Non-Myopic Zoom: Treating Coarse-to-Fine Filtering as Sequential Search

Large environmental surveys almost always run as a cascade. A cheap model scans everything at low resolution, and an expensive step (a high-resolution image, a drone overflight, a human check) is spent only on the tiles that look promising. The rule deciding which tiles get the expensive look is usually "inspect the highest-scoring tiles first." That rule is a greedy, one-step decision. It ignores two things. First, the budget remaining. Second, the fact that the outcome of one inspection tells you something about the uninspected tiles nearby. This note proposes to replace the rule with a belief model and a non-myopic choice. The belief model is a Gaussian process, a smooth probabilistic map that predicts a value and an uncertainty at every location from correlated nearby observations, placed over where the cheap model is wrong. The choice of each inspection then accounts for how its result will change the value of every inspection still to come.

The motivation is concrete. In the satellite industrial-emission pipeline of Pollution tracker, the coarse stage reached 94% held-out accuracy, yet it produced 890 candidates at 3% precision over a test region in North-Eastern India. The fine stage raised precision to 42%. The coarse model's errors almost certainly cluster by terrain and land use. A greedy cascade cannot learn this while it runs. It also never revisits tiles the coarse model scored low, so targets wrongly discarded there are lost for good. If inspections can teach the system where its coarse model is miscalibrated, a fixed fine-stage budget should find more targets and recover some of those coarse-stage misses.

Core Claim

A threshold or top-k cascade is a myopic acquisition function: a rule for choosing the next measurement that scores each option as if it were the last. The gap between that rule and a non-myopic one depends on one measurable quantity, the spatial correlation of the coarse model's errors.

We model the log-odds that tile $x$ contains a target as $g(x) = a\,s(x) + h(x)$. Here $s(x)$ is the coarse model's score, $a$ is a learned scale, and $h$ is a Gaussian-process residual with a Matérn kernel. Each fine inspection returns a binary label and updates $h$ everywhere nearby. Let $p_t(x)$ be the current posterior probability that $x$ holds a target, and let $q$ be the number of inspections left. We choose the next tile by maximizing

$$\alpha_t(x) = p_t(x) + \mathbb{E}_{y \sim p_t(y \mid x)}\Big[\textstyle\sum_{x' \in \mathrm{top}_{q-1}} p_t(x' \mid x, y)\Big],$$

where $\mathrm{top}_{q-1}$ is the set of $q-1$ tiles with the highest updated probability. The second term is the expected number of targets the remaining budget would find after seeing the outcome $y$ at $x$.

This form follows the batch reduction in BINOCULARS. That work replaces the intractable multi-step plan with an optimal batch over the remaining budget and proves the batch value lower-bounds the true sequential value. Its experiments, however, are on synthetic functions and hyperparameter tuning, with no spatial structure. Here the target is binary, so the expectation over $y$ requires only two posterior updates per candidate tile, which keeps the rule cheap.

The rule makes a falsifiable prediction. If the residual $h$ has no spatial correlation, the second term is constant across tiles, and the rule collapses to ranking by $p_t$, which is the greedy cascade. Gains should therefore grow with the length scale of the coarse model's errors, and that is the main thing the experiments test.

Positioning

The coarse-to-fine line supplies the setting and the failure we target. Hierarchical shape matching (Kokkinos & Yuille) prunes search subtrees using provable lower bounds on matching cost. Satellite cascades (Pollution tracker) and pathology agents (CPathAgent, whose zoom paths are imitated from model-generated demonstrations) decide where to zoom by fixed thresholds or learned imitation. None of these keeps a probabilistic belief that is updated by fine-stage outcomes, so pruning in all of them is irreversible.

The informative path planning literature supplies the belief machinery, and it motivates the expectation that non-myopia matters:

  • Gaussian-process maps with replanning are used in NAIVE, ε-Macro-GPO and PLUMES. PLUMES shows that greedy upper-confidence-bound sampling over-explores sub-optimal regions.
  • Continuous trajectory optimization (GP-PTO) shows that greedy planners spend expensive sensors too early in a mission. Its authors explicitly call for "adaptive query resolutions."

All of these methods sense at one fixed resolution over a flat map. None treats a cheap full-coverage scan as a prior to be corrected.

One result constrains the design. BINOCULARS warns that long lookahead over a poorly fitted early surrogate can hurt. We therefore cap the effective horizon at $\min(q, q_{\max})$ and treat $q_{\max}$ as an experimental variable, rather than assuming the full remaining budget is best.

Research Program

We will build the residual Gaussian-process cascade as follows:

  • Belief model. Logistic likelihood with a Laplace approximation, and a sparse inducing-point approximation so that continental tile counts stay tractable.
  • Candidate pool. Lookahead is evaluated over a pool formed by the top tiles under $p_t$ plus a stratified random sample of low-score tiles. Including low-score tiles is what lets the system audit regions the coarse stage would have pruned.
  • Hyperparameters. Kernel hyperparameters are refit online.
  • Overconfidence guard. A small fixed fraction of the budget is reserved for a calibration audit. If the realized hit rate on audited low-score tiles exceeds the model's predicted rate, kernel length scales are shrunk. This is our concrete answer to the irreversible-pruning problem.

A second regime adds physical travel cost. Here a drone verifies satellite candidates, so inspections cost distance as well as count. In this regime the top-$(q-1)$ sum is replaced by a short orienteering tour, the cheapest route that collects the most value within a budget. This is where the "information valley" behavior described by Singh et al. should appear: a valuable cluster sits behind tiles that are worthless on their own, and a greedy agent never crosses them.

Baselines

Do nothing. The naive reference spends the fine budget on uniformly random tiles. It sets the floor for targets found and is the only baseline that recovers coarse-stage misses without bias.

Strongest existing method. The threshold/top-k cascade of Pollution tracker inspects the highest coarse scores. Its reported deployment precision is 42% after the fine stage, with 3% before it. It is the strongest competitor because it is what practitioners actually run, and with independent tiles it is close to optimal. All methods will use the same coarse scores, the same fine-stage labels or classifier, and the same inspection budget. The cascade also gets its coarse scores post-hoc calibrated with temperature scaling on held-out data, so that calibration alone cannot explain our gain. Beating it means more true targets found at equal budget, plus non-zero recovery of targets below the coarse threshold.

Myopic Gaussian-process cascade. This is the key ablation. It uses the identical residual belief model, updated after every inspection, but drops the lookahead term ($q_{\max}=1$). The difference between it and our full method isolates non-myopia.

Lookahead without spatial coupling. A second ablation sets the kernel to white noise. The theory predicts it matches calibrated top-k, so it checks that any gain comes from spatial correlation.

Path-cost regime. Here we add PLUMES-style tree search with Monte Carlo Tree Search (MCTS) and GP-UCB replanning, each run with the same wall-clock budget per decision.

Why not just…?

Why not just calibrate the coarse classifier and keep top-k? Calibration fixes the average miscalibration but not spatially clustered errors. The 3% versus 94% gap in Pollution tracker suggests coarse confidence fails in specific terrains. This is a real threat if errors turn out spatially white. The white-kernel ablation and the synthetic correlation sweep below settle it directly. If calibrated top-k matches us at all correlation levels, the claim is wrong.

Why not just update the Gaussian process and act greedily? This captures most of the "learning where the coarse model errs" benefit. Greedy acquisition is known to fail when immediate value is low but downstream value is high, as in the information valleys of Singh et al. and the upper-confidence-bound mismatch in PLUMES, where median reward was 115 and 86 versus 206. Still, auditing a low-score tile pays off only if the rest of the budget can exploit what is learned. We expect the myopic cascade to be close in the pure-count regime and to fall behind mainly when the budget is large relative to the candidate pool, or when travel cost is present. The $q_{\max}$ sweep will show which holds.

Why not just reserve a fixed random fraction for exploration? A random audit recovers misses but does not aim them. Our method spends audit inspections where an outcome would change the most other tiles' probabilities. The comparison is our full method against top-k plus an ε-random audit, with ε tuned on a validation region to its best value.

Why not train an amortized policy instead? Single-forward-pass policies such as TNDP and CAtNIPP are fast. COPEx reports that purely amortized design policies are brittle under deployment constraints that differ from training, and each new region and coarse model here is a shift. Our per-decision cost is two posterior updates per candidate, so speed is not the bottleneck. We leave amortization as a follow-up rather than a competitor.

Proposed Experiments

Synthetic correlation sweep (the decisive check). Targets are planted on grids. A simulated coarse score has residual errors drawn from a Gaussian process whose length scale runs from zero (white) to several tiles. We plot targets found at a fixed budget against length scale for all methods. The prediction is that our gain over calibrated top-k is about zero at zero length scale and grows monotonically. A flat curve falsifies the core claim.

Industrial-emission detection. We rebuild the two-stage setup of Pollution tracker: a Sentinel-2 coarse classifier, with high-resolution labels as the fine stage. We report precision at budgets of 25, 55 and 100 inspections, the 55 matching the paper's reported 55 fine-stage candidates. We also report recall of targets below the coarse threshold.

Tree-crown detection. Using SelvaBox, the coarse layer is the 4.5 cm rasters downsampled to roughly 1 m. The target is a tile containing a large crown defined from the box annotations, and the fine stage reveals the true label. Crown size varies by site and terrain, which gives naturally correlated coarse errors.

Drone verification with travel cost. On the same SelvaBox tiles we impose a travel budget. We report targets found per kilometre, plus decision time against the 10–30 s per plan reported for GP-PTO. No reported result was found on Lacuna for any IPP method on SelvaBox or on the Pollution-tracker region, so every baseline in those settings will be run by us under identical budgets.

Horizon and robustness checks. Across all settings we sweep $q_{\max}$ from 1 to the full budget to test the BINOCULARS misspecification warning. We also deliberately misspecify the kernel length scale by a factor of four, to test whether the calibration-audit guard prevents the irreversible pruning that pure cascades suffer.