{"id":"297d39d8-55ed-4dbc-b4e9-75783928a761","arxiv_id":"2506.14260","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An oblique-axis regression tree is fitted to recent images, the learned splitting rules are used to predict the next image, and a CUSUM statistic flags changes in the drift pattern of jump location curves.","lead":"This paper proposes a decision-tree-based control chart that watches a sequence of images and flags when a gradual drifting pattern, such as a shrinking lake, changes speed or direction. Since many image streams, from satellites to medical scans, change slowly, a method that detects shifts in drift rather than only sudden jumps could support monitoring of environmental and clinical change.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The null-distribution claim in Theorem 1 depends on the ORT leaf structure extrapolating one time step ahead; this is only justified for exactly linear JLCs with constant drift velocity, a condition violated by curved JLCs and by the paper's own Aral Sea application, so Λ_tk may not center at 1…","rationale":"The reader's weakest assumption is that the ORT leaf-averaging estimator remains consistent at a future time point only when JLCs are exact line segments drifting with constant velocity. I agree, and I see this as the single most load-bearing condition for the central claim: Theorem 1's null centering n/√2(Λ_tk − 1) ⇒ N(0,1) is established by an extrapolation argument that relies on the 3-D JLC surface being a plane. The simulations provide useful empirical support under exactly the line-segment model, but they do not probe the curved case, and the real Aral Sea application uses coastlines that are not straight. The paper's own assumptions (v) and Definition 2 make the line-segment/constant-velocity conditions explicit, so this is not an internal inconsistency; it is a scope limitation that the abstract and real-data discussion exceed. The suggested experiment would settle whether the null distribution holds away from that scope. Because the reader already assigned CONDITIONAL with moderate confidence, my concern does not change the verdict; it sharpens the condition that should be attached to acceptance, namely that the central claim is established only for piecewise-linear JLCs with constant drift velocity unless the extrapolation step is proven under weaker conditions.","tokens_in":17255,"tokens_out":14212,"duration_ms":164588,"concrete_test":"Simulate in-control data with a curved JLC, e.g., a circle of radius r(t) = r0 + vt (constant radial drift) with two intensity levels, n = 128, m0 = 20, Gaussian noise σ = 0.15, and no change in the drift pattern, then run Algorithm 2 for 500 independent replications. Compute the empirical mean and variance of n/√2(Λ_tk − 1) and the in-control ARL using the same calibration procedure as in Section 4.1. If the mean departs from 0 or the false-alarm rate is substantially above the nominal level, the line-segment assumption in Theorem 1 is load-bearing and the central claim fails for curved JLCs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 1 in Supplementary S.1 reduces the null result to the claim that |f(x,y,t_{k+1}) − bf_tk(x,y,t_{k+1})| → 0 a.e. under H0. The argument uses Eqns. (S.1)–(S.3): an interior point at t_k is assumed to remain in the same region at t_{k+1}, and the leaf average at the future point is assumed to differ negligibly from the leaf average at t_k. Equation (S.3) is only valid when the JLC is a line segment drifting with constant velocity, because only then is the decision boundary α1x + α2y + α3t = c a plane that the ORT can extrapolate exactly. If the JLC is curved, the ORT approximates it by a piecewise-linear boundary; a leaf that is correctly assigned at t_k can cut across the true region at t_{k+1}, so leaf averaging pools intensities from different constant regions and E[Λ_tk | H0] ≠ 1. The same failure occurs with non-constant drift velocity, since the 3-D boundary is no longer a plane. Assumption (v) restricts JLCs to line segments, and every simulation uses straight-line JLCs, but the paper's stated goal and the Aral Sea application involve curved coastlines (Figure 9), so this is not an idle edge case. The theorem's proof is also explicitly a sketch deferred to the companion paper [Basak et al., 2025]; the extrapolation step is precisely the part not covered by that paper's consistency lemmas, as the supplementary text itself acknowledges. Thus the most load-bearing condition for the central claim is the exact-linearity/constant-velocity assumption on the JLCs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sequential monitoring scheme for detecting changes in the drift pattern of jump location curves (JLCs) in image sequences. The method fits an oblique-axis regression tree (ORT) to the most recent m0 images, extrapolates the resulting partition to the next time point, computes a normalized squared prediction error Λ_tk between the observed image and the leaf-averaged prediction, and feeds n/√2(Λ_tk − 1) into a CUSUM statistic Q_tk. Theorem 1 claims that under the null hypothesis of an unchanged drift pattern Λ_tk is asymptotically standard normal after centering and scaling, while under the alternative the statistic drifts by an amount proportional to Δ². The paper reports simulations for five scenarios—straight-line JLCs with constant or doubled drift rates, a stationary JLC change, and a case where only intensities change—and compares against the methods of Bui and Apley (2018) and Koosha et al. (2017), followed by an application to Landsat images of the Aral Sea.","tokens_in":17719,"tokens_out":2662,"duration_ms":29078,"significance":"The problem is well motivated and largely underexplored: existing image monitoring methods are not designed for gradual drift-pattern changes, and the ORT-based CUSUM construction is a genuinely new way to combine discontinuity-preserving estimation with sequential monitoring. If Theorem 1 and the numerical comparisons are reliable, the method would be a useful addition to the SPC toolkit for image streams, with the additional advantage of still detecting abrupt JLC changes when no drift is present. The paper also makes a welcome empirical effort, reporting ARL comparisons on multiple scenarios and a real-data application. However, the theoretical foundation is currently not self-contained: the key asymptotic result is deferred to a companion preprint, and the extrapolation step that is central to the null distribution is only justified under quite restrictive conditions that the paper's own real-data example appears to violate.","major_comments":[{"comment":"The null-distribution claim of Theorem 1 depends on the step 'if (x,y,t_k) ∈ int(Γ_l*) then (x,y,t_{k+1}) ∈ int(Γ_l*)' and on Eq. (S.3), both of which require that the 3-D boundary α1 x + α2 y + α3 t = c is a plane that the ORT can extrapolate exactly. This holds only when each JLC is a line segment drifting with constant velocity, as in Assumption (v) and Definition 2. For curved JLCs or nonconstant drift velocity, a leaf that is correctly assigned at t_k can cut across the true region at t_{k+1}, so the leaf average pools intensities from different constant regions and E[Λ_tk | H0] ≠ 1. Since the stated goal and the Aral Sea application (Figure 9) involve curved coastlines, the scope of Theorem 1 is narrower than the paper's central claim in Section 1.3.","section":"Supplementary Materials, S.1, Eqns. (S.1)–(S.3)"},{"comment":"The proof of Theorem 1 is only a sketch and defers the key consistency of the ORT leaf-averaging estimator to the companion preprint [Basak et al., 2025]. The supplementary text itself notes that those lemmas apply only when the evaluation time point lies within the training range, whereas the monitoring procedure predicts at t_{k+1} from data up to t_k. The extrapolation inequality (S.3) is precisely the part not covered by the companion lemmas, so the normal limit in Eq. (S.4) is not established by the cited results. The manuscript should either provide a self-contained proof of the extrapolation step under stated assumptions or restrict the theoretical claim to a regime where that step is proved.","section":"Section 3, Theorem 1 and Supplementary Materials, S.1"},{"comment":"The derivation of the asymptotic normality treats the pixelwise residuals z_tk(x,y) = (ε + δ)² as independent noncentral χ² variables and then applies a CLT to their average, but the ORT leaf averaging induces dependence across pixels within leaves, and the number of leaves K(I_tk) appears in the denominator of Eq. (4) while the proof replaces n² − K by n² without discussing the rate of K. The claim that E_tk is bounded by 2 is also stated without justification. These steps are load-bearing for the N(0,1) centering and scaling; without a more careful treatment of dependence and leaf-count asymptotics, the theorem is not fully verified.","section":"Section 3, Theorem 1; Supplementary Eq. (S.4)"},{"comment":"All simulations use piecewise constant images with straight-line JLCs and linear drift, so they are consistent with the restrictive conditions of Theorem 1 but do not test the method's robustness to curved JLCs or nonlinear drift. Given that the real-data application (Section 4.3) involves a curved coastline, a simulation with a curved JLC or time-varying drift velocity would directly address whether the monitoring statistic remains calibrated under the conditions the application requires.","section":"Section 4.1, Simulations 1–5"}],"minor_comments":[{"comment":"The text says a change in drift pattern is detected 'between the time-points tk and t(k+1)', but the statistic Q_tk is computed using the image at t_k; the indexing should be made consistent.","section":"Section 2.5, Eq. (5)"},{"comment":"In the 'for' loop, 'Fit a tree using algorithm 2.4' should read 'using Section 2.4', and the notation T_tk_rule is introduced without being defined consistently with Trule in Section 2.4.","section":"Algorithm 2"},{"comment":"The table and surrounding text contain formatting inconsistencies such as 'Koosha-et.-al.' and unclosed parentheses in the standard deviation column; these should be cleaned up.","section":"Section 4.2, Table 1"},{"comment":"The imputation equation omits the indicator or domain specification for t ∈ (t_i, t_{i+1}), and the interpolation formula's denominator (t_{i+1} − t_i) is written but the fraction is not typeset clearly; please clarify.","section":"Section 4.3.1"},{"comment":"Assumption (v) says 'Each JLCs is a line segment', but the earlier discussion in Section 2.3 says JLCs are piecewise linear with each linear part treated as a separate JLC; this should be reconciled explicitly.","section":"Section 3, Assumption (v)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central theoretical claim is not self-contained: it leans on a companion preprint for the consistency of the ORT estimator, and the extrapolation step—the one part not covered by that preprint—is exactly where the restrictive linearity assumption is needed. The empirical work is promising but narrow. I would suggest the editor ask for either a self-contained proof of the extrapolation step under clearly stated conditions (possibly including a mild smoothness or curvature condition on JLCs) or an explicit limitation statement that Theorem 1 applies only to piecewise-linear JLCs with constant drift velocity, with simulations and application framed accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper presents a genuinely new combination—using an oblique-axis regression tree fitted on a moving window of images, extrapolating its splits one step ahead, and feeding the squared prediction error into a CUSUM—to detect changes in the drift pattern of jump location curves. The simulation evidence is supportive under the paper's explicit linearity assumptions, and the Aral Sea application is a nice demonstration. I think it deserves a serious referee, but the central theory is not yet airtight.\n\nWhat's actually new: previous work monitors JLCs for abrupt changes or uses univariate drift CUSUMs; this is the first to track gradual drift patterns directly via tree-based partitions. The method preserves jump discontinuities and, when there is no drift, still detects step shifts, which is a useful flexibility. The simulations cover several drift scenarios and compare favorably to two baseline image monitoring methods; the ARL results look right.\n\nSoft spots: Theorem 1 is inherited from a companion preprint, and the sketch in the supplementary has a load-bearing step—the claim that leaf averages extrapolated to t_{k+1} remain close to those at t_k (Eqn. S.3)—that is only justified when JLCs are exactly line segments moving with constant velocity. The paper states assumption (v) (linear JLCs) clearly, but then applies the method to the Aral Sea, where coastlines are curved; the theory does not cover that case, and the paper does not discuss the mismatch. The bootstrap control limit in the real data partly dodges the problem, but the advertised broad applicability outruns the proof. Also missing: code, data, and the oblique split optimization details, so independent replication is not yet straightforward. The allowance parameter and control limit are calibrated on in-control data, which is standard SPC practice but means the reported ARLs are not a fully external validation.\n\nBottom line: for a reader working on image monitoring, this is a useful idea with a clear simulation setup, but I would not rely on the null distribution for curved boundaries until the extrapolation step is either proved under weaker conditions or the method is reframed as a heuristic. I'd send it to peer review with a request to address the theory-application gap and release code.","headline":"Novel ORT-CUSUM drift monitoring method with good simulations, but the theory's one-step-ahead extrapolation is only justified for linear JLCs and the paper should be upfront about that gap.","tokens_in":18204,"tokens_out":3808,"would_cite":false,"duration_ms":40410,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P30","62L10","62G08"],"pacs":[],"model":"deepseek-v4-flash","headline":"A CUSUM statistic built from oblique-axis regression tree predictions can detect changes in the gradual drift of jump location curves in image sequences, with an asymptotic normality result that gives it a principled control-limit…","keywords":["jump location curves","jump regression analysis","oblique regression tree","tree-based image monitoring","CUSUM control chart","drift pattern monitoring","sequential image monitoring","change-point detection"],"falsifier":"Simulate an in-control sequence in which a curved JLC, such as a circle expanding at constant rate, drifts with constant velocity, run the ORT-CUSUM with m0 = 20, and check whether the empirical mean of n/√2(Λtk − 1) stays near 0 and the in-control ARL matches the nominal level; if the statistic drifts or signals prematurely, the line-segment and constant-velocity assumption is load-bearing.","tokens_in":17086,"feed_emoji":"🛰️","tokens_out":4527,"duration_ms":43032,"temperature":0.7,"pith_summary":"The paper proposes a sequential monitoring scheme that detects changes in the gradual drift of jump location curves in image sequences. The method fits an oblique-axis regression tree (ORT) to a moving window of past images, extrapolates the tree's splitting rules to the next time point, and uses leaf-only averaging to produce a predicted denoised image. The discrepancy between this prediction and the observed image feeds a CUSUM statistic that raises a signal when the drift pattern changes. A central theorem states that under the null hypothesis of an unchanged drift pattern the normalized discrepancy is asymptotically standard normal, and under the alternative the statistic drifts by a term proportional to the squared jump deviation. Simulations show detection delays near one time point for several drift changes, and a real-data application flags a change in Aral Sea shrinkage around November 2018.","feed_headline":"Tree-based monitor catches drift-pattern changes in image streams","feed_subtitle":"A CUSUM statistic built from regression-tree predictions flags changes in gradual drift within a few frames.","key_machinery":"Oblique-axis regression tree (ORT): a decision tree that splits nodes by linear combinations of the coordinates (x,y,t), producing polytopal leaves that can track slanted jump location curves. The recursive splitting algorithm chooses splits maximizing the reduction in sum of squares; leaf-only averaging yields a jump-preserving estimate. The load-bearing mechanism is the predicted partition: the tree fitted on images up to the current time is evaluated at the next time point, and the CUSUM statistic Qtk = max(0, Qt(k−1) + n/√2(Λtk − 1) − κ) accumulates the centered, scaled prediction error.","core_discovery":"The central claim is that a CUSUM chart constructed from ORT-based prediction error can distinguish a stable drift pattern from a changed one. At each time point, an ORT is fitted to the last m0 images and its split rules are evaluated at the future time point; leaf-only averaging over the predicted leaves gives a denoised estimate of the next image. The statistic Λtk is the mean squared difference between this predicted image and the observed image, scaled by noise variance and degrees of freedom. Theorem 1 states that n/√2(Λtk − 1) converges in distribution to N(0,1) when the drift pattern is unchanged, while under a changed drift pattern the same quantity shifts by O(1)Δ², where Δ is the jump deviation associated with the change. This makes the method a drift-pattern change detector, and, when no drift is present, it also detects abrupt changes in jump location curves.","pith_inferences":["The detection delay appears tied to how quickly the drift displacement becomes visible at the pixel scale; for slow drifts, several frames will be needed before the prediction error accumulates enough to signal.","Because the estimator is consistent only when the true JLCs are line segments with constant-velocity drift during the in-control window, curved boundaries would likely bias Λtk away from 1 even under the null; a piecewise-linear approximation of curved JLCs is a natural stress test.","The ORT splitting rule could be extended to monitor drift patterns in tensor or multi-channel images, or to include spatio-temporally correlated noise by replacing the i.i.d. Gaussian assumption with a block bootstrap calibration of control limits.","The Aral Sea result, a signal near November 2018, gives a concrete target for ecological studies to confirm whether a deceleration in shrinkage began in the second half of 2018."],"forward_implications":["A change in drift rate or drift direction of a JLC raises a CUSUM signal, typically within one to four time points in the simulations.","In the absence of any drift, the same statistic still detects abrupt changes in jump location curves, so the method covers both gradual and step changes.","The method ignores intensity changes that do not move JLCs, such as uniform brightness drift, which is useful in the presence of shadows or lighting effects.","The theoretical result gives an asymptotic normal benchmark, so control limits can be set from the standard normal when the noise variance and prediction error are estimated.","The method is designed for registered, equally spaced image sequences and can be applied to three-dimensional image arrays as well."],"supporting_citations":[{"why":"Supplies the consistency lemmas for the ORT leaf-averaging estimator that Theorem 1 invokes directly.","marker":"[Basak et al., 2025]"},{"why":"Provides the framework for choosing the CUSUM allowance parameter κ and control limits.","marker":"[Qiu, 2013]"},{"why":"The closest prior drift-monitoring method; the proposed approach is positioned as a more general, JLC-aware alternative.","marker":"[Yi and Qiu, 2023]"},{"why":"Baseline comparison method for textured-surface image monitoring; the paper compares ARL performance against it.","marker":"[Bui and Apley, 2018]"},{"why":"Baseline wavelet-based GLR image monitoring method used in the simulation comparisons.","marker":"[Koosha et al., 2017]"},{"why":"Establishes the jump regression analysis model that the paper adopts for image sequences.","marker":"[Qiu, 2005]"}],"fun_headline_variants":["Tree-based CUSUM flags drift shifts in image series","Oblique regression tree detects gradual image changes","Drift patterns in images caught by tree method","Decision tree approach monitors drift in image data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme hinges on the assumption that during the in-control period the jump location curves are exact straight-line segments drifting at a constant velocity, so that a tree fitted on past images remains accurate when extrapolated one step ahead to the next image.","fun_headline_variants_meta":{"raw":{"variants":["Tree-based CUSUM flags drift shifts in image series","Oblique regression tree detects gradual image changes","Drift patterns in images caught by tree method","Decision tree approach monitors drift in image data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1589,"prompt_tokens":887,"completion_tokens":702,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":642}},"tokens_in":503,"tokens_out":702,"duration_ms":7304,"temperature":1.0,"reasoning_tokens":642,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:17:37.620650+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate an in-control sequence in which a curved JLC, such as a circle expanding at constant rate, drifts with constant velocity, run the ORT-CUSUM with m0 = 20, and check whether the empirical mean of n/√2(Λtk − 1) stays near 0 and the in-control ARL matches the nominal level; if the statistic drifts or signals prematurely, the line-segment and constant-velocity assumption is load-bearing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the consistency lemmas for the ORT leaf-averaging estimator that Theorem 1 invokes directly."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline comparison method for textured-surface image monitoring; the paper compares ARL performance against it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline wavelet-based GLR image monitoring method used in the simulation comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the jump regression analysis model that the paper adopts for image sequences."}],"review_version":1}