Machine Learning Maps Debris Flow Danger Along a Chinese Expressway

When Typhoon Butterfly swept across southern China in June 2025, four debris flows erupted in a single day along the Lianzhou-Sanshui section of the G55 Erenhot-Guangzhou Expressway in Guangdong Province, endangering one of the region's most important transport corridors. The event exposed a blind spot in Chinese hazard research: western Guangdong, blanketed in dense vegetation and historically quiet, had attracted little debris flow study, leaving engineers with assessment tools poorly suited to the terrain. A new study in Environmental Earth Sciences now fills part of that gap, combining field surveys, satellite imagery and four machine learning algorithms to map where debris flows are most likely to strike along nearly 300 kilometres of expressway.

The research team, led by Xibei Jia and colleagues, focused on the Lianzhou to Sanshui section of the G55, together with the Lianzhou Connection Expressway, the S60 Lianshan-Hezhou Expressway and the G78 Shantou-Kunming Expressway, a combined network of roughly 323 kilometres linking Guangdong with Hunan and Guangxi. The corridor crosses the southern edge of the Nanling Mountains and the plain-hill area of the Pearl River Delta, with elevations ranging from 100 to 500 metres. From northwest to southeast, the landforms alternate in a regular sequence: steep, deeply dissected low and medium mountains; low, gentle hills threaded with dense gullies; and finally flat erosion-abrasion terraces and alluvial plains covered in thick, loose deltaic deposits.

This gradient allowed the researchers to divide the route into a steeper northern section and a gentler southern section, a distinction that would prove central to their findings. The climate amplifies the danger. The region has a mid-latitude subtropical monsoon climate in which roughly 85 percent of annual precipitation falls between March and September, with June the wettest month.

Typhoons strike from July to September, with instantaneous wind speeds reaching 35 metres per second, and the rainstorms that accompany them routinely trigger mountain torrents and landslides. When the June 17, 2025 disaster struck, six meteorological stations along the expressway recorded 72-hour rainfall totals of 99.5, 164.3, 147.8, 377.7, 186.2 and 95.9 millimetres. The wide spread among these figures revealed how spatially heterogeneous the storm was, concentrating extreme loading on some slopes while sparing others only kilometres away.

To build a reliable picture of the hazard, the team identified 237 potential debris flow gullies along the route by integrating satellite imagery, unmanned aerial vehicle surveys and field investigations. Potential gullies were defined as those with favourable geomorphic and material conditions for debris flow initiation, such as steep channels, abundant loose deposits, depositional fans or direct connections to expressway drainage structures. The researchers then separated true debris flow gullies from the rest using multiple lines of evidence: gullies that erupted on June 17, 2025, gullies showing historical debris flow deposits revealed by on-site trenching, including poorly sorted gravel-clay mixtures, mud marks and depositional features clearly distinct from ordinary fluvial or residual slope deposits, and, as auxiliary evidence, interviews with nearby villagers.

In the end, 33 gullies were confirmed as debris flow gullies, a positive rate of just 13.92 percent that made the dataset highly imbalanced and posed a challenge for any classification algorithm. The team acknowledged an important limitation in this inventory. Some old debris flow deposits may have been eroded, buried by later fluvial sediments, covered by vegetation or disturbed by engineering activity, so the gullies classified as non-debris-flow features should be understood as lacking identifiable evidence during the present investigation rather than as incapable of producing future flows.

From five categories of variables, topography, geological structure, lithology, vegetation and meteorological-hydrological conditions, the researchers selected twelve impact factors: elevation difference, longitudinal gradation, bending coefficient, catchment area, slipperiness of weathering product, regional fault influence, side-slope angle, vegetation type, vegetation coverage, unstable slope, annual maximum hourly rainfall and annual maximum 24-hour rainfall. Unlike the subjective expert scoring embedded in current Chinese standards, each factor was extracted from traceable sources, including 1:500 topographic maps surveyed in the field, regional geological and tectonic maps, engineering survey data along the expressway and meteorological records. Multicollinearity checks using the Pearson correlation coefficient, variance inflation factor, principal component analysis and information gain ratio confirmed that all twelve factors could be retained: correlation coefficients ranged from -0.29 to 0.62, all variance inflation factors stayed below 3, and the first eight principal components captured more than 80 percent of the information with relatively balanced factor weights between 0.1573 and 0.2475.

The most novel methodological move was the introduction of slipperiness of weathering product as a replacement for conventional lithological classification. Sieve analyses of solid material at gully mouths showed that the debris was dominated by sediment, with few stones larger than 40 millimetres, and field investigations found that the flows were typically initiated by shallow collapses of one to three metres of surface soil that cascaded into landslide-debris flow chains. Under these conditions, the standard fourfold lithology scheme, built around bulk rock strength, offered little discriminatory power.

The new factor instead grades surface weathering products into three classes: highly slippery materials such as mudstone and granite weathering products, which soften, disintegrate and turn muddy on wetting with a sharp loss of strength; moderately slippery interbedded sandstone-mudstone products prone to softening and swelling; and low-slipperiness products of dense sandstone and limestone, which disintegrate into coarse, permeable residual soils that resist pore pressure buildup and supply little material. Four algorithms were then trained to classify the 237 gullies: a back propagation neural network, random forest, support vector machine and extreme gradient boosting, known as XGBoost. Because debris flow gullies made up fewer than 14 percent of the sample, the team applied the synthetic minority oversampling technique to boost the weight of positive cases, and wrapped normalization, oversampling and classification into a single pipeline fitted only on training data to prevent information leakage.

Crucially, instead of random train-test splitting, which can produce overoptimistic scores when nearby gullies share similar conditions, the route was divided into three spatially independent folds of 11 positive samples each, with each fold held out once for testing while the other two trained the model. Hyperparameters were re-optimized within each training round, and classification thresholds were set using training data alone before being applied to the held-out section. The results were decisive.

XGBoost achieved the best overall performance, with a mean accuracy of 0.923, precision of 0.725, recall of 0.879, F1-score of 0.794, ROC-AUC of 0.927 and PR-AUC of 0.859, comfortably ahead of the back propagation neural network (accuracy 0.871, ROC-AUC 0.842), random forest (0.775 and 0.753) and the support vector machine (0.693 and 0.783). The support vector machine illustrated a classic trap in imbalanced classification: it matched XGBoost's recall of 0.879 and even reached a perfect 1.000 in the first fold, but it flagged 16, 21 and 19 negative gullies as debris flow sites across the three folds, driving its precision down to 0.343 and its PR-AUC to 0.484. XGBoost maintained high recall while keeping false positives in check.

Its susceptibility map also matched field reality best, concentrating medium and high susceptibility zones, 48.33 percent and 11.42 percent of the corridor respectively, in the northern mountainous section, where granite residual soils, eucalyptus plantations and steep terrain dominate, while the southern alluvial plains, underlain by permeable conglomerate and sandstone and cloaked in deep-rooted forest with coverage above 90 percent, emerged as generally low susceptibility. Beneath the mapping exercise lies a coherent formation mechanism. Multi-phase tectonic activity, spanning the Caledonian, Hercynian-Indosinian, Yanshanian and Himalayan periods, fractured the bedrock and shaped the steep topography; the 33 known debris flow gullies cluster at longitudinal gradients of 200 to 400 per mille, steep enough to mobilize loose material while sustaining erosional transport.

Material supply is governed by water-sensitive weathering products, fault-fracture zones rich in loose deposits and the region's distinctive eucalyptus plantations, which grow to harvest in only five to seven years, transpire strongly, crack the soil and offer shallow roots that reinforce slopes poorly. Rainfall then acts as the trigger, softening slope materials by infiltration and scouring them by runoff, while the V-shaped gullies typical of the area funnel water into a flashy hydrological response of rapid rise and high velocity. The authors are careful about statistical weight: Fisher's exact tests for slipperiness of weathering product, regional fault influence and vegetation type yielded p-values of 0.194, 0.634 and 0.203, none reaching the conventional 0.05 threshold, so these are presented as potential conditioning factors consistent with field observations rather than proven drivers, pending validation on larger datasets.

Subject of Research: Machine learning-based debris flow susceptibility assessment along the G55 expressway in Guangdong Province, China

Article Title: Debris flow susceptibility assessment along the G55 expressway in Guangdong Province, China

Article References: Jia, X., Liu, Y., Liu, B., Ren, W., & Qin, L. (2026). Debris flow susceptibility assessment along the G55 expressway in Guangdong Province, China. Environmental Earth Sciences, 85(17), Article 443. https://doi.org/10.1007/s12665-026-13191-x

Image Credits: AI Generated

DOI: 10.1007/s12665-026-13191-x

Keywords: debris flow, susceptibility mapping, machine learning, XGBoost, Guangdong, expressway, landslide, rainfall trigger, weathering products, spatial cross-validation, typhoon, geological hazard