Social-media and newspaper reports reveal large-scale meteorological drivers of floods on Sumatra

The study addresses the question of which large-scale atmospheric drivers are most responsible for floods in Sumatra and whether these drivers can provide predictive skill. Sumatra, situated in the Maritime Continent where precipitation is among the world’s highest, experiences frequent and damaging floods that affect millions. Forecasting is challenged by incomplete flood records and reliance on short-lead local observations. Prior work highlights the strong diurnal cycle of convection and modulation by intraseasonal and equatorial wave phenomena such as the Madden–Julian Oscillation (MJO) and convectively coupled Kelvin waves (CCKWs). The purpose is to construct a more complete multi-year flood record from governmental and crowd-sourced data and to attribute flood events to large-scale meteorological phenomena, thereby assessing potential predictability and improving early warning capabilities.

Background literature indicates: (1) global and Indonesian flood vulnerability and increasing risk under climate change; (2) the Maritime Continent’s extreme precipitation and strong diurnal convection; (3) the MJO as a dominant intraseasonal (30–90 day) convective envelope propagating eastward (~5 m/s), modulating local precipitation; (4) CCKWs as faster (~12 m/s) equatorially trapped, convectively coupled disturbances that strongly modulate rainfall over the Maritime Continent and elsewhere; (5) additional drivers such as Boreal Summer Intraseasonal Oscillation and South China Sea cold surges, especially when interacting with the MJO; (6) prior recognition of CCKWs’ influence over equatorial Africa and the Maritime Continent, including phase locking with the diurnal cycle; and (7) the emerging utility of social media and local news data for flood detection and spatiotemporal characterization in Indonesia. Collectively, prior studies suggest multi-scale interactions where CCKWs embedded within or independent of the MJO can enhance precipitation, but a direct, event-based linkage to floods had not been systematically established for Sumatra.

Data and period: 2014–2018. Three independent flood datasets: (1) Twitter-derived events, (2) local newspapers, and (3) Indonesian National Board for Disaster Management (BNPB). Space-time aggregation was used to align events with large-scale meteorological features. - Twitter processing: Tweets containing flood-related Indonesian keywords (e.g., “banjir”) were collected and geoparsed to Sumatra using the TAGGS algorithm, leveraging user metadata and text-referenced locations. Tweets from >100 geotagged points were aggregated into 10 regions corresponding to administrative provinces, with Aceh and North Sumatra split into north/south subregions along the mountain range, yielding 10 regions total. Time series were binned to 3-hourly intervals, low-pass filtered with a 3-day running mean, and flood events were detected as local maxima exceeding a region-specific threshold (85th percentile of a 30-day low-pass filtered activity distribution). Multiple maxima within 3 days were merged into a single event; termination was the subsequent local minimum. - Newspapers and BNPB: Local newspaper reports were compiled and binned to the same regions; BNPB provided official flood start dates and impacts. Both were subjected to the same 3-day re-occurrence merging rule for consistency. - Meteorological datasets: Satellite precipitation from TRMM/GPM 3B42v7 (3-hourly, 0.25°×0.25°) used to compute daily rates and 5-day accumulations; ERA5 850 hPa zonal wind anomalies (hourly, averaged to 3-hourly) with seasonal cycle removed (first three harmonics over 2000–2018); MJO state characterized by RMM index (active when amplitude >1, phases 2–3 over Indian Ocean, 4–5 over Maritime Continent). Limited in-situ rain gauges (29 stations in Sumatra; case analysis uses 3 near Padang) from BMKG dataonline portal. - CCKW diagnostics: Precipitation anomalies associated with CCKWs were isolated via wavenumber–frequency Fourier filtering applied to TRMM precipitation. Trajectories were computed within 2.5°S–2.5°N by tracking space-time continuous anomalies exceeding 2.5 mm/day. A “robust” CCKW was defined as one exhibiting a well-defined equatorial trajectory between 80°E and 110°E (indicating initiation at least ~2 days upstream) with anomaly ≥2.5 mm/day; “strong” CCKWs exceeded 5 mm/day. For synoptic context, CCKW activity at 90°E was used to assess association with flood onsets. - Event aggregation: Regional flood detections across Sumatra were combined into “flooding periods” defined as continuous sequences of floods occurring anywhere on the island. Seasonal precipitation favorability was defined from a 90-day low-pass filtered area-averaged 5-day accumulation exceeding the 5-year mean. - Caveats addressed: TRMM tends to underestimate over steep orography; rain gauge data contain 20–40% gaps; geoparsing uncertainties mitigated by large sample sizes and event-level aggregation.

- Coverage and consistency: Across 2014–2018, regional flood counts were 493 (Twitter), 453 (newspapers), and 445 (BNPB), aggregating into 254, 221, and 189 island-wide flooding periods, respectively. Seasonal cycles were broadly consistent across datasets despite regional biases. - Dominant role of CCKWs: Over 91% of flooding period onsets were associated with CCKW activity at 90°E (precipitation anomaly >2.5 mm/day), and over 63% began during strong CCKWs (>5 mm/day). At least 34% of onsets were associated with robust CCKW trajectories (80°E–110°E). About 45% (66–79 out of 160) of robust CCKW trajectories were associated with flooding periods on Sumatra, and about 60% of floods were immediately preceded by a robust CCKW event. - Limited direct MJO association: Only 28–34% of flooding periods occurred during MJO phases 2–5 (favorable for Maritime Continent precipitation), and less than 23% began with MJO active over the Indian Ocean (phases 2–3). About 34–46% occurred with weak/no MJO (amplitude <1). Many floods during MJO-active times occurred when the convective envelope was away from the Maritime Continent (phases 6–1), implying local suppression. - Seasonal patterns: Northern Sumatra regions peak in boreal fall/winter; southern regions in winter/spring; central regions in fall/spring; no maxima in boreal summer, consistent with ITCZ migration. - Case study (Padang, 29 May–1 June 2017): A robust, strong CCKW, originating over East Africa ~5 days earlier, passed Sumatra while MJO was active over the Indian Ocean. CCKW precipitation anomalies exceeded 10 mm/day over Sumatra, with 5-day accumulations >100 mm near the island. Strong low-level westerlies, enhanced by Tropical Cyclone Mora and an equatorial cyclonic circulation, likely intensified orographic rainfall. The event underscores that CCKWs, embedded within or outside MJO envelopes, can trigger floods.

Linking event-based flood records to large-scale dynamics shows CCKWs as the primary dynamical precursor of Sumatran floods, far outweighing direct MJO phase effects at flood onset. Because robust CCKWs are trackable and typically identifiable at least two days before reaching Sumatra, they offer actionable predictability for flood risk, especially when combined with knowledge of the diurnal cycle and local topography. The synergy of three independent flood datasets (Twitter, newspapers, BNPB) enhances reliability, with consistent seasonal signals and concordance during major events despite differing regional sensitivities. The Padang case exemplifies how CCKWs, sometimes augmented by off-equatorial vortices or tropical cyclones, can rapidly raise 5-day accumulations and overwhelm local catchments. These findings advocate incorporating tropical wave diagnostics (CCKW tracking and MJO monitoring) into operational early warning to extend lead times beyond local gauge/radar detection.

This study constructs a five-year flood record for Sumatra by integrating governmental reports with crowd-sourced Twitter and local newspaper data, then attributes flood events to large-scale meteorological drivers using satellite precipitation, reanalysis winds, and equatorial wave diagnostics. The key contribution is demonstrating that convectively coupled Kelvin waves are associated with over 90% of flood onsets and constitute the principal dynamical precursor, while favorable MJO conditions alone explain a much smaller fraction. Operationally, tracking robust CCKWs provides 2–5 days of additional predictability for flood risk in Sumatra and potentially across the Maritime Continent. Future research should: (1) extend analyses over longer periods and broader regions; (2) improve detection of off-equatorial (dynamical equator) Kelvin waves; (3) examine interactions among CCKWs, MJO, equatorial Rossby waves, and topographically induced vortices; and (4) advance numerical prediction of CCKWs to enhance subseasonal-to-synoptic lead times.

- Observation and sampling: The analysis period (2014–2018) is relatively short. Rain gauge data have substantial gaps (20–40%). TRMM 3B42v7 underestimates precipitation over steep orography, implying lower-bound estimates of accumulation and CCKW anomalies. - Event detection: Social media geoparsing and thresholding introduce uncertainties and regional biases (internet access, media usage trends), though mitigated by aggregation and cross-validation with newspapers and BNPB. - Dynamical attribution: CCKW trajectory tracking is confined to a narrow equatorial band (2.5°S–2.5°N), potentially missing off-equatorial/dynamical-equator waves. The study emphasizes flood initiation rather than persistence or amplification phases, likely underestimating CCKW influence. - Generalizability: Findings pertain to Sumatra’s specific geography and multi-scale convection; transferability to other regions requires similar social media density and validation.