Skip to main content
Back to timeline
arXivSource publication:

Landslide inventories record observers as well as hazards: splitting occurrence from recording pays off on one earthquake, not all

Synopsis

Landslide susceptibility maps learn from inventories that are systematically biased, because landslides near roads and towns are more likely to be reported, so a classifier absorbs the observation process into the hazard signal. This work adapts ecological occupancy models to landslide inventories, treating several independent inventories of one event as repeated surveys and modelling true occurrence and recording probability separately, with a small exhaustively mapped anchor area to set the scale. Under simulated reporting bias in three earthquakes it reduces road bias and estimates the overall landslide rate more stably than standard practice, but its ranking advantage over neutralisation appears on only one earthquake, Iburi.

Source-provided article image: Occupancy Modelling of Biased Landslide Inventories: When Does Separating Occurrence from Recording Improve on Neutralisation?
arXiv

Interpretation

Several independent landslide inventories of the same event are treated as repeated surveys of the same terrain, so the model can learn separately whether a landslide truly occurred in a cell and whether that cell was recorded. Earlier practice fits hazard and recording bias inside one model and then removes the influence of bias variables such as roads and built-up density when mapping, which is neutralisation; it needs only one inventory but assumes the bias variables can be separated from the hazard signal within a single model. On real terrain, roads and near-complete reference inventories from three earthquakes with simulated reporting bias added, the occupancy model has lower road bias than standard practice in 24 of 25 paired draws across all settings.

The two processes are trained through one joint likelihood: a susceptibility network sees only terrain features and gives the probability that a landslide occurred, while a detection network sees only observation features such as road distance, built-up density and cloud, and gives the probability that the cell was recorded; only the first is used for mapping. Detection is constrained to be monotone, so being farther from a road or under more cloud can only lower the recording probability and denser built-up land can only raise it, with each standardised feature entering through a sum of increasing sigmoids with non-negative weights. On Iburi, monotone detection raised remote-quarter AUC from 0.58 to 0.72, the largest effect among the design choices, and the learned detection direction correlated with the true rules at 0.81 for the news inventory, 0.80 for the agency inventory and 0.56 for the satellite inventory.

Perspective

The results concern event-conditional susceptibility mapping for three earthquakes: the maps describe where landslides occurred in that event and are not forecasts for a future earthquake. In the simulated scenarios the recording bias is generated by known rules, and the real-inventory test uses only three satellite-mapped Gorkha inventories.

Road and population reporting bias and cloud cover are simulated, and in scenario A the detection features match the simulated rules closely, which favours the occupancy model. Only Haiti was out of sample, and it tested only the negative side of the rule; there is no out-of-sample test of a high gap. Draws within one earthquake share terrain and truth, so pooled statistics overstate generality, and the reference inventories are treated as truth although they are themselves imperfect.