GeoChrono: Benchmarking and Rethinking Long-Term Temporal Understanding in Remote Sensing

ACM Multimedia 2026
Yujie Li1, Jiancheng Pan2, Zhiwei Wei3, Jiuniu Wang4, Mugen Peng1, Wenjia Xu1,✉
1State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications
2Tsinghua University    3Hunan Normal University    4City University of Hong Kong
TL;DR

We decompose long-term remote sensing understanding into a four-level cognitive hierarchy and instantiate it as ChronoBench (12 sub-tasks, 17,689 validated QA pairs). Benchmarking reveals that mainstream MLLMs fall drastically behind human experts, with Long-Term Memory as the most critical bottleneck. We then propose GeoChrono, an MLLM that models per-location temporal trajectories, reaching 78.34% overall accuracy — over 20 points above the best commercial model.

Overview of ChronoBench's four-level hierarchy and GeoChrono's performance radar
ChronoBench decomposes long-term remote sensing understanding into a four-level cognitive hierarchy. Built upon it, GeoChrono constructs per-location temporal trajectories to model the long-term evolution of geographic parcels, achieving state-of-the-art performance and human-level accuracy on several sub-tasks.

A Four-Level Cognitive Hierarchy

The Earth evolves along space and time, yet most remote sensing interpretation methods see only static snapshots or bi-temporal pairs. Understanding the continuous, long-term trajectory of landscape transformation demands more: a model must perceive land cover, recognize changes, memorize evolution histories, and reason across time and space. Prior benchmarks organized around application scenarios have shown that MLLMs struggle with long-term remote sensing, but not which underlying competency fails. ChronoBench is built to answer exactly that: inspired by human cognition, it isolates these competencies as four progressive levels, each framed by a core question, so that a failure at any level points to a specific missing capability:

LEVEL 0

Land Cover Perception

What is here at a given time?
OP
LEVEL 1

Temporal Recognition

What changed between specified time points?
BCC · BCAC · LCC · BOC
LEVEL 2

Long-Term Memory

When did changes occur, and what is the complete evolution history?
OAM · OCM · OHM
LEVEL 3

Spatio-Temporal Reasoning

What can be inferred by comparing across locations and time?
CCME · RDC · OCO · CSOCO

Each level withholds one more piece of information from the model. The pivotal step is from Level 1 to Level 2: in Level 1 the time points are given as inputs, whereas Level 2 requires the model to recall temporal information as outputs — the watershed between perceiving change and remembering history.


Abstract

Remote sensing offers an unparalleled vantage point for observing the Earth's long-term surface evolution, yet it demands that a model not only perceive land cover at isolated moments, but also track changes, memorize evolution histories, and reason across time and space. However, existing studies lack a systematic evaluation that dissects these distinct competencies. To fill this gap, we introduce ChronoBench, a multidimensional benchmark that decomposes this task into four progressive cognitive levels (i.e., Land Cover Perception, Temporal Recognition, Long-Term Memory, and Spatio-Temporal Reasoning). ChronoBench comprises 12 sub-tasks and 17,689 rigorously validated QA pairs. Extensive evaluations reveal that mainstream MLLMs fall drastically behind human experts, with Long-Term Memory emerging as the most critical bottleneck. Motivated by this finding, we further propose GeoChrono, an MLLM with enhanced capabilities for tracing, memorizing, and reasoning about long-term geographic evolution. Leveraging the physical prior that geographic parcels remain spatially fixed while their semantics evolve, we design a Temporal Trajectory Encoder (TempEnc) that constructs per-location temporal trajectories for dedicated land cover evolution modeling, and we introduce a Coarse-to-Fine Token Compressor (C2FComp) that adaptively preserves dynamic regions while compressing the static background. To support training, we also construct ChronoInstruct, a 104K-sample instruction-tuning dataset spanning all competency levels. GeoChrono achieves state-of-the-art performance on ChronoBench, surpassing the leading commercial MLLMs by over 20%, while C2FComp reduces visual tokens by over 56% while retaining 94.6% of GeoChrono's performance.


The ChronoBench Benchmark

3,469
high-resolution images (1024×1024)
500
regions across 39 U.S. cities
17,689
rigorously validated QA pairs
12
sub-tasks over 4 cognitive levels
7.26
temporal frames per QA on average
104,949
ChronoInstruct training samples

Explore the 12 Sub-Tasks

Object Perception OP

Identify the land cover at a specified location and timestamp, grounded by an HBB box or geographic coordinates.

Q: In 2013, what is the land cover type of the region grounded by ⟨580, 256, 769, 611⟩?
A: Vegetation
Bi-Temporal Class-Level Change Recognition BCC
Q: How has the vegetation changed comparing 2018 and 2021?
A: Vegetation → buildings; vegetation → non-vegetated surface
Bi-Temporal Class-Level Area Change Recognition BCAC
Q: How did the area of vegetation change comparing 2005 vs. 2023? Select one: expanded or reduced.
A: Reduced
Long-Temporal Class-Level Change Recognition LCC
Q: What changes has the non-vegetated surface undergone in the period of 2010 to 2022?
A: Non-vegetated surface → vegetation
Bi-Temporal Object-Level Change Recognition BOC
Q: Between 2018 and 2021, what land cover changes occurred within the region ⟨315, 107, 586, 449⟩-grounded vegetation? Select ALL that apply.
A: Vegetation → buildings; vegetation → non-vegetated surface
Object Appear Memory OAM
Q: Identify the interval during which the ⟨0, 123, 181, 326⟩-grounded vegetation in 2021 began its current continuous presence.
A: 2011–2013
Object Change Memory OCM
Q: Identify the interval during which the vegetation in 2008 grounded by ⟨196, 791, 344, 908⟩ changed to other land cover.
A: 2015–2017
Object History Memory OHM
Q: Provide the chronological sequence of distinct land cover phases for the building in 2023 referred to by ⟨210, 155, 277, 172⟩.
A: Vegetation → buildings
Class-Level Change Magnitude Estimation CCME
Q: In which interval did the vegetation undergo the sharpest change?
A: 2012–2014
Region Development Comparison RDC
Q: Among Region1 ⟨58, 19, 130, 293⟩, Region2 ⟨63, 465, 148, 647⟩, Region3 ⟨589, 130, 640, 321⟩ and Region4 ⟨700, 516, 862, 672⟩, which is the most developed from 2010 to 2021?
A: Region 3
Object Construction Ordering OCO
Q: Comparing the building#1 grounded by ⟨859, 200, 995, 278⟩ and the building#2 grounded by ⟨0, 478, 96, 612⟩ in 2023, which is built earlier?
A: Building #2
Cross-Sequence Object Construction Ordering CSOCO
Q: Comparing the building#1 at ⟨838, 79, 913, 163⟩ in Region#1 in 2021 with the building#2 at ⟨33, 943, 131, 999⟩ in Region#2 in 2023, identify which was constructed first.
A: Building #1
Task taxonomy of ChronoBench with example QA pairs for all 12 sub-tasks
Task taxonomy of ChronoBench. Seven spatially grounded sub-tasks provide both HBB-box and geo-coordinate variants; answers span single-choice, multiple-choice, and ordered-sequence formats.

Reproducible, Rule-Based Construction

All QA pairs are generated by deterministic rule-based programs that traverse pre-existing, human-annotated semantic change masks — no model-generated labels in the loop — followed by a two-stage human quality-control process with independent annotators and a domain-expert supervisor. The same pipeline scales up to ChronoInstruct, a 104,949-sample instruction-tuning dataset with multiple-choice, short-text, and free-form natural-language answers.

Overview of the data construction pipeline
Overview of the data construction pipeline.
QA distribution across levels and sub-tasks in ChronoBench
QA distribution in ChronoBench (17,689 pairs).
QA distribution across levels and sub-tasks in ChronoInstruct
QA distribution in ChronoInstruct (104,949 pairs).

How Well Do MLLMs Understand Long-Term Dynamics?

92.28% vs. 57.48%

A substantial human–machine gap. Human experts confirm the tasks are unambiguous and solvable, yet the best commercial model (Gemini-3-Flash) trails them by ~35 points; open-source models peak at 46.73%.

< 5% on OHM

Long-Term Memory is the critical bottleneck. On reconstructing complete evolution histories, open-source models nearly collapse (0.45–4.55%) and RS domain models stay below 2.5%, despite being trained on temporal RS data.

Scaling ≠ solution

Parameter scaling is inconsistent. Growing InternVL-3.5 from 4B to 14B barely moves memory and reasoning scores; brute-force scaling does not bridge the gap to human cognition.


The GeoChrono Model

Temporal remote sensing sequences differ fundamentally from natural video: geographic parcels remain spatially fixed while their semantics evolve, and only a small fraction of a wide-area scene changes meaningfully across frames. GeoChrono turns both physical priors into architecture: the Temporal Trajectory Encoder (TempEnc) is the core module that directly targets the Long-Term Memory bottleneck diagnosed above, while the Coarse-to-Fine Token Compressor (C2FComp) builds on it as a further innovation for lightweight computation in long-temporal scenarios:

GeoChrono framework: Temporal Trajectory Encoder and Coarse-to-Fine Token Compressor
Overview of the GeoChrono framework, built upon Qwen3-VL-4B-Instruct.

Temporal Trajectory Encoder (TempEnc)

TempEnc decouples the spatio-temporal feature volume into per-location temporal trajectories: after spatial context aggregation, tokens sharing the same spatial index are grouped across all frames and modeled with hybrid attention — half the heads bidirectional to capture global change patterns, half causal to preserve chronological progression — plus text-guided semantic focusing that filters task-irrelevant temporal information. TempEnc adds only 0.8% memory and 0.7% FLOPs, yet lifts overall accuracy from 72.52% to 78.34%.

Coarse-to-Fine Token Compressor (C2FComp)

Exploiting change sparsity, C2FComp scores each spatial block's relevance to the text instruction, keeps full-resolution tokens for salient regions, and condenses the static background into compact coarse tokens (differentiable top-K via Gumbel-Softmax). At the 1/4 selection ratio it removes 56.3% of visual tokens and 56.5% of FLOPs while retaining 94.6% of full-model performance — beating uniform 768×768 downsampling on both compression and accuracy.


🏆 ChronoBench Leaderboard

Accuracy (%) on ChronoBench. Bold green marks the best model result per column; the human row is a reference upper bound. Coord / Box denote geo-coordinate and HBB-box grounding variants.

Method Land Cover Perception Temporal Recognition Long-Term Memory Spatio-Temporal Reasoning OA
OPAVG BCACBCCLCCBOCAVG OAMOCMOHMAVG CCMERDCOCOCSOCOAVG
CoordBox CoordBox CoordBoxCoordBoxCoordBox CoordBoxCoordBox
Human98.5295.5697.0494.8182.9682.9698.5289.6389.7898.5297.0491.1187.4185.9390.3791.7378.5297.7899.26100.0099.2698.5295.5692.28
Commercial models
Gemini-3-Flash65.0665.9665.5182.0050.5352.4555.8556.1661.3856.4458.4446.6847.7525.0021.3647.5237.7963.6468.0578.0154.3666.6759.8957.48
GPT-5.447.0740.0043.5380.0050.7653.8766.1360.0067.5747.6941.5235.3646.0926.1420.0039.2130.7330.5153.8370.6052.4369.1250.4256.29
Seed-1.6-Vision50.9873.2362.1181.9146.2753.7347.0858.4763.8737.0847.0537.2241.8520.0018.4136.8631.2238.9856.5273.2955.8379.4153.7455.48
Open-source general models
InternVL-3.5-4B45.4158.6552.0354.8332.5727.9524.5124.2038.1922.0321.496.0910.201.360.9113.3322.7633.3354.6649.2850.4952.4542.0733.25
InternVL-3.5-8B33.0845.5639.3256.2928.3128.4935.6039.3543.2125.1124.3013.9116.951.603.4116.9330.4039.5551.3553.2157.7747.5545.1136.32
InternVL-3.5-14B40.9055.9448.4264.1533.4130.6636.4330.0845.6722.3921.7615.7619.070.450.9116.4521.9528.2552.3853.8348.0649.5141.4237.76
Qwen3-VL-4B35.1941.3538.2766.9431.1336.7723.9616.7342.0722.7618.9518.9422.522.954.5517.5421.3023.1647.8347.8354.8553.9239.5335.10
Qwen3-VL-8B39.4041.9640.6869.2129.1543.1531.9829.2647.1229.7428.5617.4820.002.950.9120.5223.7428.8154.6653.0055.8347.0642.8039.19
Qwen3-VL-32B42.8644.6643.7675.4244.2947.9045.6846.1657.8829.9230.7325.0327.953.864.5524.0628.7832.7751.7659.8359.7162.2547.2346.73
Remote sensing domain models
TEOChat-7B32.6336.9934.8148.8514.234.2117.2724.2530.0724.7527.834.116.621.591.1414.6417.4027.6850.9353.219.2222.0633.3526.82
EarthDial-4B36.9940.3038.6554.8318.044.3417.7224.0933.0826.2026.2017.4817.750.681.3618.5615.6122.0348.4446.3841.7552.4536.2530.11
DVLChat-4B48.5746.9247.7485.9031.8137.1834.2130.4654.4829.5626.9324.3729.932.502.0522.9117.2425.4250.7251.7658.7455.3940.5944.07
Ours
GeoChrono83.9193.3888.6593.6665.9176.6677.8380.2783.0372.7187.0468.6175.1031.5932.7368.1062.6061.5861.4992.5575.2492.1672.9278.34

Want your model listed? Open an issue or pull request with your predictions at IntelliSensing/GeoChrono.

Zero-Shot Generalization

MethodDVL-BenchCDVQA
BCA-SingleBCA-MultiCSE-Single
TEOChat35.18.717.050.0
EarthDial62.220.330.952.1
DVLChat64.921.331.343.7
GeoChrono72.942.135.759.2

C2FComp: Efficiency vs. Performance

Selection ratioFLOPsVisual tokensOAPerf. ratio
w/o C2FComp100%100%78.34100.0%
768×768 resize−48.75%−43.75%73.9594.4%
1/2−39.22%−37.50%74.5895.2%
1/4−56.53%−56.25%74.1194.6%
1/16−68.49%−70.31%72.6092.7%
1/64−71.34%−73.83%71.5691.4%

Qualitative Results

Side-by-side cases reveal how baselines fail: they default to visually salient classes when the queried target is small, over-select options in multi-choice change recognition, and anchor memory answers near the reference year instead of tracing back to the actual event. GeoChrono is the only model that answers all seven perception, recognition, and memory cases correctly.

Qualitative comparison on perception, recognition, and memory sub-tasks
Qualitative results on Land Cover Perception (OP), Temporal Recognition (BCC, LCC, BOC), and Long-Term Memory (OAM, OCM, OHM). GeoChrono is the only model that correctly answers all seven tasks.

GeoChrono is not perfect either. On Object History Memory, when the ground-truth evolution chain reaches length four or above, it exhibits phase omission: brief reversions to vegetation amid an overall development trajectory are skipped, because dormant vegetation in off-season imagery closely resembles non-vegetated surface. Sharpening this ambiguous class boundary is a key direction for long-horizon temporal memory.

Failure cases on Object History Memory showing phase omission
Representative failure cases on Object History Memory: intermediate vegetation phases are omitted from the predicted evolution chain.

BibTeX

@inproceedings{li2026geochrono,
  title     = {GeoChrono: Benchmarking and Rethinking Long-Term Temporal
               Understanding in Remote Sensing},
  author    = {Li, Yujie and Pan, Jiancheng and Wei, Zhiwei and Wang, Jiuniu
               and Peng, Mugen and Xu, Wenjia},
  booktitle = {Proceedings of the 34th ACM International Conference on
               Multimedia (ACM MM)},
  year      = {2026}
}