Satellites and aircraft now map forests down to individual tree crowns, and repeat surveys could in principle tell us when each tree dies. That would turn biomass mapping from a single snapshot into a measure of carbon turnover. The natural statistical tool is survival analysis: a proportional hazards model estimates how covariates such as height, crowding or drought exposure multiply the instantaneous risk of death, and it correctly handles trees still alive when observation ends. There is a catch that neither literature has faced. From above, a tree can vanish from the record for two reasons. It may have died, or the sensor may no longer see it because it is shaded by neighbours, lost in a saturated dense canopy, or simply missed by the crown detector. The trees most likely to vanish for the second reason are small, suppressed understory trees, and ground data show these are also the trees most likely to die. The standard proportional hazards model assumes that dropping out of observation is unrelated to risk. Remote sensing breaks this assumption, and the violation grows with exactly the structural heterogeneity that the biomass literature already struggles with.
We propose a grouped-time proportional hazards model that separates true death from non-detection. The chance that a living crown is seen in a given survey is modeled as a function of canopy structure. The model is anchored by a small set of field-remeasured trees, and it is fit to crowns tracked across repeat airborne surveys. The claim is that this correction changes the ecological conclusions, not just the fit statistics. Mortality rates and hazard ratios estimated from imagery should be biased in proportion to canopy complexity when detection is ignored, and should agree with field mortality once detection is modeled.
Core Claim
Two open problems in the file turn out to be the same problem seen from two sides. Biomass remote sensing lists "the missing understory" and residual saturation in dense canopies as known limits. Proportional hazards modeling lists informative censoring as a known way its estimators become biased. When crowns are tracked across repeat surveys, occlusion and saturation become informative censoring. The detector's error rate is not small: the DeepForest pipeline behind NEON Crowns reports 72.4% recall at 50% overlap, so on the order of a quarter of crowns are missed in any single year. Those misses concentrate in lower canopy layers. Developing Tree Mortality Models Using Bayesian Modeling Approach fits multilevel logistic models to field larch plots and finds that DL, the competitive load from larger neighbours, is positively associated with mortality. The same trait that raises a tree's risk of dying therefore also lowers its chance of being seen.
A naive analysis has two options, and both fail. If a missing crown is scored as dead, mortality is inflated for small trees and structurally complex stands. If the crown is treated as censored at the survey where it was last seen, the highest-risk trees are quietly removed. We expect the resulting bias in stand-level mortality to follow the structural-complexity ranking that Improving Aboveground Biomass Estimation in Lowland Tropical Forests across Aspect and Age Stratification observed for biomass error, where accuracy degraded from homogeneous pine to mixed broadleaf.
Research Program
The unit of analysis is an individual crown tracked across annual airborne surveys at National Ecological Observatory Network (NEON) sites. These surveys supply co-registered RGB imagery and LiDAR canopy height models. We will apply the released DeepForest detector from NEON Crowns to each survey year. Crowns are matched across years by overlap and by consistency of the height model. Each crown then has a history: seen, not seen, or seen with a large height drop. The 1.73 m height error reported for that pipeline sets the threshold for calling a height drop real rather than noise.
The model has two linked parts. The death part is a grouped-time proportional hazards model, which is the discrete-interval form of the Cox model and handles irregular survey spacing without assuming a shape for risk over time:
$$P(\text{crown } i \text{ dies in interval } k \mid \text{alive at start}) = 1 - \exp\left(-\exp(\alpha_{g,k} + \beta^T x_i + u_s)\right)$$
In this expression, $\alpha_{g,k}$ is a baseline log-hazard for interval $k$ that is allowed to differ by stratum $g$ (forest type crossed with stand age class), $x_i$ are crown covariates (height, crown area, height relative to neighbours, local LiDAR height percentiles, spectral stress indices), $\beta$ are shared log hazard ratios, and $u_s$ is a site-level random effect, or frailty, that absorbs unmeasured site conditions. Stratifying the baseline while sharing $\beta$ across continuous structure addresses the tension in the file directly. The file contrasts discrete stratification, as in the Xishuangbanna study, with continuous LiDAR structure, as in Mapping Forest Aboveground Biomass Using Multisource Remotely Sensed Data and the GEDI deep-ensemble work. Here the strata set when risk is high and the structure sets who is at risk. The frailty term follows the random-effect proportional hazards models in Robust inference for univariate proportional hazards frailty regression models and Consistency in a Proportional Hazards Model Incorporating a Random Effect.
The detection part is a logistic model for $p(z_{ik})$, the probability that a living crown is seen in survey $k$. Here $z$ describes occlusion: rank in the local height distribution, canopy closure, and within-window height variance. Detection and death are identified from two sources. The first is crowns that go missing and then reappear, which proves they were alive while unseen. The second is a calibration subset of crowns matched to mapped, remeasured field stems whose true status is known. We will fit the joint likelihood by marginalizing over the unknown alive/dead state of unseen crowns, which amounts to a two-state hidden Markov chain per crown. Fast risk-set computations of the kind described in FastSurvival keep this tractable at millions of crowns.
The final output is a biomass-loss rate for each stratum. Each crown's per-interval death probability is weighted by its allometric biomass, estimated from height and crown area, and the weighted probabilities are summed.
Baselines
The do-nothing reference is the stratum-level mortality rate computed directly from imagery, with any crown missing in a later survey counted as dead. This is what a crown-detection map naively implies. It sets the scale of the problem, and we expect it to overstate mortality most in complex broadleaf stands.
The second reference treats disappearance as censoring. A standard stratified Cox model with site frailty is fit to the same crowns, and each crown is censored at its last detection. This is the strongest off-the-shelf survival method a practitioner would reach for. It isolates our key ingredient, because it shares every covariate, stratum and frailty with the full model and differs only in lacking the detection part. Beating it means the full model's stratum mortality rates sit closer to field rates, and that the gap between the two models widens with canopy complexity.
The third reference is a multilevel logistic mortality model over the full survey window, of the kind in Developing Tree Mortality Models Using Bayesian Modeling Approach. That paper reports an AUC of 0.832 on field-measured larch plots. It is not directly comparable, because it uses field data from a different forest. We therefore reimplement the model on our crowns with identical covariates and a two-level random intercept.
For flexible learners, we will run a random survival forest and a DeepSurv network (Deep Survival; on METABRIC it reports a C-index of 0.643 against 0.630 for linear Cox and 0.624 for random survival forests) on the same naively censored data. Each learner gets the same feature set and an equal hyperparameter search budget, with site-blocked cross-validation for all methods. We found no reported result on Lacuna for crown-level mortality hazard models driven by remote sensing, so every number in the comparison will be produced by us under one protocol.
Why not just…?
Why not just use a more flexible survival learner such as a random survival forest or DeepSurv? Flexibility addresses a misshapen risk function, not a biased sample. If suppressed trees are missing from the risk set because they are unseen, any learner trained on that risk set inherits the bias. We test this directly by running these learners on naively censored data. If they match the detection-aware model on agreement with field mortality, our claim fails and we will say so.
Why not just improve the crown detector until misses are negligible? Better detectors help, but occlusion is physical. Airborne optical and LiDAR sensors under-sample shaded layers even with perfect algorithms, which the file lists as an open problem. The strongest tropical detector in SelvaBox reaches only 48.60 on its raster F1 at 75% overlap. We will run an explicit check: rerun the analysis with detections restricted to high-confidence crowns and to all crowns. If detection error drives the bias, the naive estimates should shift with the detection threshold, while the corrected estimates should stay stable.
Why not just use capture–recapture survival models, which ecologists already use to separate death from non-detection? This is the real ancestor of the proposal, and we adopt its logic openly. What is new is the setting and the parameterization. Detection is predicted from sensor-derived canopy structure rather than estimated per animal or per occasion, death is modeled as proportional hazards with stratified baselines and spatial frailty, and the model runs over tens of millions of machine-detected crowns calibrated against field stems. If a plain capture–recapture model with constant detection performs as well, the lesson is that structure-dependent detection is unnecessary. The ablation that fixes $p$ to a constant tests exactly this.
Why not just difference biomass maps between years? Differencing adds the errors of both snapshots together. It also reads saturated dense canopies as stable, as the 550 t/ha ceiling of the optical map in the Madagascar GLAS study illustrates. And it cannot attribute loss to individual trees or covariates. We will still report map-differenced loss as a sanity reference.
Proposed Experiments
The first check is a simulation with known ground truth. We generate synthetic stands with a known hazard, including a true positive effect of suppression, and with detection that falls with canopy rank. Naive scoring should reverse or inflate the suppression effect, censoring at the last sighting should shrink it, and the full model should recover it within its confidence interval. Performance will be reported across a range of detection rates, bracketing the 72% recall reported for NEON Crowns.
The decisive real-data check uses NEON sites where mapped field stems are remeasured. One part of the plots serves as calibration and is held out by site from a validation part. On the validation plots we compare stratum mortality rates and biomass-weighted loss from each method with field-observed rates. The main metrics are absolute error in annual mortality rate and in biomass loss (Mg/ha/yr). Secondary metrics are the time-dependent concordance index and the integrated Brier score computed against field status. The proposal succeeds if two things hold. First, the naive estimators' error rises with a structural-complexity index, namely LiDAR height variance, the variable ranked high in Mapping Forest Aboveground Biomass Using Multisource Remotely Sensed Data. Second, the detection-aware model flattens that trend.
A third check targets hazard ratios. On the calibration plots, the field-estimated effect of relative height on mortality serves as a reference, and we ask which remote-sensing estimator reproduces its sign and magnitude. Finally, we will test whether the proportional hazards assumption holds by interacting covariates with interval indicators across drought and non-drought years. Where it fails, the stratified baselines absorb the time variation, and the result is reported rather than hidden.
References
- NEON Crowns: a remote sensing derived dataset of 100 million individual tree crowns
- Developing Tree Mortality Models Using Bayesian Modeling Approach
- Improving Aboveground Biomass Estimation in Lowland Tropical Forests across Aspect and Age Stratification: A Case Study in Xishuangbanna
- Mapping Forest Aboveground Biomass Using Multisource Remotely Sensed Data
- Global canopy height regression and uncertainty estimation from GEDI LIDAR waveforms with deep ensembles
- Robust inference for univariate proportional hazards frailty regression models
- Consistency in a Proportional Hazards Model Incorporating a Random Effect
- FastSurvival: Hidden Computational Blessings in Training Cox Proportional Hazards Models
- Deep Survival: A Deep Cox Proportional Hazards Network
- SelvaBox: A high‑resolution dataset for tropical tree crown detection
- Interest of Integrating Spaceborne LiDAR Data to Improve the Estimation of Biomass in High Biomass Forested Areas