Unreported sites are censored, not negative
Most sinkhole susceptibility studies frame the task as binary classification over observed sinkholes and sampled non-sinkhole locations. The paper argues this introduces label inaccuracy: many reported depressions are not field-verified sinkholes, and many susceptible sites go unreported because of reporting bias, limited monitoring, or incomplete records. Treating an unreported location as a confirmed negative therefore encodes an assumption the data cannot support.
The proposed framework instead represents each site as either having a reported subsidence or being right-censored. Using calendar year as the timescale, the observed time is the minimum of the first reported subsidence year and the final observation year, and the event indicator marks whether a subsidence was actually reported. Censored sites stay in the risk set up to their censoring year, contributing the information that no event occurred by then rather than being treated as hard negatives. This is what makes the output continuous and time-aware instead of a static class label.
Three modalities enter through separate encoders
The factors governing sinkhole formation span heterogeneous data modalities that differ in resolution, temporal cadence, and scale. The model keeps them separate rather than flattening them into one feature vector. A multilayer perceptron encodes tabular attributes such as lithology, soil type, and hydraulic properties. A convolutional network encodes spatial patches of raster layers around each site. A recurrent network encodes annual precipitation summaries as a multivariate sequence up to the prediction year.
The three embeddings plus a spatial context term are combined by a fusion module implemented as concatenation followed by a linear projection, and a final multilayer perceptron maps the fused representation to a log-relative hazard. The full hazard is the baseline hazard multiplied by the exponential of that term. Parameters are estimated by minimizing the negative Cox partial likelihood with Breslow ties, so only sites with observed events contribute to the sum over events.
Susceptibility is a calibrated 10-year probability
After training, the baseline cumulative hazard is estimated from training sites using the Breslow estimator. For any spatial unit, the survival probability over a K-year horizon starting from year t is computed from the difference in cumulative hazard across that horizon, and the raw complementary incidence probability is one minus that survival probability.
The paper notes that because subsidence events are rare, this raw probability is poorly calibrated in absolute terms: the network orders sites well, but the raw magnitudes do not match observed event frequencies. A monotone isotonic calibration map, fit on validation windows with the pool-adjacent-violators algorithm, is therefore composed with the model output. This step removes the calibration error at negligible cost to discrimination and lets the outputs be read as absolute 10-year probabilities.
The county split is the more defensible estimate
The study evaluates under two holdout regimes. The random split assigns individual sliding windows to train, validation, and test uniformly at random, which measures interpolation: because a site's windows overlap heavily in their 10-year inputs, nearby windows of one site can appear in both training and test. The county, or geologic-unit, split assigns whole geologic units to a single fold, so held-out sites are spatially disjoint from training and the result measures generalization to unseen terrain.
The pretrained multimodal model reaches a C-index of 0.780 on the random split versus 0.703 on the county split. The paper attributes the 0.077 gap to interpolation between near-duplicate windows of a site, since adjacent windows share 9 of 10 observation years, and states that the county number is the more defensible estimate for unseen terrain. ROC-AUC is identical across regimes at 0.899, which locates the gap in censoring-aware time-concordance rather than in binary discrimination.
Fusion with precipitation drives the gain
Fusing the static modalities with precipitation is the primary driver of performance, improving both C-index and ROC-AUC relative to the time-series-only baseline. Under the county split, the time-series-only model reaches a C-index of 0.696 and ROC-AUC of 0.888, while the multimodal model reaches 0.695 and 0.904. Under the random split, the time-series-only model reaches 0.743 and 0.879, while the multimodal model reaches 0.780 and 0.899.
The calibrated risk assigned to confirmed sinkholes withheld entirely from training rises to roughly 7 times background. The paper reports a lift of 7.1 times under the random split and 6.9 times under the county split for the multimodal model, compared with 4.4 times and 6.4 times for the time-series-only model. That these unseen sinkholes still score high, and concentrate in the west-central karst belt rather than being uniformly raised, is presented as evidence that the model captures genuine susceptibility rather than memorized locations, consistent with subsidence being controlled by subsurface lithology that precipitation alone cannot localize.
Pretraining adds no consistent benefit
The framework includes an autoencoder pretraining stage in which each modality encoder is trained to reconstruct its input before the survival stage, with the decoder discarded afterward. The ablation results show this yields no consistent gain. Under the random split, the pretrained multimodal model matches the non-pretrained model on C-index and ROC-AUC at 0.780 and 0.899, while lift is slightly lower at 7.0 times versus 7.1 times. Under the county split, pretraining raises C-index from 0.695 to 0.703 but leaves ROC-AUC essentially unchanged and lift slightly lower at 6.8 times versus 6.9 times.
The paper's explanation is that fine-tuning recovers equivalent encoders regardless of initialization. This matters for readers comparing architectures: the reported benefit of multimodal integration comes from combining modalities, not from the pretraining stage, and the county-split C-index difference of 0.008 is small relative to the 0.077 gap between split regimes.
What the statewide map does and does not show
The study covers Florida and combines three public datasets: subsidence incident locations from the Florida Geological Survey, annual precipitation from PRISM for 1895 to 2025, and the hydraulic-head difference between the surficial water table and the Floridan Aquifer System. After screening the subsidence records, 6,302 sinkhole and non-sinkhole sites were retained. Each site was converted into annual sliding samples using a 10-year precipitation observation window followed by a 10-year prediction horizon, producing 557,946 samples with a low event rate.
Scored on each site's most recent decade, the resulting statewide map concentrates high susceptibility along the west-central karst belt, coincident with the confirmed-sinkhole distribution without being fit to those labels. The paper's claim is that the method effectively ranks sinkhole risk and captures spatial variation in occurrence. The map is a susceptibility surface over a fixed 10-year horizon, not a dated forecast of individual collapses, and the county split remains the relevant benchmark for reading it on terrain the model has not seen.
Where the county split number should anchor judgment
The most decision-relevant comparison in this paper is not multimodal versus time-series-only in isolation, but which split regime a reader trusts. The random split's 0.780 C-index is inflated by overlapping windows of the same site, and the paper itself identifies the county split's 0.703 as the more defensible estimate for unseen terrain. Any assessment of whether this framework is ready for a new region should therefore start from the spatially blocked number, not the interpolation number.
The second priority is the calibration step. The raw Cox probabilities rank well but are miscalibrated, and the isotonic map is what converts them into absolute 10-year probabilities; without it, the ranking is still usable but the magnitudes are not. The third is the ablation result: since encoder pretraining adds no consistent gain, the value of this framework rests on modality-specific encoders plus cross-modal fusion, and the lift of roughly 7 times background on confirmed sinkholes withheld from training is the figure that best summarizes what that fusion buys.
Sources used in this answer
Multimodal Deep Survival Analysis for Sinkhole Susceptibility
A multimodal Cox proportional hazards framework that treats unreported locations as right-censored rather than negative reaches a C-index of 0.703 on a spatially blocked county split and 0.780 on a random split, with ROC-AUC of 0.899 under both regimes, and assigns confirmed sinkholes withheld from training roughly 7 times background calibrated risk.