From Pixels to Hierarchical Sequences: Quadtree Mask Encoding for Vision-Language Binary Change Detection

• Existing decoder-based VLMs separate dense prediction from autoregressive reasoning, complicating unified end-to-end token-level optimization.Flat patch-grid serialization also allocates tokens uniformly, causing spatial adhesion and poor handling of small fragmented changes.Source: §1 · pp. 1–2
• QUAKE recursively refines heterogeneous regions while collapsing uniform areas, preserving fine boundaries and compressing redundant background.Its grammar makes structural errors observable during autoregressive generation.Source: §1 · p. 2
• QUAKE-CoT pairs hierarchical mask sequences with chain-of-thought rationales describing change type, location, and extent.A progressive curriculum first teaches valid sequence generation, then integrates bi-temporal rationales.Source: §1 · p. 2
• 78.31% accumulated F1 and 73.18% per-image F1 were achieved, surpassing decoder-based baselines and uniform text-as-mask serializations.Reasoning evaluations also found more faithful bi-temporal explanations.Source: §1 · p. 2
• Grammar-gated dual-reward reinforcement learning jointly optimizes format validity, pixel-level mask accuracy, and semantic reasoning fidelity.Source: §1 · p. 2

• Existing remote-sensing VLMs commonly rely on auxiliary mask decoders or sparse coordinate cues for pixel-level prediction.Sparse cues delegate dense mask generation to external segmentation systems.Source: §2 · p. 2

Problem Formulation

• QUAKE-CD generates reasoning, summary, and quadtree-mask spans, then deterministically decodes valid mask sequences into binary change maps.The sequence uses the tripartite form ythink, yanswer, and yseg, with yseg drawn from the formal language LQ.Source: Problem Formulation · p. 2, Quadtree Mask Encoding · p. 3

QUAKE Encoding

• QUAKE partitions masks into 32×32 tiles, using uniform tokens or recursively expanded quadtree nodes until the minimum leaf size ℓ.The four node bits identify foreground-containing top-left, top-right, bottom-left, and bottom-right quadrants.Source: Quadtree Mask Encoding · p. 3
• QUAKE grammar supports linear-time syntactic validation, structural grid checking, and deterministic non-parametric decoding.These checks distinguish syntactic validity, structural validity, and semantic correctness.Source: Quadtree Mask Encoding · p. 3

QUAKE-CoT Dataset

• QUAKE-CoT pairs deterministic mask encodings with teacher-generated traces covering change type, location, extent, and plausible cause.The dataset contains 27,509 training samples and a 9,301-sample test set formed from validation and test splits.Source: Supervised Fine-Tuning with Curriculum · p. 4

Grammar-Gated Dual-Reward RL

• The dual-reward framework combines decoded-mask IoU rewards with a reasoning reward model to optimize spatial and semantic outputs together.Preference pairs use ground-truth traces, empirical rollouts, and hard negatives, while the combined reward is weighted by µ.Source: Grammar-Gated Dual-Reward RL · p. 4

Dense Change Mask Generation

• 78.31% accumulated F1 surpasses BIT by 1.37% and remains within 3.13% of ChangeFormer, while 73.18% per-image F1 exceeds all specialist models.QUAKE-CD also achieves the highest recall under both evaluation protocols and substantially outperforms PixelLM and LISA-7B.Source: Dense Change Mask Generation · p. 5
• 74.27 pairs/s is at least 1.85× faster than decoder-based VLMs, while accumulated F1 improves over Text4Seg and RSUniVLM by 53.55% and 15.97%.Throughput remains close to flat serialization methods despite QUAKE's additional hierarchical structure.Source: Dense Change Mask Generation · pp. 5–6
• QUAKE-CD exceeds the same-8B baseline by 31.25% in human reasoning score and leads the 235B model by 13.20%, indicating stronger grounding than scaling alone.It also achieves 49.30 BLEU-1 and 40.24 METEOR, while semantics-aware metrics distinguish capacity from grounded supervision.Source: Reasoning Quality Evaluation · p. 6

Ablation Studies

• F1 peaks at ℓ=8 as reducing ℓ from 16 to 2 expands average sequences from 57 to 523 tokens without monotonic quality gains.Very small leaves increase grammar violations and optimization difficulty, whereas ℓ=16 erodes boundary precision.Source: Ablation Studies · pp. 6–7, Figure 3 · p. 6
• Initializing Stage 2 from Stage 1 improves both F1 measures by roughly 4%, while Quadtree Reward adds about 1% accumulated F1 and raises reranker score from 68.05 to 69.72.The full dual-reward model reaches a reranker score of 71.26.Source: Figure 4 · p. 6, Ablation Studies · p. 7
• QUAKE-CD resolves dense small-building changes more accurately than external-decoder baselines and Text4Seg, whose predictions over-cover broad regions in dense scenes.Its textual descriptions specify change type, location, and extent.Source: Qualitative Analysis · p. 7, Figure 5 · p. 7

5 Conclusion

• QUAKE-CD formulates dense remote-sensing change detection as structured autoregressive generation under a syntax-verifiable mask grammar.The framework combines quadtree masks, grounded reasoning traces, staged curriculum learning, and grammar-gated dual-reward reinforcement learning.Source: §5 · p. 7

A Quadtree Mask Encoding Details

• The grammar is linearly checkable, and deterministic decoding reconstructs masks symbolically without neural overhead during inference or reinforcement-learning reward computation.Validity checks cover wrappers, dimensions, run lengths, recursion depth, and child counts.Source: A Quadtree Mask Encoding Details · p. 10, Decoding Algorithm D · p. 10, Formal Grammar of LQ · p. 10
• A quadtree node recursively expands only active quadrants in top-left, top-right, bottom-left, bottom-right order until reaching the minimum leaf size.Uniform tiles are represented directly, while heterogeneous regions receive finer spatial decomposition.Source: Formal Grammar of LQ · p. 10, Decoding Algorithm D · p. 11
• The encoding is recall-preserving because any leaf containing foreground forces foreground rendering, with precision loss confined to boundary leaves of size ℓ.This construction makes boundary quantization the source of false-positive expansion rather than foreground omission.Source: Encoding Algorithm E · p. 10
• 100.0% foreground recall, 91.2% accumulated F1, and 88.0% per-image F1 validate round-trip encoding fidelity on QUAKE-CoT with ℓ=8.The decoded masks serve as supervision targets for supervised fine-tuning and reward signals for reinforcement learning.Source: Encoding Fidelity and Compression… · p. 11, Table 3 · p. 11
• ℓ=8 balances compression against precision: ℓ=2 inflates sequence length, whereas ℓ=16 reduces spatial resolution and increases per-leaf false positives.The encoder recursively subdivides heterogeneous 32×32 tiles until the minimum leaf size is reached.Source: Encoding Algorithm E · p. 10, Encoding Fidelity and Compression… · p. 11, Table 4 · p. 11

B QUAKE-CoT Dataset Details

• QUAKE-CoT pairs pixel-level change masks with chain-of-thought reasoning traces describing grounded visual evidence.The dataset builds on established bi-temporal remote-sensing benchmarks and adds hierarchical mask sequences with teacher-distilled traces.Source: B QUAKE-CoT Dataset Details · p. 11

• QUAKE-CoT draws image pairs and binary masks from SYSU-CD, LEVIR-CD, and LEVIR-CD+, covering urban changes and building additions or removals.Source: Data Sources and Curation · p. 11
• Each mask is converted into a quadtree sequence after 256×256 normalization, including 16-tile partitioning for 1024×1024 LEVIR images with T=32 and ℓ=8.Source: Data Sources and Curation · p. 11
• Teacher traces compare the bi-temporal images using change categories and spatial reasoning rather than binary judgments, then undergo human cross-verification against the images and ground-truth masks.Four graduate students validate the traces, with 98.7% judged factually consistent and 1.3% revised.Source: Data Sources and Curation · p. 11
• Filtering removes inconsistent traces and ambiguous boundary-noise tiles while retaining 20% of fully empty-mask tiles to reduce unchanged-sample imbalance.Source: Data Sources and Curation · p. 12
• Held-out test splits follow the source benchmarks, and human assessments use common rubrics for factual consistency and spatial grounding.Source: Data Partitioning · p. 12, Data Sources and Curation · p. 12
• The final corpus contains 36,810 aligned image-pair, mask, and reasoning-trace triples, with traces averaging 60.4 words and 84 Qwen3-VL tokens.Source: Dataset Statistics · p. 12

Licenses

• QUAKE-CoT is released under the licenses of SYSU-CD, LEVIR-CD, and LEVIR-CD+, including the associated imagery, quadtree sequences, and reasoning traces.Source: Licenses and Limitations · pp. 12–13

Limitations

• Teacher distillation may propagate preferred phrasings, global-appearance cues, or culturally specific descriptors despite human cross-verification.Source: Licenses and Limitations · p. 13
• The source benchmarks over-represent building additions and removals relative to natural changes such as vegetation succession, waterbody fluctuation, and fine-grained land-cover conversion.Source: Licenses and Limitations · p. 13
• Removing tiles dominated by sub-tile boundary noise underweights legitimate micro-changes, so QUAKE-CoT should not be the sole benchmark for sub-pixel or thin-linear change detection.Source: Licenses and Limitations · p. 13

Preference Pair Construction

• Preference training anchors the reward model with ground-truth and fabricated hard-negative responses, while rollout comparisons retain pairs only when their IoU gap exceeds the margin threshold.The resulting preference set is intended to prioritize spatial fidelity over surface phrasing.Source: Hard-Negative Generation · p. 13, Preference Pair Construction · p. 14

Dual-Reward Specification

• The reward pipeline combines grammar and structural validation with Tversky-based mask accuracy and a learned score for bi-temporal reasoning traces.Grammar failures receive zero quadtree reward, while the content score weights false negatives more strongly than false positives using α=0.3 and β=0.7.Source: Dual-Reward Specification · p. 14

Oracle Performance Upper Bound

• 10.35% and 9.70% higher accumulated and per-image F1, respectively, result from QUAKE’s hierarchical encoding over flat text-as-mask encoding.Both representations preserve 100.00% foreground recall, while accumulated precision rises from 68.04% to 84.03% and per-image precision from 65.23% to 81.43%.Source: Oracle Performance Upper Bound · p. 15, Table 9 · p. 16
• 85.75% accumulated F1 and 82.52% per-image F1 of the representation oracle exceed the strongest alternative by 8.76% and 5.35%, respectively.QUAKE-CD’s normalized precision remains above 84% under both evaluation regimes.Source: Oracle Performance Upper Bound · pp. 15–16, Table 10 · p. 16
• 78.10% F1 on LEVIR-CD and 67.62% on LEVIR-CD+ demonstrate tiled inference on original 1024 × 1024 images without high-resolution fine-tuning or architectural modification.Recall is 98.61% on LEVIR-CD and 85.82% on LEVIR-CD+.Source: Scalability to High-Resolution Imagery · p. 16, Table 11 · p. 16
• Qualitative tiled predictions recover large changed structures but merge adjacent buildings and expand into surrounding regions, matching the reported high-recall, lower-precision behavior.The examples include curvilinear residential development and rows of newly constructed buildings.Source: Scalability to High-Resolution Imagery · p. 16

Generalization on Out-of-Domain Dataset

• The external evaluation shifts spatial resolution, sensor, and scene appearance while preserving native resolution and the same tiled deterministic-decoding protocol.WHU uses 0.2 m/pixel imagery versus 0.3 m/pixel for LEVIR-CD and 0.5 m/pixel for SYSU-CD.Source: Generalization on Out-of-Domain Dataset · pp. 16–17
• 69.10% per-image F1 surpasses the strongest specialist, ChangeFormer at 65.29%, while 71.29% accumulated F1 remains within 4.07% of it.QUAKE-CD’s accumulated recall is 78.40%, the best in the table.Source: Generalization on Out-of-Domain Dataset · p. 17, Table 12 · p. 17
• 13.04 and 12.45 points separate QUAKE-CD from the strongest VLM baseline, RSUniVLM, on accumulated and per-image F1, respectively.RSUniVLM reaches 58.25% accumulated F1 and 56.65% per-image F1.Source: Generalization on Out-of-Domain Dataset · p. 17
• 78.10% and 82.44% of the accumulated and per-image representation-oracle F1 scores are attained without adaptation.The oracle is computed by encoding and deterministically decoding WHU ground-truth masks with QUAKE.Source: Generalization on Out-of-Domain Dataset · p. 17
• Attention visualizations concentrate token-conditioned responses within designated tiles and aggregated responses on changed structures across varied object scales, shapes, and backgrounds.The overlays are consistent with predicted masks and ground-truth annotations.Source: Qualitative Attention Visualization · p. 18

F Failure Mode Analysis

• Grammar invalidity denotes unparsable <seg> spans, whereas structure mismatch denotes parsed grids whose dimensions differ from the expected H/32 × W/32 shape.Examples include malformed quadtree nodes, undersized run-length counts, and unmatched delimiters.Source: F Failure Mode Analysis · pp. 18–19
• Grammar invalidity falls from 1.26% after curriculum supervision to 0.13% after grammar-gated dual-reward RL.The grammar gate assigns zero reward to ungrammatical outputs.Source: F Failure Mode Analysis · pp. 18–19, Table 16 · p. 19
• Structure mismatch decreases from 12.64% to 2.17% after reinforcement learning, showing that the structure score penalizes grid-dimension drift beyond grammar validation.Pad/crop fallback is triggered exactly when a structure mismatch is detected.Source: F Failure Mode Analysis · p. 19, Table 16 · p. 19
• The QUAKE system prompt specifies 32×32 tiles, raster ordering, band-wise lines, and delimiter-based tile sequences for structured mask generation.These rules define the spatial layout that structure validation checks.Source: G Prompt Templates · p. 22, Figure 12 · p. 22
• Prompted teacher traces, hard negatives, reranking, and GPT pairwise scoring support QUAKE-CoT construction and preference-based evaluation.The resulting samples pair grounded reasoning and answer traces with QUAKE segmentation sequences.Source: G Prompt Templates · p. 19, Figure 13 · p. 23, Figure 14 · p. 24