Hydro-Meteorological Landslide Inventory for Sustainable Urban Management in a Coastal Region of Brazil

Abstract

Comprehensive, standardised, multi-temporal inventories of rainfall-induced landslides linked to soil moisture remain scarce, especially in tropical regions. Addressing this gap, we present a multi-source urban inventory for Brazil’s Baixada Santista region (1988–2024). A key advance is the introduction of geographical and temporal confidence classifications, which indicates precisely how each landslide’s location and occurrence date are known, thereby addressing a previously overlooked criterion in Brazil’s landslide data treatment. The inventory comprises 2534 records categorised by spatial (G1–G3) and temporal (T1–T3) confidence. Notable findings include the following: (i) confidence classifications enhance inventory reliability for research and early warning, though precise temporal data remains challenging; (ii) multi-source integration with UAV validation is key to robust inventories in urban tropical regions; (iii) soil moisture complements rainfall-based warnings, but requires local calibration for satellite-derived estimates; (iv) data gaps and biases underscore the need for standardised landslide documentation; and (v) the framework is transferable, providing a scalable model for Brazil and worldwide. Despite limitations, the inventory provides a foundation for (i) susceptibility and hazard modelling; (ii) empirical thresholds for early warning; and (iii) climate-related trend analyses. Overall, the framework offers a sustainable, practical, transferable method for worldwide and contributes to strengthening disaster information systems and early warning capacities.

1. Introduction

Landslides represent a significant hazard in mountainous and densely urbanised regions worldwide, causing substantial human, economic, and environmental losses each year [1]. Globally, rainfall is recognised as the predominant trigger of shallow landslides, particularly in tropical and subtropical environments where intense precipitation events frequently occur [2,3]. Climate change is expected to exacerbate these risks by altering precipitation regimes, increasing the frequency of short-duration, high-intensity rainfall events, and thereby intensifying slope instability hazards in vulnerable regions [4]. In this context, effective landslide risk management increasingly depends on highly accurate, spatially and temporally resolved inventories for the development of reliable landslide early warning systems (LEWSs) [5,6].

A cornerstone of effective LEWSs is the availability of a detailed, reliable, and spatially explicit landslide inventory—a systematic record of past events including location, timing, type, trigger, and relevant physical attributes [7]. Methodological guidelines emphasise the need for consistent classification, high spatial accuracy, and precise temporal data [3,8,9,10]. Recent advances have leveraged high-resolution remote sensing, semi-automated mapping, and integration with rainfall datasets to enhance completeness and operational utility [11,12,13,14]. In parallel, studies have expanded the use of empirical hazard-assessment models [6,15,16] and underscored soil moisture as a critical landslide-triggering factor [17,18,19,20,21]. However, as noted by [7], only a few operational LEWSs worldwide integrate soil moisture observations alongside rainfall, yet where implemented, this combination improves forecasting skill and reduces the number of false alarms [7,22,23].

Despite several international initiatives producing multi-temporal landslide inventories linked to rainfall data [24,25,26], there remains a notable absence of publicly available catalogues that systematically incorporate soil moisture conditions for rainfall-induced events. The inclusion of antecedent soil moisture conditions, increasingly recognised as a critical factor for improving landslide prediction [21,27,28,29], has yet to be routinely incorporated into inventory structures, particularly in tropical regions like Brazil.

In the Brazilian context, recent studies have begun to address event-specific compilations focused on space [30,31,32,33,34,35] and even the inventory quality’s impact on landslide susceptibility maps [12], but neglecting temporal variables. However, comprehensive and publicly accessible catalogues that explicitly integrate rainfall and soil moisture conditions remain limited, particularly for high-risk urban areas. The Baixada Santista region, located on the southern coast of São Paulo State (Brazil), provides a pertinent example of this need. The area has experienced a long history of rainfall-triggered landslide disasters, including major events in Santos (1928), across the region in the 1950s and 1970s, and major debris flows in Cubatão (1985) that affected industrial infrastructure [36]. Recently, rainfall-induced landslides in 2013, the fatal event of March 2020, and the severe 2023 São Sebastião carnival event highlight the growing hazards associated with extreme precipitation under changing climatic conditions [37,38]. Alongside these headline events, scattered, recurrent landslides are routinely recorded by municipal and state civil defence agencies, yet these remain largely under-compiled and poorly integrated with hydro-meteorological data.

As demonstrated by successful LEWSs internationally [39,40,41], systematically recording landslide occurrences linked to both rainfall and soil moisture conditions is essential for improving risk understanding, threshold derivation, and for tracking the influence of climate variability on landslide activity [4,7].

This paper addresses this data gap by developing a high-resolution, hydro-meteorological landslide inventory for the Baixada Santista region. Specifically, we aim to shed light on the research question: Can a multi-source inventory integrating rainfall and soil moisture improve data reliability for landslide hazard modelling and early warning in data-scarce tropical regions? In response to this, we hypothesise that explicitly incorporating antecedent soil moisture alongside rainfall will provide more reliable triggering condition datasets than rainfall-only inventories. The inventory integrates event records from civil defence reports, news archives, scientific publications, and satellite-based validations with rainfall and soil moisture measurements from both ground-based and remote-sensing platforms, using a spatial and temporal confidence classification adapted from established international catalogues [3,24,42]. The research design comprised five interconnected steps: (i) the compilation of a harmonised, multi-source landslide inventory; (ii) the assignment of locational and temporal confidence levels to each event, following internationally recognised frameworks; (iii) the linkage of each landslide record to corresponding rainfall and soil moisture measurements; (iv) descriptive analysis of temporal trends, spatial patterns, and associated triggering conditions; and (v) discussing the inventory’s operational and scientific value for early warning systems, climate-related monitoring, and data-driven susceptibility and hazard modelling. Although centred on Baixada Santista, the methodological framework is transferable to other data-scarce, landslide-prone settings, offering a replicable approach for advancing empirical risk models and operational LEWSs.

2. Materials and Methods

This study adopted a multi-step methodological framework to compile and analyse a comprehensive inventory of rainfall-induced landslides in the Baixada Santista region. The approach integrated diverse data sources, including landslide records, hydro-meteorological variables, and geoenvironmental datasets. The methodology encompassed data preprocessing steps such as cleaning, validation, and the classification of events based on spatial and temporal confidence levels. Additionally, soil moisture data were incorporated to assess their operational relevance for landslide hazard assessment. All datasets were spatially integrated using Geographic Information System (GIS) techniques.

The input data consisted of four primary categories: landslide occurrences, rainfall records, soil attributes, and geoenvironmental parameters. These datasets were preprocessed to support the development of the landslide inventory, subsequent analyses, and the implementation of a web-based application. As a final product, two inventory structures were generated: the Landslide Inventory Date-based (LID) and the Landslide Inventory Time-based (LIT) structures, both of which were incorporated into the application prototype. The overall methodological workflow is illustrated in Figure 1 and further elaborated in the following sections.

2.1. Data and Test Site

This paper focuses on the Baixada Santista region, located on the southeastern coast of São Paulo state (Brazil). The test site encompasses the municipalities of Santos, São Vicente, Cubatão, and Guarujá, with an area of 716.51 km², collectively forming one of the country’s most socioeconomically important coastal zones. The region holds strategic value due to the Port of Santos, the largest in Latin America, and hosts a dense industrial and petrochemical complex in Cubatão [36,43].

Geographically, Baixada Santista lies between the Serra do Mar mountain range and the Atlantic Ocean, with urban settlements extending onto coastal plains and steep hillslopes (Figure 2). The region’s geology is dominated by Precambrian crystalline rocks overlain by thick colluvial and residual soils, particularly along the Serra do Mar escarpment, where landslide susceptibility is high [44,45,46].

The climate is classified as humid subtropical (Cfa), with an average annual rainfall exceeding 2000 mm, concentrated mainly between December and March [47]. The combination of intense rainfall events, steep terrain, and rapid urbanisation has historically triggered numerous destructive landslides, particularly affecting informal settlements on hillslopes [48].

The intense urban pressure and ongoing occupation of landslide-prone zones, despite risk management policies, continue to pose significant challenges for local authorities. Baixada Santista’s long history of rainfall-induced landslides has made it a priority region for hazard monitoring and early warning initiatives in Brazil.

2.1.1. Landslide Data Gathering

An accurate and comprehensive landslide inventory is a critical component for hydro-meteorological hazard assessment and early warning system development [3,7]. In the absence of an official or standardised regional database for the Baixada Santista region, a multi-source data gathering strategy was adopted to assemble a detailed catalogue of rainfall-induced landslide events. This process involved the following key tasks:

• The systematic collection and mapping of landslide occurrences from institutional and local civil defence reports, verified news articles, and previous research, all covering the period from 1988 to 2025 for the cities of Cubatão, Guarujá, Santos, and São Vicente. These records provided fundamental information including date, location, and descriptive accounts of each event (

  Table 1. Landslide multi-source information gathered from different institutions.

  Source	Date Range	Observations
  São Paulo State Geological Institute (IG)	1990–2018	Data from the whole state, includes news reporting, local civil defence data, and more
  Santos Municipal Civil Defence	2014–2023	Local data, occurrences recorded when called upon by society
  Cubatão Municipal Civil Defence	1998–2024	Local data, occurrences recorded when called upon by society
  Gathered by the author from news reporting	Scattered	Information completion conducted by the author according to the news, such as landslide time/period

  );
• Field validation that was conducted through targeted site visits and UAV-based assessments to verify event locations, characterise slope conditions, and fill data gaps. Ten areas were selected based on accessibility and visited across the study region, including scars from the 2020 event. Additionally, validation in Cubatão leveraged prior fieldwork from [46], while the consistent UAV-based assessments conducted by the Civil Defence of Santos provided high-accuracy data for recent events, significantly reducing the need for extensive field surveys;
• The spatial verification and adjustment of event locations using QGIS, based on address records, fieldwork validation, and the interpretation of high-resolution satellite imagery via PlanetScope and Google Earth Pro’s time series. The adjustment was performed on a 1:25.000 map scale.

For each recorded landslide, the inventory included the following information: (i) date and time or time range of occurrence, when available; (ii) precise or approximate geographic coordinates and location; and (iii) landslide type, if identified.

2.1.2. Rainfall and Soil Moisture Network

To characterise the hydro-meteorological conditions preceding each landslide event, multiple rainfall and soil moisture datasets were compiled, considering both their temporal resolution and availability relative to the landslide inventory’s timeline. Recognising that seasonal fluctuations in soil moisture are amongst the principal factors triggering landslides [4,49], this study integrates soil moisture conditions into the landslide inventory—a feature rarely incorporated in inventories of this nature.

Rainfall data were sourced from a combination of ground-based and satellite-based systems to provide maximum spatial and temporal coverage. Given the limited coverage of ground-based monitoring, especially for earlier events and peri-urban areas, satellite-based products such as Rainfall Estimates from Rain Gauge and Satellite Observations (CHIRPS) and Integrated Multi-satellitE Retrievals for GPM (IMERG) were integrated to supplement rainfall and soil moisture records, filling gaps in historical gauge data and providing estimates where ground measurements were unavailable.

Similarly, ECMWF Reanalysis 5-Land (ERA5-Land) soil moisture data were utilised to provide spatially and temporally consistent soil moisture information for periods and locations not covered by the in situ geosensor network, which only became operational from 2019 onwards. Whilst reanalysis products lack the precision of local measurements, they offer suitable temporal and spatial coverage for assessing broader antecedent soil moisture conditions within the study area. The complete set of hydro-meteorological datasets utilised in this study is summarised in

Table 2. Hydro-meteorological datasets used in this study.

Data Type	Dataset	Time Range	Temporal Resolution	Source
Rainfall	Manual rain gauge	1940–present	Daily	ANA 1
	Manual rain gauge	1940–present	Every 3 hrs	SABESP 2
	Automatic rain gauge	2015–present	10 min	CEMADEN 3
	CHIRPS satellite estimate	1981–present	Daily	Climate Hazards Center
	IMERG satellite estimate	2000–present	30 min	NASA 4
Soil moisture	Automatic in situ geosensor network	2019–present	10 min	CEMADEN
	Era5-Land reanalysis	1981–present	Hourly	Copernicus

¹ National Water Agency. ² State of São Paulo Sanitation Company. ³ State of São Paulo Sanitation Company. ⁴ National Aeronautics and Space Administration.

.

This multi-resolution, multi-source integration strategy ensured that each landslide event was contextualised within its hydro-meteorological environment as accurately as possible, accommodating the inherent uncertainties and temporal inconsistencies present within historical datasets.

2.1.3. Geoenvironmental Parameters

To characterise the physical context of each landslide occurrence and further support the establishment of hydro-meteorological thresholds, a set of geoenvironmental and physical parameters were integrated into the inventory. These parameters aimed to capture the influence of terrain morphology, geological setting, and hydrological units on landslide-triggering mechanisms. The following datasets and procedures were considered:

• Watersheds and slope units: Watershed boundaries and slope units were derived from a 5 m resolution contour lines from a Digital Elevation Model (DEM) produced using data from the Brazilian Geological Survey database using the r.watershed algorithm in GRASS GIS, providing a detailed representation of topographic variability across the study area. The delineation allowed for the assignment of each landslide to its respective hydrological unit and slope compartment;
• Slope type classification: Slopes were preliminarily classified as “unmodified” or “anthropogenically modified” based on human-induced alterations, using field observations, urban mapping, and high-resolution imagery interpretation. This classification is crucial for differentiating between areas affected by direct anthropogenic interventions and those retaining their natural characteristics, especially in densely urbanised regions where human-induced changes are a key landslide trigger, such as in Hong Kong and Chongqing [16,41,50];
• Lithological map: A lithological map was integrated into the inventory, associating each event with its geological substrate and soil behaviour related to different rainfall and soil moisture thresholds. Lithological information was available at different cartographic scales: 1:50,000 [44] for the municipalities of Cubatão, Santos, and São Vicente, and 1:250,000 [51] for Guarujá. The difference in map scale was considered a source of spatial uncertainty in the lithology parameter attribution;
• Land use and land cover map: Obtained from MapBiomas Collection 9 (2023) [52], which provides annual, multi-temporal, and 30 m resolution land cover and land use classifications across Brazil derived from Landsat imagery. It enabled the assessment of anthropogenic modifications and land cover patterns in the study area.

All parameters were spatially joined to the landslide inventory using GIS techniques in QGIS version 3.34 and Python programming language version 3.10, and uncertainties in parameter attribution were carefully noted, especially in areas where data resolution or the cartographic scale varied.

2.2. Inventory Compilation

2.2.1. Data Preprocessing

The multi-source landslide, rainfall, and soil moisture datasets were filtered to enable integrated analysis. Due to incomplete records and the absence of consistent spatial or temporal reporting standards, events lacking date, address, or neighbourhood information were excluded from the inventory compilation. Moreover, cross-referencing checks were performed to verify and reduce duplicated data shared amongst the sources.

Given the notable limitations inherent to multi-source data integration in densely populated urban and peri-urban environments, an uncertainty classification system was developed to quantify both the geographical precision and temporal accuracy of each landslide record. This classification was adapted from mapping standards proposed by [24,42], more suitable to larger-scale mapping. We adjusted them to suit the finer spatial context of urban environments in Baixada Santista.

1.

Geographic confidence classification

Landslide records were first georeferenced as points, and each was assigned a geographic confidence class based on the locational precision and reliability of the source information (

Table 3. Landslide geographic confidence classification, adapted from [42].

Geographic Confidence Class	Code	Description
High	G1	Known point; precisely mapped via high-resolution imagery, remote sensing, GPS, field survey, or precise report
Medium	G2	Approximate location from address and number of the house or geocoding
Low	G3	Estimated based on indirect info from any other source

). Landslides’ coordinates were then validated and, where necessary, adjusted through address verification, field campaigns, and the inspection of high-resolution time-series satellite imagery via Google Earth Pro.

2.

Temporal confidence classification

The temporal confidence class indicated the precision of each record’s time of occurrence and was categorised into three classes, as shown in

Table 4. Landslide temporal confidence classification, adapted from [24].

Temporal Confidence Class	Code	Description
High	T1	Precise time of occurrence available (recorded hour and minutes)
Medium	T2	Approximate period of the day available (morning, afternoon, night, dawn) 1
Low	T3	Date is known or approximate date information available

¹ For T2 cases, landslide occurrence times were assigned to the start of the reported time period as follows—dawn: 00:00–5:59, morning: 06:00–13:59, afternoon: 14:00–18:59, and night: 19:00–23:59.

. This classification reflects the varied nature of the documentary and observational sources available.

Both geographic and temporal classification enabled subsequent analyses to incorporate both spatial and temporal uncertainty, ensuring appropriate consideration of data quality and reliability throughout the threshold analysis and early warning model development stages.

For T1 events, where the exact time was available, rainfall and soil moisture values were extracted for defined antecedent periods preceding the recorded occurrence time. For events classified as T2 and T3, where precise times were unavailable, occurrence times were standardised according to locally adopted conventions for periods of the day.

Table 5. Temporal classification framework for T2, T3, and date events based on reported time periods.

Category	Time Range	Assigned Time for Events Without Known Time
Dawn	00:00–05:59	Events reported as occurring at dawn were assigned 00:00
Morning	06:00–13:59	Events reported as occurring in the morning were assigned 06:00
Afternoon	14:00–18:59	Events reported as occurring in the afternoon were assigned 14:00
Night	19:00–23:59	Events reported as occurring at night were assigned 19:00
Date only	YYYY-MM-DD	Events reported with a specific date but no time information were assigned to the day before

presents the time assignments applied for these categories based on typical usage in Brazilian reporting contexts.

Data harmonisation included the removal of implausible soil moisture values (e.g., negative values), aligning datasets to common temporal resolutions and gap-filling using satellite-derived ERA5-Land soil moisture reanalysis, as described below.

3.

Slope modification classification criteria

Landslide points were further classified based on land cover (MapBiomas Collection 9), slope (>5°), and proximity to urban infrastructure. A 20 m buffer around built-up areas was employed as a proxy for anthropogenic influence, in combination with a visual assessment of vegetation removal or terrain reshaping. Based on these criteria, slopes were categorised as either “anthropogenically modified” (evidence of human intervention such as vegetation clearance or earthworks) or “unmodified” slopes. This classification approach recognises that modified slopes can occur outside urban cores and that natural slopes may persist within urbanised regions.

4.

Soil moisture sensors and satellite data agreement

To assess the agreement between satellite-derived soil moisture estimates and in situ measurements, a set of statistical metrics and threshold values were adopted based on established practices in remote sensing validation studies [53,54]. The following criteria were applied:

• Mean Bias Error (MBE): The mean bias error quantifies the systematic difference between the satellite and in situ measurements, indicating whether the satellite product consistently overestimates or underestimates soil moisture values relative to ground truth. Positive values denote overestimation, while negative values indicate underestimation;
• Root Mean Square Error (RMSE) and Mean Absolute Error (MAE): These metrics describe the overall and average discrepancies, respectively, between the satellite and ground-based measurements. The RMSE gives greater weight to larger errors due to its squared term, while the MAE represents the mean of absolute differences, providing an intuitive average error magnitude;
• Pearson Correlation Coefficient (r): The Pearson correlation coefficient was used to measure the strength and direction of the linear relationship between the two datasets. A threshold of r ≥ 0.30 was adopted as the minimum acceptable level of linear association for reliable agreement at each station and sensor depth;
• Coefficient of Variation Difference (CV_diff): This metric captures the relative difference in variability (as a percentage of the mean) between the satellite-derived and in situ soil moisture datasets. A CV_diff below 20% was set as the threshold to ensure that the relative spread of values remains comparable across both data sources, an essential check given the different observational natures of the two datasets.

These criteria were evaluated individually for each geosensor station and soil depth. Stations meeting both the correlation (r > 0.30) and variability (CV_diff ≤ 20%) thresholds were elected to proceed with the gap filling. From this elected subset, an average additive bias correction was calculated for each satellite-derived soil moisture product and geosensor depth combination. These bias corrections were subsequently applied to adjust the satellite data, improving its alignment with the in situ observations at locations where model performance was demonstrably consistent [55].

5.

Assembling the hydro-meteorological landslide inventory

After preprocessing, a comprehensive landslide dataset was created, linking each event to corresponding rainfall and soil moisture conditions (Figure 1), based on spatial proximity and temporal confidence. A procedure in Python was developed for each landslide event to match the nearest rainfall and soil moisture stations based on the following criteria in order:

• Looks up for stations with at least min_records of 20 days of valid rain data in the 30 days before the landslide date;
• Tries first stations in the same city with that data count;
• Falls back to the nearest station, ignoring city but still meeting data count;
• If none meet the threshold, it falls back to the nearest station regardless of data count.

This approach maximised data completeness while retaining spatial proximity, providing representative hydro-meteorological conditions for each event. Non-landslide days within the study period were retained as non-events, forming a complete landslide (1) and non-landslide (0) dataset for subsequent modelling.

2.2.2. Time Resolution-Based Datasets

Given the varying temporal resolution of landslide occurrence records, a two-tier approach was implemented to assess the influence of antecedent conditions and short-term triggering rainfall events. The dataset was stratified into two subsets:

• Daily resolution model: The first set of models examined the role of cumulative rainfall and antecedent soil moisture at a daily resolution. Events with exact dates only (T3), enabling only daily summary variables;
• Finer resolution: Events with precise timestamps or approximate time periods within the day (T1 and T2), allowing for the extraction of finer temporal scale hydro-meteorological variables.

This stratification facilitated the comparison of predictive skill using coarse versus refined temporal data, reflecting practical data constraints in landslide monitoring.

2.2.3. Rainfall Data Featuring Engineering and Completeness

To generate hydro-meteorological variable precursors to landslides, a suite of antecedent rainfall and soil moisture variables was derived, including the following:

• Cumulative rainfall totals over 1, 3, 7, 14, and 30 days;
• Moving maximum rainfall intensities at 10 min, 30 min, 1 h, 3 h, 6 h, and 12 h intervals;
• Antecedent soil moisture conditions via daily, 3-day, and 7-day averages and maxima;

Data completeness was evaluated for each rain gauge station and month by calculating the percentage of expected records effectively collected within that period. The completeness percentage was computed as follows:

$$C_{i,j}= {\displaystyle \frac{N_{i, j}^{obs}}{N_{i, j}^{exp}}} \times 100$$

(1)

where

• $C_{i,j}$ is the completeness percentage for station i in month j;
• $N_{i, j}^{obs}$ is the number of unique records observed in month j for station i;
• $N_{i, j}^{exp}$ is the total expected number of records for that month (either number of hours for Cemaden network or number of days for EMAE and SABESP stations).

For Cemaden stations, the expected number of records per month was determined by the total number of hours in the month. For EMAE and SABESP stations, completeness was assessed based on daily precipitation totals, using the number of days in the corresponding month.

2.3. Inventory Analysis and Web Application

Following the compilation and processing of the landslide inventory and hydro-meteorological data, a series of exploratory analyses were conducted to characterise landslide occurrence patterns, assess data completeness, and evaluate hydro-meteorological precursors. These included the following:

(i)

Temporal and spatial distribution assessments, analysing annual, decadal, and monthly landslide frequencies and their geographic distribution across municipalities. Temporal confidence and geographical accuracy classifications were summarised to appraise dataset quality;

(ii)

Seasonal trends analysis, comparing monthly landslide occurrences with regional average rainfall to identify peak activity periods and confirm wet season patterns typical of the tropical coastal region;

(iii)

Hydro-meteorological data completeness evaluation, calculating the percentage of expected rainfall and soil moisture records retrieved per monitoring station and month, using the completeness metric from Section 2.1.3. Heatmaps illustrated spatial and temporal monitoring gaps;

(iv)

Soil moisture data validation and bias correction, comparing ERA5-Land reanalysis soil moisture against in situ sensor data. Performance was assessed via Pearson’s correlation, the Mean Absolute Error (MAE), and the coefficient of variation difference (CV_diff) to identify suitable stations and depths for correction.

To support the integrated use of the landslide catalogue and hydro-meteorological records, a web-based application was developed using the Streamlit library in Python. The tool enables the following:

(i)

Interactive visualisation of individual landslide events alongside antecedent rainfall and soil moisture conditions;

(ii)

Filtering and querying by temporal, spatial, geomorphological, and confidence parameters;

(iii)

Operational and research support, offering municipal agencies, researchers, and planners a platform to assess landslide hazard contexts in near real time or retrospectively.

This exploratory analysis and application prototype provide a foundational resource for the ongoing development of data-driven landslide early warning systems and risk assessment frameworks for the Baixada Santista region.

3. Results

3.1. Landslide Inventory Overview

The final inventory includes 2530 rainfall-induced landslide records spanning 1988–2024. Figure 3 illustrates the spatial distribution of global landslide occurrences with at least the date assigned across the Baixada Santista region, alongside inset donut charts summarising event classifications by geographical confidence, temporal confidence, and landslide type. Figure 4 shows the inventory filtered with those landslides that have an accurate time (T1) and time period (T2) assigned, with the same donut charts set up.

3.1.1. Landslide Inventory with Date (LID)

For this dataset, all the landslide events were kept. In terms of locational accuracy, 67.2% of the events were classified as G2 (high confidence based on address or neighbourhood-level verification), 22.5% as G1 (precise geocoded coordinates from field verification or high-resolution imagery), and 10.3% as G3 (approximate locational attribution based on general area or city reference).

Regarding temporal confidence, the majority (81.2%) were assigned to T3 (date only), while 10.9% were categorised as T2 (approximate time periods), and 7.9% as T1 (precise occurrence time). Landslide types were covered predominantly by translational movements (88.1%), followed by rockfalls (9.7%) and fewer debris flows (0.9%).

This visual and categorical overview highlights the predominance of translational slides in the region’s recent landslide history and the ongoing challenges in obtaining precise temporal data for empirical modelling and operational early warning.

Spatially, Cubatão (40.8%) and Santos (40.7%) accounted for the majority of records, reflecting their continuous and systematic data provision throughout the study period. In contrast, Guarujá (13.4%) and São Vicente (5.1%) contributed fewer records, largely due to gaps in obtaining information about specific periods in local reporting.

3.1.2. Landslide Inventory with Time (LIT)

From the complete inventory, 472 events were classified with precise or approximate occurrence times (T1 and T2). It is worth highlighting that most of the landslides in the LIT had higher locational accuracy than in the LID, with 71.0% G1, 19.9% G2, and 9.1% G3 classifications. Temporal confidence was split between 58.5% for T2 and 41.5% for T1, reflecting a reasonable proportion of events with precise occurrence timing.

Translational landslides were the predominant type in this subset, accounting for 96.4% of the events, followed by rockfalls (2.5%) and debris flows (0.6%). In terms of spatial distribution, Guarujá (37.7%) contributed the largest share of time-attributed records, followed by Santos (27.1%), Cubatão (26.9%), and São Vicente (8.3%). This pattern reflects not only differences in landslide occurrence, but also in reporting practices and data availability across municipalities. Guarujá was the municipality most affected by landslides in recent events, such as the 2020 disaster, which made it possible to collect more accurate information.

Notably, this subset shows an even stronger predominance of translational slides than the full database, reinforcing the typological pattern typical of rainfall-induced landslides in the region’s colluvial slopes.

3.1.3. Temporal Distribution

The annual and decadal distribution of landslide events was analysed (Figure 5). The annual distribution (Figure 5a) reveals a clear increase in recorded landslides starting in the early 1990s, with notable peaks in 2010 (13.9%), 2011 (11.4%), 2020 (12.5%), and 2013 (7.3%), corresponding to significant rainfall-induced events and improvements in reporting protocols. Other smaller annual peaks appear in 1994 (2.8%), 1999 (3.0%), and 2009 (6.7%).

Decadal aggregation (Figure 5b) indicates that 41.2% of events occurred in the 2010s, followed by 31.3% in the 2000 s, 14.5% in the 1990s, and a rising share in the current 2020s (12.9%), reflecting recent extreme rainfall years and ongoing data collection. The very low proportion in the 1980s (0.1%) likely reflects incomplete records from early years.

It is important to note that the observed temporal pattern likely results from a combination of factors such as (i) actual increases in landslide occurrence during extreme rainfall years, (ii) progressive improvements in event reporting, especially from the late 2000s onward, and (iii) landslides without dates assigned not being considered in this inventory. These influences should be considered when interpreting apparent trends in landslide frequency over time.

The spatial distribution of landslide records reveals uneven data availability across the four municipalities. Cubatão and Santos exhibit continuous and consistent reporting throughout the study period, reflecting well-established local monitoring and civil defence systems. In contrast, Guarujá and São Vicente show significant temporal gaps and inconsistent reporting between 2012 and 2019 (Figure 5c). Both cities’ landslide records are sparse, lacking documented events likely due to limited local data collection or access.

3.2. Field Validation and Landslide Characterisation

Field campaigns were conducted between March and April 2025 across the municipalities of Guarujá, Santos, and Cubatão. UAV flights were performed using a DJI Mini 2, allowing for the capture of high-resolution aerial imagery of recent and older landslide scars. Some examples can be seen in Figure 6. This approach enabled the identification and visual characterisation of shallow translational landslides, as well as occasional soil and rock slides, within densely vegetated and urbanised areas.

This field validation was essential not only for refining the precise geographic location of events—classified as G1 confidence—but also for verifying landslide typologies and material characteristics, which are often difficult to determine through satellite imagery alone, even when using very high-resolution sources like Google Earth.

3.3. Geoenvironmental Characterisation

Landslide occurrences were predominantly associated with anthropogenically modified slopes (55.5%), reflecting extensive urban modification, while unmodified slopes accounted for 44.5%. Urban land use dominated (47.4%), followed by forest formations (35.9%) and mixed-use mosaics (13.3%). Other land cover types were minimal. Lithologically, ophthalmic migmatites (AcMg, 29.1%), mica schists (PSpX, 18.1%), and neosome stromatic migmatites (AcMn, 12.4%) comprised the majority of substrates linked to landslides. These results suggest that the concentration of landslides in anthropogenically modified slopes and urbanised areas reflects not only the physical susceptibility of these environments, but also the greater exposure and reporting frequency in such locations. While urban expansion often alters slope stability through vegetation removal and unregulated construction [56], it is important to note that landslides in natural escarpments are historically significant, but not taken into account in this work due to a lack of precise temporal data. Additionally, the predominance of landslides in migmatites and mica schists aligns with previous susceptibility studies in the region [46], reaffirming the inherent lithological predisposition of these substrates to slope failures. Figure 7 graphically showcases the geoenvironmental variables associated with landslides.

3.4. Rainfall and Soil Moisture Conditions

3.4.1. Monthly Landslide Frequency and Seasonal Trends

Figure 8 presents the distribution of recorded landslide occurrences per month alongside the region’s average historical rainfall patterns. As expected, landslide activity exhibits a strong seasonal trend, with a marked concentration during the summer months. The highest number of events was recorded in March (732 landslides), followed by January (536), February (432), and December (261). These months typically coincide with the region’s peak rainfall season, reinforcing the established correlation between intense rainfall episodes and slope instability in the study area.

From April onwards, landslide occurrences decline sharply, reaching minimal values during the dry season months (May–September), where monthly counts remain below 50 events. A secondary, modest increase is observed in October (94) and November (98), reflecting the onset of pre-summer rains.

This distribution highlights a classic tropical coastal pattern, where cumulative rainfall and consecutive rainy days in the summer months act as primary triggers for shallow translational landslides. No significant lag effects were observed—landslide peaks align closely with rainfall maxima. Notably, certain months with substantial rainfall (like April) exhibit relatively lower landslide numbers, likely reflecting variations in rainfall intensity thresholds, antecedent soil moisture, and data completeness for those periods.

3.4.2. Rainfall and Soil Moisture Data Completeness

The average monthly data completeness (%) was computed for each rain gauge station across the study area, considering the percentage of expected records effectively collected during the period of analysis. Within the CEMADEN network, the average monthly completeness values ranged from 45.9% to 96.9%. Stations in Santos, Guarujá, and Cubatão generally showed higher completeness, with Santos 354850010A and Guarujá 351870101G reaching averages above 96%, while others like São Vicente 355100902G fell below 50%. In contrast, the EMAE rain gauges located in Cubatão exhibited excellent coverage, with both stations 2346160 and 2346468 achieving 100% completeness across the evaluated period. The SABESP SABOO station (2346279) in Santos also demonstrated very high data availability, averaging 99.8% completeness.

Some CEMADEN stations exhibited significant data gaps, with average monthly completeness falling below 70% in several cases, notably São Vicente 355100902G (45.9%). These gaps indicate periods of missing rainfall records, potentially compromising the reliability of rainfall–landslide associations in those areas. Caution is advised when interpreting results for locations with intermittent coverage. Figure 9 highlights these stations and the periods with reduced data availability.

The soil moisture dataset comprises measurements from 16 stations, with overall data completeness ranging from approximately 65% to 86%. The mean completeness was 77.8% (±5.9%), indicating moderate variability across stations. Station 351350401G exhibited the highest completeness, consistently recording over 81% of expected data across all sensor depths, reflecting robust sensor reliability. In contrast, station 355100904G had the lowest completeness (65%), with significant gaps exceeding 38% at multiple depths, suggesting potential sensor or transmission issues. Most stations achieved between 70% and 85% completeness, although data gaps varied by sensor level and location, emphasising the necessity for targeted maintenance or gap mitigation to improve dataset quality. Figure 10 illustrates the overall soil moisture data completeness across 16 monitoring stations.

3.4.3. Soil Moisture Data Validation and Gap Filling

The comparison between satellite-derived soil moisture estimates from the ERA5-Land reanalysis dataset and in situ geosensor measurements revealed notable variability in performance across stations and soil depths. Pearson correlation coefficients (r) ranged from −0.18 to 0.71, with a median value of 0.33. Higher correlation values were typically observed at deeper soil levels and at stations situated within homogeneous land cover areas.

The Mean Absolute Error (MAE) values ranged between 1.24% and 15.78% volumetric percentage points, with generally lower errors associated with stations exhibiting higher correlation values. Stations where the coefficient of variation difference (CV_diff) remained below 20% and r exceeded 0.5 were classified as accepted for bias correction purposes. Figure 11 presents the metric outputs in a box plot graphic set.

Only one geosensor station (351350401G) met the established reliability thresholds (Pearson’s r > 0.5 and CV_diff ≤ 20%) for comparison with ERA5-Land soil moisture estimates. At this station, the swvl3 layer demonstrated acceptable agreement at the sm_level1 (r = 0.51, CV_diff = 5.61%), and swvl4 at the sm_level4 (r = 0.71, CV_diff = 0.43%). Based on these results, average additive bias corrections were computed for these specific ERA5 soil moisture layers and sensor depths.

3.5. Application Records and Visualisation

An application was developed with the help of the Python library streamlit to store BrazInv. The application is publicly available, with some landslide events associated with their antecedent rainfall and soil moisture conditions, and it will be further developed with more data available throughout the Ph.D. research of the author. The columns included in the dataset are presented below:

• datetime_brazil: Date and time of record in Brazilian local time;
• rain_mm_24 h, rain_mm_48 h, rain_mm_72 h: Cumulative rainfall totals in millimetres over the preceding 24, 48, and 72 h, respectively;
• geo_confidence: Geographical confidence classification (G1–G3) indicating locational accuracy of the landslide record;
• temporal_confidence: Temporal confidence classification (T1–T3) indicating date/time accuracy of the landslide occurrence;
• city: Municipality name;
• address: Street address of the landslide site;
• neighbourhood: Local neighbourhood or district name;
• latitude, longitude: Geographical coordinates of the landslide site in decimal degrees;
• slope_type: Slope category where the landslide occurred (e.g., cut, natural, or fill);
• landslide_id: Unique identifier for each landslide record;
• landslide_occurred: Boolean flag indicating if a landslide occurred (1) or not (0);
• ws_id: Watershed identification code where the landslide site is located;
• lulc: Land use and land cover classification of the landslide site;
• rainfall_station_code: Code identifying the associated rainfall monitoring station;
• sm_station_code: Code identifying the associated soil moisture station;
• sm_l1–sm_l6: Soil moisture values (%) at six soil depth layers, varying 0.5 m amongst each one, ranging from 0.5 m to 3 m.

A graphic plot of a single landslide event and its triggering rainfall and soil moisture antecedent conditions was enabled, considering the LID dataset as an example of visualisation—also available in the application (Figure 12).

4. Discussion

4.1. Strengths and Limitations of the Inventory

This study presents the most comprehensive rainfall-induced landslide inventory ever compiled for the Baixada Santista region, integrating multi-source records from 1988 to 2024, including temporal and spatial attributes with verified classifications. A major strength lies in the systematic categorisation of geographical (G1–G3) and temporal (T1–T3) confidence, enabling nuanced use in subsequent empirical and operational applications such as rain threshold validation and landslide susceptibility mapping. The field validation campaigns using UAV imagery were pivotal in refining the locational accuracy of key events and clarifying landslide typologies, especially where satellite imagery (e.g., Google Earth) proved insufficient.

An additional operational advantage is the integration of soil moisture data, both from in situ geosensors and bias-corrected ERA5-Land estimates, which opens valuable opportunities for empirical, rainfall-based landslide early warning modelling. This integrated hydro-meteorological perspective remains rare in inventories in the literature and enhances the inventory’s applicability for landslide forecasting systems.

Nonetheless, limitations persist. The temporal completeness is uneven, with most records classified as T3 (date only), and only 18.8% of the inventory is assigned with precise (T1) or approximate (T2) times. Spatially, data availability is heterogeneous across municipalities, with Cubatão and Santos consistently represented, while Guarujá and São Vicente exhibit significant gaps, mainly related to data availability. Moreover, older records prior to the 1990s remain scarce, reflecting the absence of systematic reporting before this period. A further limitation concerns the attempted gap-filling of soil moisture geosensor data using ERA5-Land reanalysis values. Only one station (351350401G) achieved the minimum reliability threshold (r > 0.5, CV_diff ≤ 20%), substantially restricting the operational use of uncorrected satellite-derived soil moisture for real-time landslide early warning in this tropical and coastal setting. This constraint is likely influenced by the coarse spatial resolution of ERA5-Land relative to the highly localised soil and land cover variability of the region, a challenge also observed in similar environments [54,57]. However, despite its valuable temporal coverage, ERA5-Land’s spatial resolution may not capture the fine-scale soil moisture variations critical for localised landslide initiation zones. Future work should consider integrating higher-resolution sources, such as Sentinel-1 SAR products, or densifying in situ sensor networks to enhance spatial accuracy.

To address the temporal limitations of landslide records, we recommend implementing more rigorous and standardised data collection protocols to improve the temporal accuracy of event reporting. In relation to the uneven availability of spatial data, particularly in municipalities with significant gaps such as Guarujá and São Vicente, strengthening partnerships with local authorities and expanding monitoring infrastructure is essential. For the integration of soil moisture data, further local calibrations and an increased density of in situ sensors are advised to enhance data reliability. Lastly, the scarcity of records prior to the 1990s could be mitigated through the digitisation and compilation of historical landslide records from municipal archives and other local sources.

4.2. Gaps Addressed in Brazil and International Landslide Data Practices

Brazil has long faced challenges in establishing consistent, standardised, and spatially explicit landslide databases. The BrazInv addresses several of these limitations by introducing an explicit confidence classification for both temporal and locational attributes, incorporating field-validated G1 points via UAV imagery and digitising historical municipal records. Additionally, the soil moisture integration represents an operational leap forward, enabling landslide event characterisation alongside hydro-meteorological drivers—a methodological innovation largely absent in Brazilian national practice.

The public availability of the BrazInv application further represents an important transparency and data-sharing contribution, supporting not just scientific research, but also municipal risk management. This approach helps to tackle the chronic issue of limited data accessibility that has historically constrained landslide risk assessment in the country.

4.3. Comparison to International Inventories and Highlights

When compared to established international landslide inventories, BrazInv distinguishes itself through its detailed regional focus, integration of soil moisture data, and the use of spatial and temporal confidence classifications at a granular scale.

Table 6. Comparative summary of landslide inventories.

Inventory	Location	Time Frame	Rainfall Data Linked	Confidence Classification	Soil Moisture Data Linked	Public Availability
BrazInv	Baixada Santista (Brazil)	1988–2024	Yes	Geographical and temporal	Yes	Partially open
ITALICA [24,25]	Italy	1996–2021	Yes	Geographical and temporal	No	Open access
Italian IFFI Inventory [58]	Italy	Historical–present	No	No	No	Open access
NASA Global Landslide Catalog (GLC) [59,60]	Global	2007–present	No	No	No	Open access
USGS Landslide Inventory [61]	USA	Historical–present	No	Geographical	No	Open access
British Geological Survey (BGS) [62,63]	United Kingdom	Historical–present	Yes 1	No	No	Open access
New Zealand National Landslide DB [64]	New Zealand	Historical–present	No	No	No	Open access
São Paulo State IG Inventory	São Paulo State (Brazil)	Historical–present	No	No	No	Open access
Nicaragua Landslide Database [65]	Nicaragua	1826–2003	Yes	No	No	Limited

¹ The BGS landslide database is linked to rainfall data from MetOffice, with monthly updates.

summarises the key characteristics of BrazInv and selected global inventories for context.

While Table 6 summarises the key attributes of selected global inventories, their compilation methodologies differ markedly. The USGS database standardises disparate local-scale inventories by harmonising core attributes and applying a semi-quantitative confidence ranking to account for variability in data quality and mapping precision. The British Geological Survey (BGS) National Landslide Database compiles over 17,000 records from geological maps, surveys, public reports, and media sources, with each entry verified through a quality assurance process to ensure accurate location and descriptive detail. ITALICA, by contrast, focuses on rainfall-induced landslides in Italy, integrating news reports and institutional data with a unique emphasis on high spatial and temporal accuracy, supported by an automated workflow for reconstructing triggering rainfall conditions.

BrazInv shares common limitations with these inventories, including the underreporting of minor or historical events and biases towards more visible landslide types. However, it advances current practices by systematically incorporating hydro-meteorological parameters—such as rainfall accumulation and soil moisture—into each event record, offering improved context for landslide triggering conditions. Despite these enhancements, BrazInv faces similar challenges of temporal incompleteness and typological bias, particularly towards translational landslides.

4.4. Potential Applications

The refined inventory provides an empirical foundation for several applications:

• Operational LEWSs; the subset of landslides with T1/T2 times and G1/G2 locations can be used for threshold definition and the validation of rainfall and soil moisture-based warnings, which is a critical gap in current Brazilian systems, which largely rely on rainfall-only thresholds with limited empirical landslide occurrence data for calibration;
• Empirical and physical modelling, by supporting hydro-meteorological correlation studies and landslide susceptibility mapping and offering calibrated inputs for models requiring precise temporal and spatial landslide data, a method applied effectively in regions such as Italy (e.g., [42]);
• The inventory also holds operational and scientific value for climate-related landslide monitoring, given its long temporal coverage of rain gauges and landslides coupled with the integration of rainfall and soil moisture information. The observed seasonal patterns and rainfall–landslide relationships enable further assessments of potential shifts in landslide seasonality and frequency under changing climatic conditions. Additionally, the inventory’s structure facilitates data-driven susceptibility and hazard modelling initiatives in urbanised, landslide-prone regions, providing robust, confidence-classified data for probabilistic or machine learning-based approaches;
• Public policy and civil defence planning; by identifying municipalities and decades with data gaps, the inventory guides future monitoring investments and risk management priorities. In addition to enhancing public transparency and facilitating access for decision-makers, the BrazInv application is structured to support its integration into municipal contingency planning protocols. Its design also enables use in public risk communication campaigns, fostering citizen education and the participatory mapping of landslide hazards. Hence, we highly encourage continuous investment in the current network to maintain and even amplify it for more accurate landslide forecasting.

4.5. Transferability of the Framework

The methodology applied in developing BrazInv, integrating multi-source records, confidence classification, field validation with UAV imagery, and hydro-meteorological data linkage, is highly transferable to other Brazilian regions facing similar data inconsistencies. Its alignment with internationally recognised inventory standards (e.g., IAEG suggested practices [24,25,66]) positions it as a model framework for broader application.

However, its successful replication depends on local institutional engagement, consistent data availability, and perhaps access to UAV resources for field validation. As shown in this study, UAV-assisted validation was crucial to refining G1 event locations and typologies. Future inventories in other regions should incorporate this as standard practice, particularly in densely vegetated or urbanised slopes where landslide scars remain visually inaccessible from conventional aerial or satellite imagery.

Overall, BrazInv advances landslide inventory practice in tropical, data-scarce settings by combining multi-source records, temporal and spatial confidence classifications, UAV-based validation, and the integration of both rainfall and soil moisture metrics. The predominance of landslides on anthropogenically modified slopes underlain by migmatites and mica schists mirrors patterns observed in other humid tropical regions, where a combination of lithological susceptibility and urban encroachment amplifies hazard potential. These geoenvironmental patterns, already documented in regional susceptibility studies, are now linked here to a multi-decadal hydro-meteorological context, offering a richer empirical basis for hazard modelling.

While temporal incompleteness and typological biases persist, the framework demonstrates that targeted field verification, systematic classification, and the integration of hydro-meteorological drivers can substantially improve the reliability and applicability of inventories. By aligning with international best practices, yet adapting to local institutional realities, BrazInv provides a replicable model for inventory development in comparable settings across Latin America and other tropical coastal regions. As such, it not only addresses a long-standing data gap in Brazil, but also contributes to the global discourse on operationalising landslide data for early warning, susceptibility mapping, and climate-related hazard assessment.

5. Conclusions

This study developed the most comprehensive rainfall-induced landslide inventory to date for the Baixada Santista region (1988–2024), integrating multi-source records, hydro-meteorological data, and UAV-assisted field validation. Two key contributions are observed: (i) the introduction of detailed geographical and temporal confidence classifications enabling transparent precision assessment, and (ii) linking soil moisture to the rainfall-induced landslides inventory, addressing a significant gap in Brazil’s landslide inventory reliability and in the literature, respectively. The main outcomes include the following:

(i)

Compilation of 2534 landslide records systematically classified by spatial (G1–G3) and temporal (T1–T3) confidence, enabling transparent data precision assessment;

(ii)

Demonstration that UAV field validation markedly improves locational accuracy and typological consistency, especially in complex terrain;

(iii)

Integration of soil moisture data, highlighting its potential for operational early warning while identifying the limitations of satellite-derived products without local calibration in tropical coastal environments;

(iv)

Identification of spatial and temporal reporting gaps, informing future monitoring and data management priorities;

(v)

Development of a prototype of a publicly accessible platform to showcase BrazInv, promoting transparency and usability for both research and risk management.

The inventory offers substantial operational and scientific value by supporting (i) landslide early warning systems through threshold definition and validation; (ii) climate-related landslide monitoring by enabling the assessment of long-term trends and seasonality; and (iii) data-driven susceptibility and hazard modelling initiatives in urbanised, landslide-prone areas. From these findings, we conclude the following:

(i)

Confidence classifications significantly enhance inventory reliability for both research and early warning applications, though precise temporal data remains challenging;

(ii)

Multi-source integration combined with UAV validation is essential for producing detailed and robust inventories in urbanised tropical regions;

(iii)

Soil moisture data can complement rainfall-based warnings, but requires local calibration to be operationally effective when using satellite reanalysis estimation;

(iv)

Persistent data gaps and biases highlight the need for systematic, standardised landslide documentation at municipal and national levels;

(v)

The methodology and framework are transferable worldwide and provide a model for other Brazilian regions and similar international contexts.

Overall, we are able to conclude that this work provides a leap forward in landslide risk management by improving data reliability and transparency, supporting more effective early warning systems, and contributing to disaster risk reduction, aligned with the Sendai Framework for Disaster Risk Reduction. To ensure the inventory’s long-term sustainability, it is essential to establish institutional roles and secure dedicated funding and coordination by local authorities and scientific agencies.