Satellite carbon accounting now says that thinning of standing forests (selective logging, fire, drought and wind mortality) releases as much or more carbon than outright clearing. The same accounting cannot say who or what did the thinning. A forest knocked down by a storm and a forest selectively logged can lose similar biomass and look alike in coarse microwave data, yet one is a natural disturbance and the other is a human emission that carbon policy must assign to someone. This proposal starts from a simple physical observation. Downed stems, felling gaps and logging trails are elongated objects lying at angles, and the pattern of those angles differs by cause. A windstorm lays trees down in a shared direction. Selective logging fells trees toward skid trails, the narrow tracks used to drag logs out, so stems align locally with a branching trail network and point every which way at the scale of a whole plot. Drought mortality drops trees with no preferred direction at all.
We propose to measure this directly. The plan has three parts. First, train an oriented object detector on centimeter-resolution drone imagery to find downed stems, treefall gaps and trail segments as rotated boxes. Second, summarize each plot's detections as an angle-statistics "orientation field." Third, test whether that field separates logging, windthrow and drought mortality better than biomass loss, spectral change or object counts alone. A secondary component handles the misuse risk raised for fine-grained forest maps. Field devices sign compact orientation summaries with a lattice-based homomorphic signature whose security proof does not assume idealized hash functions. An auditor can then verify plot-level aggregates without ever receiving the coordinates of individual trees.
Core Claim
The orientation distribution of elongated disturbance debris is a measurable, cause-specific signature, and it can attribute canopy degradation to human or natural drivers where biomass-only signals cannot. The novelty sits in the angle-statistics attribution, not in a new backbone. If a simple instance-segmentation pipeline recovers angles as well as an orientation-aware detector, the attribution claim still stands, and we will report that honestly rather than defend the architecture.
Why This Gap Exists
The carbon literature motivates the question. Using passive microwave vegetation optical depth (a ~25 km satellite measure of woody water content that tracks biomass), Human influence on Amazon's aboveground carbon dynamics intensified over the last decade downscaled carbon change to 100 m and attributed −0.42 ± 0.08 PgC of 2010–2020 loss to degradation. Forest degradation contributes more to carbon loss than forest cover loss in North American boreal forests found degradation losses of −162.21 Tg C/yr against −28.65 Tg C/yr from cover loss. The synthesis in Recent advances and challenges in monitoring and modeling of disturbances in tropical moist forests names the remaining failures: no instrument resolves skid trails or treefall gaps, and spectral or dielectric pixels cannot separate natural mortality from illegal logging. These works define the target. Our method does not replace their continental budgets. It supplies plot-scale attribution evidence that such budgets currently infer only from land-cover transition maps.
The vision literature supplies the tools but has pointed them at the wrong objects. ReDet builds rotation-equivariant backbones whose features rotate along with the input, and pools rotation-invariant features for classification. FCOSR assigns training samples with ellipses fitted to slender rotated boxes. RQFormer aligns transformer attention to rotated regions. All are evaluated on ships, planes and vehicles. The one tropical tree benchmark, SelvaBox, annotates 83,137 crowns with horizontal boxes and uses standard detectors.
We deliberately do not make crowns the focus. Crowns are roughly isotropic, so rotated boxes should help them little, and we treat that as a control experiment rather than a contribution. The objects that are genuinely elongated, densely packed and arbitrarily oriented in a damaged forest are fallen stems and trails. Those are exactly the regime oriented detectors were built for.
RQFormer's own stated limitation constrains our design. Attention restricted to the rotated box discards surrounding context. For a fallen log, that context is diagnostic: the stump, the gap at its root end, or the trail it was dragged toward all lie along its long axis, just outside the box.
Research Program
Data. We will build an oriented-annotation layer for disturbance objects. Annotators draw rotated boxes for downed stems, treefall gaps and trail segments on a subset of SelvaBox orthomosaics (CanopyRS/SelvaBox), and add rotated crown boxes on a smaller sample for the control experiment. For attribution, we need plots with known disturbance history. We will collect drone orthomosaics at comparable resolution over documented selective-logging areas, documented blowdowns, and drought-mortality plots. We will cross-reference these with the 30 m Tropical Moist Forests degradation layer used by the Amazon study, so that every plot is confirmed as degraded and only the cause is unknown to the model. Collecting this ground-truth set is the main cost and risk of the project. Without it, the attribution claim cannot be tested.
Detector. We start from a strong oriented detector (ReDet and FCOSR as the two candidate bases) and add one ingredient: an axial context band. Beyond pooling from the predicted rotated box, the detection head pools from two strips that extend the box along its major axis by a learned fraction of its length. These strips capture the stump-and-gap end and the trail end. Angles are predicted as the doubled-angle vector $(\cos 2\theta, \sin 2\theta)$. A fallen stem's axis is the same at $\theta$ and $\theta+\pi$, so this encoding treats them as identical and removes the jump between $-\pi/2$ and $\pi/2$ that both the detection literature and the orientation-alignment literature flag as a source of unstable training.
Orientation field. For a plot with detected stems at angles $\theta_i$, define the alignment
$$S = \Big|\tfrac{1}{N}\textstyle\sum_{i=1}^{N} e^{2i\theta_i}\Big|,$$
where $N$ is the number of stems and $S \in [0,1]$ is near 1 when stems share an axis and near 0 when angles are uniform. We compute $S$ at several spatial scales. We also compute the mean angle between each stem and the tangent of the nearest detected trail segment. Our hypotheses are as follows. Windthrow should give high $S$ at all scales. Logging should give high stem-to-trail alignment but low plot-scale $S$. Drought mortality should give low values for both. A small classifier (a gradient-boosted model or logistic regression) on these features predicts the cause. We will keep it small on purpose, so that any success comes from the features rather than from model capacity.
Attested aggregation. SelvaBox warns that tree-level maps could aid illegal logging or land grabbing, but stops at a policy recommendation. Our pipeline exports only per-tile angle histograms and counts. These are vectors, and plot-level statistics are their sums. Efficient Post-Quantum Secure Network Coding Signatures in the Standard Model gives a linearly homomorphic signature: anyone can add signed vectors, and the sum still verifies. Its security rests on a lattice hardness problem without modeling hash functions as ideal random oracles, which matters for audit records that must stay trustworthy for decades. Drones sign tile histograms, an untrusted aggregator sums them, and an auditor verifies that the plot statistic is a genuine combination of authorized tiles while never seeing tile locations.
We take on the scheme's known cost, a public key of size $2mn\log q$ (where $n$ and $m$ are the lattice dimensions and $q$ the field size). We will need the larger field $q=2^{16}$ so that summed counts do not wrap around. This component is an engineering demonstration of feasibility, not a new cryptographic result.
Baselines
The naive reference for detection is SelvaBox's own horizontal-box pipeline (DINO with a Swin-L backbone), retrained on our disturbance classes. Angles are then recovered as the principal axis of the image pixels inside each horizontal box. On crowns, the cited model reports 48.60 RF1₇₅ in-domain (an F1 score that counts a detection as correct only at overlap ≥ 0.75) and 41.91 RF1₇₅ zero-shot on BCI50ha, versus 34.97 for Detectree2 (SelvaBox). For fallen stems, gaps or trails, no reported result was found on Lacuna for any method, so we establish the first numbers ourselves.
The strongest existing oriented detectors are ReDet and FCOSR. On the general benchmarks, the file reports 80.10 mAP on DOTA-v1.0 and 90.46 mAP on HRSC2016 for ReDet (ReDet), 79.25 and 90.13 for FCOSR (FCOSR), and 80.92 mAP multi-scale on DOTA-v1.0 for RQFormer. We run every detector on identical splits, with the same backbone family where possible, the same training schedule and the same hyperparameter-search budget. On the general benchmarks, our modified heads must stay within about 1 mAP of the published numbers, to show the context band does not hurt rigid objects.
Beating the oriented baselines means higher rotated-box mAP and lower axis error on downed stems, measured on held-out sites rather than held-out tiles. The key ablation removes the axial context band, and also replaces it with an equal-area isotropic ring. This isolates whether oriented context, not simply more context, is what helps.
For attribution, the do-nothing reference is predicting the majority cause. Three stronger references follow. The first is a biomass-and-spectral classifier using canopy height loss and multispectral change from the same imagery. The second is a counts-only classifier using the number and area of detected gaps and stems with no angles. The third is a classifier on the 30 m degradation-layer attributes. Beating them means a statistically clear gain in balanced accuracy under leave-one-site-out cross-validation.
For aggregation, the reference is per-tile signatures plus a trusted aggregator. We compare sign, combine and verify time, and key size, against the 0.437 ms combination time reported at $q=2^8$ by Xie et al.
Why not just…?
Why not just segment the debris and fit minimum-area rectangles? This is the real threat. Instance masks yield orientation from shape moments, and segmentation may handle crooked, broken stems better than boxes do. We include Mask R-CNN–style segmentation with rectangle fitting as a full baseline. If it matches the oriented detectors on axis error, we will say the detector architecture is unnecessary and rest the contribution on the orientation field. That attribution claim does not depend on which detector produced the angles.
Why not just use microwave biomass and 30 m degradation maps, as continental studies do? These data already show that degradation dominates, but by the authors' own account they cannot resolve trails or gaps, and they confound natural and human causes (Recent advances and challenges…). Our degradation-layer baseline tests this directly. The honest limit is coverage: drone plots cannot map a continent. We position the orientation field as calibration evidence for attribution inside coarse budgets, not as a replacement for them.
Why not just apply oriented detectors to tree crowns, the obvious extension of SelvaBox? Crowns are near-isotropic, so we expect rotated boxes to add little there. We test this on the re-annotated crown subset. A null result would confirm that the value of orientation lies in debris, which is our premise.
Why not just encrypt the data or use ordinary signatures? Encryption protects data in storage, but the auditor still needs a verifiable plot-level number. Ordinary signatures force the auditor either to trust the aggregator or to receive every tile. Homomorphic signing removes both requirements. Whether that is worth a large public key is an empirical question, and we answer it with measured costs on edge hardware.
Decisive Checks
The project succeeds if three things hold. First, under leave-one-site-out evaluation, orientation-field features improve cause attribution over both the counts-only and biomass-spectral classifiers. Second, the improvement disappears when angles are randomly shuffled within each plot. This shuffle control is the single most informative experiment, because it isolates angle information from object density. Third, angle error on held-out sites is low enough that alignment $S$ estimated from detections tracks $S$ computed from manual annotations.
The project fails, and we will say so, if the shuffled control performs as well as the real angles. That would mean the signal lies in how many objects there are, not how they are arranged. The detector ablation and the cryptographic cost measurements are secondary results reported either way.
References
- Human influence on Amazon’s aboveground carbon dynamics intensified over the last decade
- Forest degradation contributes more to carbon loss than forest cover loss in North American boreal forests
- Recent advances and challenges in monitoring and modeling of disturbances in tropical moist forests
- ReDet: A Rotation-Equivariant Detector for Aerial Object Detection
- FCOSR: A Simple Anchor-Free Rotated Detector for Aerial Object Detection
- RQFormer: Rotated Query Transformer for end-to-end oriented object detection
- SelvaBox: A high‑resolution dataset for tropical tree crown detection
- Efficient Post-Quantum Secure Network Coding Signatures in the Standard Model