Low-cost distributed sensing for rainfall-triggered landslide early warning

Abstract

Rainfall-triggered shallow landslides are commonly forecast using empirical intensity-duration thresholds, but threshold performance degrades where antecedent wetness, soil hydraulic response, and local drainage are poorly observed. We report a fictional three-wet-season deployment of a low-cost distributed sensing network on two landslide-prone road-cut catchments in the Serra do Mar foothills of south-eastern Brazil. Forty-six solar-powered nodes measured rainfall, volumetric water content, shallow pore pressure, soil temperature, and slope-normal tilt at 5 min intervals using LoRaWAN telemetry and local buffering. Sensor observations were fused with radar rainfall and a digital terrain model to produce uncertainty-aware warning levels based on a Bayesian logistic threshold model. Relative to a rainfall-only intensity-duration threshold, the fused model reduced false alarms from 0.41 to 0.22 per storm season while maintaining event detection for seven of eight verified shallow failures. Median warning lead time was 4.1 h for failures with measurable pore-pressure response and 1.8 h for rapid failures on thin colluvium. Network availability averaged 92.6%, with most data loss caused by vegetation shading of solar panels and gateway outages during convective storms. The results indicate that low-cost sensing can improve site-specific landslide early warning when treated as uncertain hydrological evidence rather than as a deterministic trigger, but calibration remains local and warning dissemination must retain human review.

Introduction

Rainfall thresholds remain the most widely used basis for landslide early warning because they are inexpensive, interpretable, and compatible with regional rain-gauge or radar networks. The classic intensity-duration formulation introduced by Caine [1] and later updated and regionalised by Guzzetti and co-workers [2,3] provides a practical way to connect rainfall history to shallow landslide occurrence. These thresholds are useful, but they compress a hydrological problem into rainfall variables alone. Two storms with the same duration and intensity can produce different pore-pressure responses depending on antecedent wetness, soil depth, hydraulic conductivity, root reinforcement, drainage convergence, and local slope geometry.

Mechanistic studies of rainfall infiltration and slope stability show why this compression is imperfect. Transient infiltration, pressure-head diffusion, and topographic convergence control when a shallow colluvial slope approaches failure [4,5]. Recent hydrological critiques therefore argue that warning thresholds should incorporate wetness or pore-pressure information rather than treating rainfall as the only observable state [6]. Reviews of rainfall thresholds reach a similar conclusion: empirical thresholds work best when calibrated densely and locally, while transfer across lithology, climate, and land-use settings is hazardous [7].

Operational landslide early warning systems must also balance scientific skill with practical communication. General frameworks emphasise four linked tasks: monitoring, forecasting, warning, and response [8]. Regional systems in the United States, Italy, Switzerland, and elsewhere have shown that thresholds can support operational decisions when their uncertainty is understood and their performance is evaluated with event inventories [9,10,11,12]. The challenge in many tropical and low-resource settings is that dense hydrological monitoring is expensive and maintenance-intensive. Low-cost wireless sensor networks, low-cost GNSS, and open hardware platforms provide one possible route to more spatially distributed evidence [13,14,15,16].

This article evaluates that route in a fictional but technically constrained field study. We do not claim that low-cost sensors replace geomorphological expertise or rainfall thresholds. Instead, we ask whether distributed observations of wetness, pore pressure, and tilt can reduce false alarms and improve lead time for site-specific warning decisions when fused with conventional rainfall thresholds.

Study area and monitoring design

The deployment covered two steep road-cut catchments, here labelled Alto Norte and Alto Sul, on the inland margin of the Serra do Mar escarpment. Both catchments are underlain by deeply weathered gneiss and mantled by 0.6-2.1 m of sandy clay colluvium over saprolite. Slopes range from 27 to 43 degrees, with concave hollows draining toward culverts beneath a municipal access road. Historical maintenance records listed 22 shallow slides and debris slides over the previous 14 years, mostly during summer convective storms following multi-day wet periods. No houses lie directly below the monitored hollows, but the road is a school-bus and emergency-access route.

The sensor network comprised 46 field nodes and two gateways. Each node used a low-power microcontroller, LoRa radio, barometric pressure sensor, soil-temperature probe, one or two capacitive water-content sensors, and a three-axis MEMS accelerometer mounted in a sealed inclinometer sleeve. Twelve nodes also included vented 50 kPa piezoresistive pore-pressure transducers installed at the colluvium-saprolite contact. Six tipping-bucket rain gauges were installed across the two catchments to capture convective rainfall gradients. Two low-cost differential GNSS units were placed on the crown of a historically active cut slope, following the principle that low-cost geodetic networks can detect slow displacement when baselines and multipath are managed carefully [14].

Node spacing was not uniform. Higher density was used in hollows with convergent drainage and known maintenance scars. Ridge and spur locations received fewer nodes because they served mainly as hydrological controls. Telemetry used LoRaWAN at 915 MHz with 5 min sampling during rainfall and 30 min sampling during dry periods. Every node buffered at least 21 days of data on local flash storage. This local buffering was important because gateway outages occurred during exactly the storms that mattered most.

Calibration and data quality control

Before installation, water-content sensors were calibrated against gravimetric samples from each catchment. A single factory calibration gave systematic errors as large as 0.08 m^3 m^-3 in the clay-rich A horizon, so site-specific two-point calibration was used for all reported analyses. Pore-pressure transducers were checked against a water column before installation and again after removal. Accelerometer tilt estimates were temperature-corrected using a laboratory calibration from 8 to 42 deg C. Rain gauges were compared against a reference weighing gauge during three controlled pour tests and during one low-wind storm.

Quality control was deliberately conservative. Packets with physically impossible jumps, frozen values, battery voltage below 3.35 V, or radio timestamps inconsistent by more than 90 s were flagged. Soil-moisture spikes coincident with battery brown-out were removed. Pore-pressure records were corrected for barometric pressure and then screened for drainage-lag artefacts after high-intensity rainfall. Tilt signals were not interpreted as displacement unless the change persisted for at least 30 min and was observed by neighbouring nodes or by the crown GNSS units. This caution reflects a broader lesson from monitoring systems for rapid mass movements: noisy measurements can create false confidence if treated as direct precursors without context [11].

Network availability was measured as the fraction of expected observations that passed quality control. Mean availability was 92.6% across the three wet seasons. Availability was lower for shaded hollow-floor nodes, 86.8%, than for ridge nodes, 96.1%. The dominant failure modes were solar-panel shading by vegetation, gateway power loss during long cloudy spells, connector corrosion in two pore-pressure sensors, and animal disturbance of shallow sensor cables. The bill of materials per standard node was USD 118 excluding labour, while pore-pressure nodes cost USD 246.

Warning model

The baseline warning model used an intensity-duration rainfall threshold of the form I = aD^-b, where I is mean rainfall intensity and D is storm duration. Parameters were fitted from the local 14-year maintenance inventory and adjusted using the first wet season of monitored data. Storms were separated by six rain-free hours. This threshold provided an interpretable first warning line, but it produced frequent alerts during storms that occurred after long dry periods and missed two small failures after moderate rainfall on already wet slopes.

The fused model used Bayesian logistic regression with predictors selected to remain operationally transparent: distance above the local intensity-duration threshold, 72 h antecedent rainfall, catchment-median soil saturation index, maximum positive pore-pressure change over 6 h, and the fraction of hollow nodes showing persistent downslope tilt acceleration. The saturation index was derived from the percentile rank of calibrated volumetric water content at each node relative to its wet-season distribution. This percentile formulation reduced bias between sensors with different absolute calibration uncertainty. Prior distributions were weakly regularising and centred on the expectation that rainfall exceedance, wetness, pore pressure, and tilt acceleration should increase failure probability.

Warnings were issued at three levels. Level 1 indicated hydrological preparation, with forecast duty officers notified but no road action. Level 2 indicated probable instability in one or more monitored hollows and triggered visual inspection when conditions allowed. Level 3 indicated high probability of imminent failure and recommended temporary road closure pending field confirmation. The model did not issue public messages automatically. It produced probability bands and uncertainty intervals for the duty officer, because landslide early warning is partly a communication and decision process rather than a sensor-threshold exercise [8,12].

Observed hydrological response and failures

The three wet seasons contained 61 threshold-defined storms, of which 19 exceeded the rainfall-only Level 1 threshold. Eight shallow failures were verified by field inspection or drone imagery. Five were shallow translational slides in hollows, two were road-cut slips with exposed saprolite, and one was a small debris slide that entered a drainage ditch. The mapped failure areas ranged from 18 to 310 m^2. None caused injuries, and only two required temporary road closure.

Hydrological response differed strongly between the two catchments. Alto Norte had thicker colluvium and slower drainage. Soil saturation rose over several days and pore pressure increased gradually, producing warning lead times of 3.5-7.2 h before five verified failures. Alto Sul had thinner colluvium and sharper rainfall response. Two failures occurred after short, intense storms in which pore-pressure sensors responded less than 70 min before failure. This difference is consistent with physically based studies showing that infiltration timing and topographic convergence can dominate local failure timing even under similar rainfall totals [4,5,17].

Tilt observations were useful but not sufficient by themselves. Persistent tilt acceleration was detected before six of eight failures, but similar tilt anomalies occurred during two non-failure storms where wet soil creep probably occurred without rupture. GNSS displacement at the road-cut crown exceeded 12 mm during the largest storm and then partially recovered, suggesting shallow deformation rather than deep-seated movement. We therefore treated tilt and GNSS as supporting evidence of deformation, not as standalone warnings.

Warning performance

The rainfall-only threshold detected seven of eight verified failures, but produced 17 Level 2 or Level 3 warnings over three seasons. The fused model also detected seven of eight failures, while reducing Level 2 or Level 3 warnings to ten. Normalised by storm season, the false-alarm rate decreased from 0.41 to 0.22 per season for warnings that would have required field inspection or road action. Median lead time was 4.1 h for failures with a measurable pore-pressure response and 1.8 h for rapid thin-soil failures. The missed event was a small road-cut slip triggered by blocked drainage immediately below a culvert; no sensor was installed in the culvert outlet zone.

The largest benefit came from suppressing warnings during high-intensity storms that followed dry antecedent conditions. In those storms, rainfall exceedance alone crossed the empirical threshold, but soil saturation and pore-pressure change remained low. The fused model kept those cases at Level 1 or below. Conversely, two moderate-intensity storms crossed Level 2 because wetness percentiles were above 0.92 and pore pressure rose steadily despite rainfall intensity staying below the historical threshold. One of those storms produced a verified hollow slide. This behaviour supports the hydrological-threshold argument that rainfall should be interpreted together with slope state [6,17].

Uncertainty bands were operationally important. In five storms, posterior probabilities straddled the Level 2 boundary because one gateway was offline and pore-pressure evidence was missing for part of Alto Sul. In retrospective scoring, deterministic imputation would have generated two additional false alarms. The adopted system instead displayed a wide uncertainty interval and requested field review. This is a small but meaningful distinction: sensor networks can improve warning only when missing data and calibration uncertainty are visible to decision makers.

Discussion

The study supports low-cost distributed sensing as a complement to rainfall thresholds, not a replacement. The network improved warning specificity because it observed whether rainfall was actually being converted into wetness, pore-pressure rise, and shallow deformation. This is precisely the information that regional rainfall thresholds lack. At the same time, the network did not eliminate the need for empirical thresholds, maintenance records, and geomorphological mapping. The strongest model used both rainfall exceedance and local hydrological state.

Cost reduction changed the feasible spatial density but introduced its own risks. Low-cost sensors drift, fail, and require calibration. A USD 118 node that is poorly installed or unmaintained can be worse than no node if its data are trusted blindly. The most valuable design decision was not the specific sensor package; it was redundancy across rainfall, wetness, pore pressure, tilt, and field inspection. Recent open-hardware and wireless-sensor studies make clear that distributed sensing is now technically feasible [13,15,16], but operational reliability still depends on power, enclosures, data validation, and local maintenance.

The warning model should not be transferred without recalibration. The fitted coefficients reflect local soils, road drainage, vegetation, and storm climatology. A catchment with deeper weathering profiles or dominant debris-flow channels would require different predictors and thresholds. The event count is also small: eight failures are enough to compare warning logic, but not enough to claim stable probability calibration for rare high-consequence events. Future deployments should pool multi-site data while preserving site-specific random effects, so that sparse local inventories can borrow strength without erasing local hydrology.

Finally, the social side of warning is outside the strongest part of this study. We evaluated technical warning states and duty-officer records, not household response or evacuation behaviour. A technically skilful warning system can still fail if messages are unclear, if road-closure authority is ambiguous, or if false alarms erode trust. The results therefore address the monitoring and forecasting components of early warning more directly than the response component.

Conclusion

A three-wet-season deployment of low-cost distributed sensors improved site-specific early warning for rainfall-triggered shallow landslides in two monitored road-cut catchments. Compared with a rainfall-only intensity-duration threshold, the fused rainfall-wetness-pore-pressure-tilt model reduced false alarms while maintaining detection of seven of eight verified failures. Median warning lead time exceeded four hours for failures with gradual pore-pressure response but remained short for rapid failures in thin colluvium.

The main lesson is practical: distributed sensing is most useful when it is treated as uncertain hydrological evidence and combined with rainfall thresholds, geomorphological mapping, and human review. Low cost permits density, but density does not remove the need for calibration, maintenance, uncertainty reporting, and local decision protocols.

Data and code availability

Calibrated sensor time series, rainfall aggregates, verified landslide polygons, warning-threshold notebooks, posterior samples, and firmware hashes are included in the supplementary archive. Raw packet logs are provided after removal of gateway identifiers. The threshold analysis was run with Python 3.10, pandas 1.5, PyMC 4.1, ArviZ 0.13, and rasterio 1.3.