ChangeChat-105k: RS Change Analysis Dataset

• ChangeChat-105k is a structured dataset that extends traditional remote sensing change detection to an instruction-following, dialogue-oriented paradigm using bi-temporal imagery.
• It employs a hybrid pipeline combining rule-based generation and GPT-assisted methods to create diverse, accurate instruction–response pairs across multiple tasks.
• The dataset supports various evaluation tasks—including binary classification, quantification, localization, and multi-turn conversations—advancing interactive RS change analysis.

Searching arXiv for ChangeChat-105k and closely related RSICA work. ChangeChat-105k is a large-scale instruction-following dataset for bi-temporal remote sensing image change analysis, introduced as the data foundation for remote sensing image change analysis (RSICA) in “DeltaVLM: Interactive Remote Sensing Image Change Analysis via Instruction-guided Difference Perception” (Deng et al., 30 Jul 2025). It pairs bi-temporal satellite image patches with 105,107 instruction–response items spanning structured and open-ended change understanding. Its design explicitly extends classical change detection and change captioning into a dialogue-style setting in which a model can answer questions about whether change occurred, what changed, how much changed, where change is located, and how a sequence of follow-up questions should be resolved from the same image pair.

1. Conceptual position within remote sensing change analysis

ChangeChat-105k was created to support RSICA, a paradigm defined as integrating “the semantic grounding of change captioning with the interactive and reasoning capabilities of VQA” for bi-temporal remote sensing imagery (Deng et al., 30 Jul 2025). Traditional change detection yields pixel-level masks, change captioning yields one-shot textual descriptions, and remote sensing visual question answering typically addresses single images rather than temporal differences. ChangeChat-105k is designed to close that gap by turning bi-temporal change analysis into an instruction-conditioned language task.

The dataset is therefore not a generic conversational corpus. Its basic unit is an image-pair-centered instruction–response sample grounded in remote sensing change signals. The interaction space is deliberately heterogeneous: short fixed-format answers such as “yes” or “no” coexist with count expressions, grid-based spatial outputs, descriptive captions, open-ended answers, and multi-turn dialogues. A plausible implication is that the dataset is intended not merely to measure caption fluency, but to supervise instruction sensitivity, temporal reasoning, and discourse continuity over the same visual evidence.

As far as the paper claims, ChangeChat-105k is the first large-scale bi-temporal remote sensing instruction-following dataset with multi-task and multi-turn coverage (Deng et al., 30 Jul 2025).

2. Source datasets and construction pipeline

ChangeChat-105k is derived from two existing remote sensing change datasets: LEVIR-CC and LEVIR-MCI (Deng et al., 30 Jul 2025). LEVIR-CC provides 10,077 bi-temporal image pairs, each with five human-written change captions, on $256 \times 256$ RGB patches at $0.5$ m/pixel resolution, mainly in urban and suburban scenes. LEVIR-MCI extends this foundation with pixel-level binary change maps, which are used to derive binary change labels, object counts, and localization signals.

The generation pipeline is explicitly hybrid. Rule-based generation is used for structured tasks, namely change captioning, binary change classification, category-specific change quantification, change localization, and part of the multi-turn dialogue construction. GPT-assisted generation is used for open-ended question answering and the more open conversational portions of multi-turn data (Deng et al., 30 Jul 2025). The rationale is division of labor: rule-based stages preserve deterministic correctness for low-level labels, while GPT contributes linguistic diversity and more natural question–answer phrasing for semantically richer tasks.

A central design constraint is that GPT is never given the raw images. Instead, it receives structured summaries derived from ground-truth annotations, including change captions, object counts, and contour information. This grounding mechanism constrains open-ended language generation to the visual evidence already encoded in the source datasets. The paper presents this as an implicit quality-control mechanism against hallucinated changes.

3. Interaction types and annotation schema

ChangeChat-105k contains six interaction types covering both structured and free-form remote sensing change analysis (Deng et al., 30 Jul 2025).

Interaction type	Training	Test
Change captioning	34,075	1,929
Binary change classification	6,815	1,929
Category-specific change quantification	6,815	1,929
Change localization	6,815	1,929
Open-ended QA	26,600	7,527
Multi-turn conversation	6,815	1,929

Change captioning wraps an existing human-written LEVIR-CC description into an instruction–response pair with the fixed instruction “Please briefly describe the changes in these two images.” Binary classification asks whether change occurred and derives “yes” or “no” directly from the change map. Category-specific quantification computes the number of changed roads and buildings by applying OpenCV contour detection to binarized change maps. Change localization imposes a $3 \times 3$ spatial partition,

$$P = \{\mathrm{TL}, \mathrm{TC}, \mathrm{TR}, \mathrm{CL}, \mathrm{CC}, \mathrm{CR}, \mathrm{BL}, \mathrm{BC}, \mathrm{BR}\},$$

and marks a block as changed when more than $5\%$ of its pixels are labeled changed.

Open-ended QA and multi-turn conversation extend these structured signals into natural-language supervision. In the GPT-assisted stages, prompts may include the five change captions from LEVIR-CC, a JSON-style object containing change counts such as roads and buildings, and normalized contour coordinates for changed instances (Deng et al., 30 Jul 2025). This allows questions about the “main change,” the number of newly added structures, or their approximate placement in the scene. Multi-turn dialogues are designed to progress from easier to harder reasoning, for example: first deciding whether change exists, then counting changed objects, then producing a detailed description based on the preceding analysis.

The annotation schema can be formalized as

$D_{\text{train}} = \{(I_j, P_j, T_j)\}_{j=1}^M,$

where $I_j$ contains the image pair $(I_{t_1}, I_{t_2})$, $P_j$ is the instruction, and $T_j$ is the target response (Deng et al., 30 Jul 2025). The learning target is therefore the mapping

$0.5$0

with $0.5$1 ranging from a binary token to a count statement, a set of grid codes, or a multi-sentence explanation.

4. Data representation and task semantics

At the image level, each sample is based on a bi-temporal pair

$0.5$2

with $0.5$3 in the raw LEVIR-derived data (Deng et al., 30 Jul 2025). The paired structure makes temporal comparison intrinsic rather than auxiliary. ChangeChat-105k thus differs from single-image VQA datasets by requiring cross-time semantic alignment between two views of the same scene.

For each image pair, the dataset provides or implies several annotation layers: five human-written change captions, a pixel-level binary change mask, object counts for changed categories such as roads and buildings, normalized object contours, and grid-based localization labels derived from the change map (Deng et al., 30 Jul 2025). These are not merely side metadata; they are the intermediate semantic representations from which the six interaction types are constructed.

The task semantics are correspondingly multi-resolution. Binary classification operates at the whole-pair level. Quantification moves to object-instance reasoning. Localization adds coarse spatial grounding through symbolic cell identifiers rather than dense masks. Captioning and open-ended QA require sentence-level semantic aggregation over the same changes. Multi-turn dialogue introduces history dependence: the current answer must remain compatible with previous questions and answers about the same image pair.

The multi-turn component is especially significant because it changes the unit of supervision from isolated prompts to dialogue state. In this setting, the current instruction can include a history of earlier turns, and the correct answer may depend on antecedents such as “these buildings” or “the new road” (Deng et al., 30 Jul 2025). This suggests that the dataset is intended to supervise not only visual difference recognition but also context-sensitive reference resolution under temporal imagery.

5. Role in DeltaVLM and benchmark function

ChangeChat-105k is the training corpus on which DeltaVLM is instruction-tuned for RSICA (Deng et al., 30 Jul 2025). In that system, the image pair is processed by a bi-temporal vision encoder, converted into difference features, refined by a visual difference perception module, aligned with the instruction by an instruction-guided Q-former, and decoded by a frozen LLM. The dataset provides the diversity of instructions and targets needed to train this pipeline across captioning, classification, quantification, localization, open-ended QA, and dialogue.

Within DeltaVLM, the training objective is standard token-level cross-entropy over target tokens $0.5$4,

$0.5$5

where $0.5$6 is a one-hot ground-truth token and $0.5$7 is the predicted distribution at position $0.5$8 (Deng et al., 30 Jul 2025). The significance of ChangeChat-105k here is that a single instruction-conditioned formulation supports multiple output regimes without changing the basic modeling interface.

The benchmark protocol reflects this unification. Captioning is evaluated with BLEU-1/2/3/4, METEOR, ROUGE-L, and CIDEr; binary classification uses Accuracy, Precision, Recall, and F1; quantification uses MAE and RMSE; localization uses Precision, Recall, F1, Jaccard similarity, and Subset accuracy; and open-ended QA is evaluated with captioning-style text metrics (Deng et al., 30 Jul 2025). The dataset therefore functions not only as training supervision but also as a composite benchmark for interactive change analysis.

A common misconception would be to treat ChangeChat-105k as a pure dialogue dataset analogous to open-domain chat corpora. Its actual role is narrower and more technical: it is an instruction-following benchmark over bi-temporal remote sensing evidence, with dialogue added as one interaction mode rather than as the sole organizational principle.

6. Scope, limitations, and access

The dataset inherits several constraints from LEVIR-CC and LEVIR-MCI, and these materially shape its coverage (Deng et al., 30 Jul 2025). First, its visual domain is dominated by urban development, especially changes involving roads and buildings. Other land-cover transitions, such as flooding, agriculture, or deforestation, are largely absent. Second, it is limited to high-resolution RGB imagery at $0.5$9 m/pixel; there is no SAR, multispectral, or multi-sensor component. Third, localization is coarse, using a $3 \times 3$0 grid rather than dense grounding within the language supervision. Fourth, the temporal model is bi-temporal only; sequences with three or more time points are not represented.

The linguistic layer also has a mixed provenance. Base change captions are human-written in LEVIR-CC, but open-ended QA and parts of the conversational data are generated by ChatGPT from structured summaries rather than directly authored from the images (Deng et al., 30 Jul 2025). The paper presents this as grounded generation rather than free hallucination, because GPT never receives raw pixels. Even so, this setup may introduce LLM-style phrasing and biases into the open-ended portions of the corpus.

ChangeChat-105k is publicly distributed together with code, dataset files, and pretrained weights through the DeltaVLM project repository (Deng et al., 30 Jul 2025). The paper additionally documents the training-time image preprocessing used with DeltaVLM, including random cropping that removes $3 \times 3$1–$3 \times 3$2 of image content, random rotation within $3 \times 3$3, and resizing to $3 \times 3$4 for compatibility with the ViT-g/14 encoder. For reuse with other models, the paper’s practical guidance is to preserve the bi-temporal pairing, keep instruction–response formatting intact, and retain multi-turn dialogue history where applicable.

In summary, ChangeChat-105k is best understood as a structured, multi-task, instruction-following corpus that redefines remote sensing change analysis as an interactive vision–language problem rather than as a one-shot detection or captioning task (Deng et al., 30 Jul 2025). Its importance lies less in raw scale alone than in the specific combination of pixel-grounded change evidence, task diversity, and multi-turn supervision that makes RSICA trainable as a unified benchmark.