respectively for the presence of bare soil prior to the landslide (A04), for the regrowth of vegetation after the event (A21), and for subsequent remediation work (A30). These landslides were deliberately delimited based solely on Sentinel-2’s true color composition, without resorting to either the acquisition of the GEE interface (available considerably later in cases b and c) or high-resolution images such as those from PlanetScope, which are useful for discriminating against the presence of bare soil prior to the landslide (case a). This test showed that true color composition images alone are not sufficient to guarantee accurate delimitation, while they are effective when used in combination with PlanetScope and GEE interface acquisitions (Figure 9).

In case a, this limitation led to the identification of an incorrect PI, consequently generating an incorrect EI; in case b, the poor delimitation produced the same PI but an incorrect EI; in case c, although slightly deviating from the correct delimitation, the EI remained unchanged.

The A24 landslide presents a peculiar situation: out of four identified time intervals, only two show common dates, indicating low temporal overlap. This discrepancy is probably due to the landslide area buffer touching adjacent land also affected by the landslide. The presence of multiple disturbed surfaces may alter the detected backscattering values, making it difficult to identify time intervals consistent with landslide initiation. Finally, the area of A46 undergoes subsequent ground settlement after the landslide event. This intervention may lead to suboptimal manual delineation of the landslide area, causing a source of error in the estimation of backscattering changes.

Regarding “EI Ampitude”, the results show time intervals ranging from a minimum of 3 to a maximum of 12 days. The shortest intervals (close to or less than 6 days) were recorded in the presence of acquisitions from both Sentinel-1 orbits (ascending and descending), with acquisition dates well-spaced between the two orbits. In contrast, longer intervals (close to or equal to 12 days) are associated with scenes from only one orbit, or with cases where, although acquisitions from both orbits are available, the dates are very close. These observations highlight how the combination of orbits and the temporal distance between acquisitions directly influences the temporal resolution of the final interval obtained from the methodology. At present, on Google Earth Engine (GEE), Sentinel-1 remains the most convenient SAR dataset for continuous monitoring due to its open access, frequent revisit time, and global coverage. However, we do not exclude the potential integration of future SAR missions, which could offer improved spatial and temporal resolution compared to Sentinel-1. Once these missions’ data become available on GEE, they could enhance our ability to monitor and date landslide events, as well as other rapid environmental changes.

As a result, “Interval Reduction (%)”, which measures as a percentage how much the radar data allows the time interval initially defined with the optical data (PI) to be reduced, also shows strong variability among the different cases analyzed. When the Interval Reduction % is low, the contribution of Sentinel-1 data allows only a slight refinement of PI, reducing it by a few days. This may be due to both the good initial quality of the optical data, which already provides accurate indication, and the limited availability of well-distributed radar acquisitions over time. High Interval Reduction (%) values suggest that integrating Sentinel-1 radar data significantly sharpens the temporal accuracy of landslide assessments. This variability highlights the importance of integrating optical data with radar data and underscores how the efficiency of the radar data in supporting temporal analysis is closely related to the frequency and distribution of acquisitions, as well as the observation conditions of the optical sensors.

Regarding the overall performance of the methodology, the results show a Hit Rate of 0.98, with 45 landslides correctly dated out of a total of 46. This high success rate confirms the effectiveness of the proposed method in identifying a time interval containing the actual triggering date of landslides that occur on slopes densely covered with vegetation.

The only case in which the methodology did not correctly identify the interval was landslide A09. In this case, although there are 4 out of 4 backscattering intervals with common dates, the EI does not include the actual date of triggering. The reason lies in the fact that the trigger date coincides with the first date of two backscattering intervals, which, by script setting, is systematically excluded from all intervals to avoid false positives due to the nature of the interval extremes. This choice stems from the need to keep the process fully automatic. In an operational context, it is necessary to prevent dates placed at the extremes of the interval from being considered as potential landslide dates. The exclusion of the first day of each interval was required, as not doing so would have caused errors in other cases. This is because the intervals are consecutive: the same day would have been included both as the last day of one interval and as the first day of the next, creating potential overlaps and false positives. However, this rule introduces a possible methodological limitation, especially in those cases where the landslide occurs near the first acquisition of the intervals found by the backscattering analysis.

The scenarios of the other investigated landslides are presented in the Supplementary Materials (Figures S1–S6).

4.2. Comparison with the Literature

Other authors (see [25]) pursue a similar goal to ours, namely, the dating of recent landslides through the combined use of Sentinel-2 and Sentinel-1 data. The use of atmospherically uncorrected Sentinel-2, level 1C images causes obvious disturbances due to aerosols, water vapor, and clouds. In our study, the use of atmospherically corrected Sentinel-2 level 2A images proves to be particularly efficient in correctly localizing the real vegetation loss, reducing the risk of misinterpreting NDVI decreases due to cloud cover. For Sentinel-1 data, the authors of [25] select, for each area of interest, a single “dominant” orbit characterized by the highest average backscattering values over the observation period. Although this choice is an effective solution aimed at reducing the variability of the radar signal, the exclusion of an orbit may result in the loss of useful information. In our approach, using both orbits allows further refinement of landslide date estimation due to higher temporal density. In addition, in [25], some estimated time intervals do not coincide with the actual date of landslide initiation, and a definitive interval containing the event cannot be identified. In contrast, in our approach, even in the few cases where the results derived are inconsistent with the days of landslide collapse, we were able to determine a final time interval that includes the actual date of the event. This results in a high percentage of correct dates, providing a useful method even in cases where not all backscattering changes coincide with the landslide trigger.

Compared with the work [67], their dates inconsistent with landslide initiation are due to the presence of other phenomena that cause ground deformation, generating both increases and decreases in the backscattering coefficient. This is because relatively long-time intervals are chosen in their approach, which increases the risk of including variations in the radar signal not directly related to the landslide event. In our approach, on the other hand, by restricting the backscattering analysis exclusively to the period when a change in vegetation is detected with Sentinel-2, we can automatically identify the interval in which the change in radar signal due to landslide initiation occurs, focusing on when the change occurs rather than how it manifests. This allows us to focus the analysis on only the time interval at which the event occurs, thus reducing any ambiguity associated with the heterogeneous nature of radar responses produced by other factors.

Similar methodologies are commonly employed for both spatial and temporal detection of landscape changes. For instance, the authors of [19] focus on automatically identifying seismic-induced landscape alterations by analyzing vegetation variations with Sentinel-2 data. Their approach relies on the Atmospherically Resistant Vegetation Index (ARVI) and involves visually setting a vegetation index threshold to classify bare areas. While this method enables a qualitative distinction between vegetative cover and bare ground, it introduces subjectivity through manual threshold definition.

A comparable strategy is seen in [29], where an operator determines NDVI variation thresholds based on visual interpretation of time series data. In contrast, our study adopts a fully automated procedure for assessing vegetation loss via optical satellite imagery, eliminating subjective decisions and arbitrary thresholding. Additionally, [29] applies a stringent cloud-filtering criterion, selecting images with a maximum cloud cover of 5%.

Instead, in our study, we apply a more flexible threshold, initially at 50%, which is progressively and manually lowered only for larger landslide areas to avoid errors due to exceeding the computational capacity of GEE. Finally, while [25,29] use a distributed workflow across different environments (software and online), the proposed approach can be applied using a single script that needs to be launched within the GEE platform, taking advantage of the ability to handle large volumes of data online in an efficient and replicable manner.

In summary, compared to previous studies, our study differs in parts of methodology and applications. Unlike [25], our method uses Sentinel-2 Level 2A data and integrates both radar orbits, ensuring better localization of vegetation loss and greater temporal density in identifying the date of the landslide. Compared to [67], which uses wide time intervals with the risk of including deformations unrelated to the landslide event, our study limits the radar analysis to the period of vegetation change detected automatically by Sentinel-2, reducing ambiguities due to external phenomena. Compared to [19,29], which are based on manually defined vegetation index thresholds and are therefore subject to arbitrary interpretations, our approach adopts a fully automated procedure for assessing vegetation cover loss, eliminating subjectivity. Finally, while [29] applies a very restrictive criterion for cloud cover, our method uses a more flexible threshold, reduced only in cases of large-scale landslides, and ensures a unique and efficient workflow that can be managed entirely by GEE, without the need for integration with external environments.

In conclusion, our approach represents an improvement on previous work: the procedure is fully automated, reproducible, and free from subjectivity, reducing issues related to arbitrary choices and ensuring results that can be extended to different scenarios.

4.3. Perspectives for Future Research

The inventory used in this study includes landslides of varying sizes and types, providing a diverse range of scenarios, although only over vegetated areas. The applicability of the method to vegetated slopes is limited by dependence on vegetation variations, which excludes sparsely vegetated, bare, or rocky terrain. To evaluate the method’s applicability in such contexts, future studies will focus on extending the methodology to these environments, integrating alternative optical indices, such as those based on soil moisture, to constrain the time interval of landslide occurrence even in the absence of vegetation cover. This approach is particularly useful for studying landslides where variation in NDVI alone may not clearly circumscribe the event. In addition, this could reduce the ambiguity arising from the possible exclusive use of radar backscattering, which, as pointed out by [67], can be influenced by other factors unrelated to the landslide event, and consequently provide a noisy signal.

In the present study, knowing the location and boundaries of landslides, we try to temporally identify events to accurately determine a trigger interval. However, a possible future research direction could be to automate a similar process to identify and recognize the spatial limits of landslide areas.

This can be achieved by comparing average pixel values from optical and radar imagery, leveraging both pre- and post-event acquisitions, to automatically detect regions of vegetation cover and morphological changes. By trying to make the spatial recognition process as objective and automatic as the temporal one, the limitations highlighted by [19,29] could be overcome. This spatial analysis approach enables efficient mapping of landslide extents while simultaneously integrating temporal detection with automated, objective land classification, thereby broadening the methodological applicability.

5. Conclusions

In this study, an automated methodology was developed and tested for dating recent landslide events in areas characterized by the presence of vegetation cover. This procedure enables the identification of a narrow temporal window, spanning just a few days, within which the landslide initiation date can be accurately constrained.

The methodology enables overcoming the main limitations associated with the use of optical satellite data, such as persistent cloud cover and the consequent scarcity of available observations, which are an obstacle to the continuous monitoring of landslide events. This critical issue is overcome through the integration of SAR data, which, being independent of atmospheric conditions, ensures temporal continuity of observations. This synergy between optical and radar acquisitions enhances the ability to date landslide events even in complex climatic contexts.

The procedure leverages the cloud-based Google Earth Engine platform, which enables seamless online access and processing of satellite imagery. This eliminates the need to download large datasets locally, thereby enhancing analytical efficiency and significantly reducing processing time.

Due to the ability to easily share the developed script, the methodology is easily replicable and applicable to different landslide areas characterized by vegetation cover, making the approach versatile to multiple spatial contexts. Furthermore, the intuitive interface ensures accessibility for users with limited technical expertise, promoting the widespread adoption of this technique as an effective tool for land monitoring.

A distinctive feature of this methodology is its full automation: the entire workflow, from the processing of optical and radar satellite data to the determination of a probable time interval, is executed autonomously, without requiring operator intervention or manual evaluation. This feature ensures greater objectivity of results, reducing the risk of user-caused errors.

The results from applying the methodology to the constructed inventory highlight the correct functionality of the proposed approach. Specifically, the process allows the estimated time interval of occurrence to be identified for about 98% of the analyzed landslides, demonstrating a high degree of robustness of the methodology even when applied to different types of landslides. Thus, the proposed procedure emerges as an effective tool to support landslide risk monitoring and analysis activities, helping to improve the timeliness and accuracy of event dating. However, the methodology is applied to landslides occurring in vegetated areas and larger than approximately 10 Sentinel-2 pixels, with potential difficulties in delineating landslide boundaries due to vegetation regrowth also affecting the sensitivity of the buffer.

Looking forward, a methodological extension could involve the development of automatic procedures not only for dating but also for the spatial recognition of landslide bodies by integrating multispectral data, radar, and digital terrain models to delineate the area affected by landslide phenomena.