When the railway is a floodplain

Two orals, one problem: water on Thai railway tracks is not a binary GIS layer. It is an operating rule.

• ICCBEI 2023 (oral): Risk assessment of railway tracks in floodplain area using digital surface model and computer vision (paper PDF, slides).
• ICRT 2024 (oral): Integrating digital surface model for flood risk assessment of railway networks in Thailand.

Project writeup, figures, poster: Railway flood risk mapping. The Elsevier chapter on track management under extreme events is in press; this post stays with the orals.

Question

Embankments block flow. Monsoon does the rest. The useful output is not a red polygon. It is: at this kilometre, at this water depth, do we run, crawl, or close.

Data and method

Inputs we actually used:

• digital surface models for elevation and the track in the terrain
• climate records 2005–2016 (twelve years)
• drainage-basin descriptors, slope, elevation
• computer vision on the surface where the track sits in the floodplain

Four risk-factor families: climate, basin, slope, elevation. Most location-years do not flood, so a classifier trained on raw counts learns to say zero. We used ADASYN on the minority flooded cases.

The label is an operation-based Flood Risk Index, tied to water on the structure:

FRI	Water	Operations
0	none	normal
1	0–30 cm (ballast to sleeper bottom)	caution
2	31–60 cm (sleeper to railhead)	reduced speed
3	61–70 cm (over railhead)	minimal operations
4	>70 cm	no operations

Several classifiers were tried. Extra Trees was the one we kept. On the project evaluation: 0.98 accuracy on training, 0.71 F1 on the 2015–2016 test years. That drop is the honest result. Rainfall features dominated importance: average rainfall, average rainfall per day, maximum rainfall per day.

The maps in the project page run north–south along the network, ground in green, rail in brown, water in blue. I am not regenerating them here.

What I would not do again

Training accuracy of 0.98 next to test F1 of 0.71 is a split and a class problem, not a victory lap. I would lead with the test F1 and with FRI-class recall, especially FRI 3–4 — those are the closures. I would also stop calling rainfall “the most important factor” as if that were a surprise; the contribution is translating rainfall and terrain into centimetres the operator already uses.

Same method rule as buckling: the index has to be a sentence in the control room.