Preprint
Article

This version is not peer-reviewed.

Cloud-Aware EO–SAR Collection: Recovering Optical Passes with Radar Under Partial Cloud Cover

Submitted:

20 September 2026

Posted:

21 September 2026

You are already at the latest version

Abstract
Optical (EO) collection from low Earth orbit fails when cloud opacity exceeds the sensor threshold, so an EO-only policy wastes access windows under partial cloud. We study a small mixed constellation (3 EO + 2 SAR) tasking three fixed areas of interest (AOIs) under a spatiotemporal cloud field. EO succeeds only when site opacity is below 0.35; SAR succeeds in weather at lower quality weight (0.55). On a MIXED preset, EO-only greedy achieves 3.6 successful collects per episode and 33.3% confirmed AOI coverage, while a phenomenology heuristic that collects with EO when predicted clear and with SAR when cloudy reaches 33.8 collects/episode and 100% coverage—about +30 collects/episode and +67 percentage points in coverage. An ablation shows that removing the discrete tip channel changes throughput by only ∼0.2 collects/episode; the operational content is dual-sensor tasking, not the tip bits. Shared-policy proximal policy optimization (PPO) with local observations and invalid-action masking learns from untrained performance (∼3 collects/episode) to about 42 collects/episode on MIXED, matching or exceeding the heuristic. Under partial cloud, EO-only tasking loses most successful collects; a dual-sensor policy that uses SAR when EO is blocked recovers on the order of +30 collects/episode and full AOI coverage in this simulation. The discrete tip channel is secondary; shared PPO can learn the dual-sensor rule. Results are from a controlled simulation, not flight data.
Keywords: 
;  ;  ;  ;  ;  ;  ;  

1. Introduction

Maintaining useful Earth observation (EO) coverage from low Earth orbit (LEO) requires more than geometric access. Optical sensors are blocked by cloud; synthetic aperture radar (SAR) is weather-tolerant but typically offers different resolution, interpretability, and cost. Commercial and government operators increasingly describe multi-sensor tip-and-cue workflows in which one phenomenology identifies an opportunity and another is tasked for follow-up [3,4,5]. Much of the published literature on EO–SAR, however, addresses image fusion after collection rather than the closed-loop collection policy that decides which sensor should image under evolving cloud [6].
Cloud-aware optical scheduling has a substantial history, including offline/online frameworks that avoid cloudy targets for agile optical satellites [1] and multi-satellite planners that incorporate historical cloud statistics [2]. Reinforcement learning (RL) and multi-agent RL (MARL) have been applied to Earth-observation scheduling under resource and timing constraints [7,8,9]. What remains less quantified in a controlled mixed-sensor setting is the magnitude of the gap between EO-only greedy tasking and a dual-sensor rule when cloud opacity straddles the optical threshold—and whether a shared-policy learner with only local observations can recover that rule.
This paper isolates that question in a small simulation:
1.
a mixed EO–SAR environment with honest optical failure under cloud;
2.
empirical evidence that EO-only tasking collapses under partial cloud while a dual-sensor policy recovers on the order of + 30 collects per episode (the discrete tip channel is secondary);
3.
shared-policy PPO that learns the dual-sensor rule and can exceed the hand-coded phenomenology heuristic on MIXED.
We do not claim flight readiness, numerical weather prediction fidelity, or identity custody under inter-satellite link clutter; those are out of scope.

3. Problem Formulation

Consider N LEO satellites (default N EO = 3 , N SAR = 2 ) and K fixed AOIs (default K = 3 ). Time is discrete with step Δ t = 60  s and horizon H = 300 steps. Satellite i is in access of site k when the nadir angular separation is at most 16 ∘ . Each site has cloud opacity c k , t ∈ [ 0 , 1 ] . An EO collect at ( i , k , t ) succeeds if and only if c k , t < τ with τ = 0.35 . A SAR collect succeeds under cloud with quality weight q SAR = 0.55 and mild extra degradation proportional to opacity.
Primary metrics:
  • successful collects per episode;
  • quality-weighted collects (EO weight 1.0 , SAR weight ≈ 0.55 );
  • confirmed coverage: fraction of AOIs successfully imaged at least once.
With only three AOIs, coverage saturates for any competent SAR user; throughput and quality are more informative on CLEAR.

4. Methods

4.1. Environment and Cloud Field

The environment is a PettingZoo ParallelEnv. Orbits are circular Keplerian at 550 km altitude and 51 . 6 ∘ inclination with RAAN spread. AOIs are placed in a Mediterranean / North Africa theater (nominal locations near Tunis, Tripoli, and Antalya) with small jitter. Cloud opacity evolves on a latitude–longitude grid via an AR(1) temporal update, Gaussian spatial smoothing, eastward advection, and moment restoration so that the MIXED preset straddles τ . Local agents observe noisy cloud estimates only when in access (MIXED noise std. 0.08 ). Privileged mode sees true opacity and is used only as an upper reference.
Presets: CLEAR (mean 0.08 , std 0.06 ), MIXED ( 0.45 , 0.22 ), HEAVY ( 0.78 , 0.12 ).

4.2. Actions, Observations, and Reward

Discrete actions: idle; collect site k with own sensor; tip partner sensor type for site k. Tips are short-lived partner cues (TTL 8 steps), not global cloud broadcasts. Local observations include identity, EO/SAR type, energy, last success/fail, and per-site access, local cloud estimate, recent collect, and pending tip flags. Training reward shapes successful collects, confirmations, and penalties for cloud-failed EO, idle-in-access, redundant collects, and SAR-when-clear. Paper metrics are collect counts, quality, and coverage—not training return. PPO masks collect actions without access; without masking and a wasted-access penalty, early training collapsed to idle (not reported as a result).

4.3. Controllers

Heuristics (dedicated evaluation path; never labeled as PPO):
  • EO-only greedy: EO collects if access and predicted clear; SAR idles.
  • SAR-only greedy: SAR collects whenever in access; EO idles.
  • Phenomenology: predicted clear → EO; predicted cloud → SAR collect and/or tip; optional EO confirm.
  • Both sensors, no tip: same sensor rules with tips disabled.
  • Privileged oracle: true cloud with phenomenology structure (not deployable).
Shared-policy PPO: local observations, MLP, invalid-action masking. Iteration 0 is untrained weights on a separate RLlib path.

4.4. Experimental Protocol

Heuristic tables: seeds { 11 , 22 , 33 , 44 , 55 } , 300 steps. PPO training evaluation: seeds { 11 , 22 , 33 } . Primary scalars: successful collects/episode, quality-weighted collects/episode, confirmed coverage.

5. Results

5.1. Preset Comparison

On CLEAR, EO-only, phenomenology, and oracle match on throughput; SAR-only matches count but scores lower quality (19.1 vs. 35.2 quality-weighted collects/episode). On MIXED, EO-only collapses to 3.6 collects/episode while phenomenology recovers 33.8 (+30.2 collects/episode; +66.7 percentage points coverage). On HEAVY, EO-only records zero successes; SAR-capable policies remain near 35 collects/episode.
Table 1. Successful collects per episode (confirmed coverage in parentheses). Heuristics only; five seeds.
Table 1. Successful collects per episode (confirmed coverage in parentheses). Heuristics only; five seeds.
Method CLEAR MIXED HEAVY
Random 16.2 (100%) 10.0 (86.7%) 8.8 (86.7%)
EO-only greedy 35.2 (100%) 3.6 (33.3%) 0.0 (0%)
SAR-only greedy 35.0 (100%) 35.0 (100%) 35.0 (100%)
Phenomenology 35.2 (100%) 33.8 (100%) 35.0 (100%)
Privileged oracle 35.2 (100%) 37.2 (100%) 35.0 (100%)
Figure 1. Collect throughput and EO-weighted quality across CLEAR, MIXED, and HEAVY (heuristics only).
Figure 1. Collect throughput and EO-weighted quality across CLEAR, MIXED, and HEAVY (heuristics only).
Preprints 234170 g001

5.2. Cloud Sensitivity

EO-only records 35.2, 15.8, 3.6, and 0.0 collects/episode at mean opacities 0.08 , 0.35 , 0.45 , and ≥ 0.62 , respectively. Phenomenology remains near 34–39 across the sweep.
Figure 2. Successful collects per episode versus mean cloud opacity. Vertical line: EO threshold 0.35 . EO-only collapses past the threshold; phenomenology tracks SAR-capable policies.
Figure 2. Successful collects per episode versus mean cloud opacity. Vertical line: EO threshold 0.35 . EO-only collapses past the threshold; phenomenology tracks SAR-capable policies.
Preprints 234170 g002
Figure 3. MIXED ablation. Tips change throughput by ∼0.2 collects/episode; dual-sensor use explains the gap relative to EO-only.
Figure 3. MIXED ablation. Tips change throughput by ∼0.2 collects/episode; dual-sensor use explains the gap relative to EO-only.
Preprints 234170 g003

5.3. Ablation: Tips Versus Dual-Sensor Tasking

Removing tips costs about 0.2 collects/episode. The large gap versus EO-only is explained by using SAR when EO is blocked, not by the discrete tip messages.
Table 2. MIXED ablation (successful collects per episode).
Table 2. MIXED ablation (successful collects per episode).
Method Collects / episode
EO-only greedy 3.6
SAR-only greedy 35.0
Both sensors, no tip 33.6
Phenomenology (with tips) 33.8
Privileged oracle 37.2

5.4. Shared-Policy PPO on MIXED

On the same three seeds, EO-greedy scored 6.0 and phenomenology 34.7 collects/episode. Trained PPO exceeds the heuristic on MIXED. Return continued to rise after collect count saturated.
Table 3. Shared PPO on MIXED (evaluation seeds 11/22/33). Iteration 0 = untrained weights.
Table 3. Shared PPO on MIXED (evaluation seeds 11/22/33). Iteration 0 = untrained weights.
Iteration Collects / episode Notes
0 3.0 ± 0.8 untrained
1 22.0
2 42.0 coverage 100%
16 42.0 ± 4.3 quality 25.1; return 2734
Figure 4. PPO collects/episode and return versus iteration on MIXED. Iteration 0 is untrained. Dashed lines: EO-greedy and phenomenology.
Figure 4. PPO collects/episode and return versus iteration on MIXED. Iteration 0 is untrained. Dashed lines: EO-greedy and phenomenology.
Preprints 234170 g004

5.5. Example Rollout

Figure 5. Example MIXED phenomenology episode: AOI opacities, EO success/fail, SAR success, and tip events.
Figure 5. Example MIXED phenomenology episode: AOI opacities, EO success/fail, SAR success, and tip events.
Preprints 234170 g005

6. Discussion

Under partial cloud, EO-only tasking loses most successful collects once opacity exceeds the optical threshold. A dual-sensor policy that uses SAR when EO is blocked recovers on the order of $+30$ collects/episode and full AOI coverage on MIXED in this simulation. The ablation shows phenomenology with tips (33.8) and both-sensors without tips (33.6) are nearly identical, so the operational content is cloud-aware sensor selection, not the tip channel alone. SAR-only matches raw counts on MIXED/HEAVY but forgoes high-quality EO collects on CLEAR (quality 19.1 vs. 35.2). Local cloud estimates are noisy: EO success when attempted is about 29% for local EO-greedy versus about 80% for the privileged oracle on MIXED. Shared PPO learns the dual-sensor structure from an untrained baseline near EO-greedy to about 42 collects/episode, modestly above the hand-coded heuristic on the reported seeds. This is a simulation result, not a flight claim.

7. Limitations and Reproducibility

The propagator is circular Keplerian with a 16 ∘ access cone, not a full force model or sensor-specific FOV. Clouds are an AR(1) advecting field, not numerical weather prediction. SAR is a scalar all-weather quality weight, not NESZ or incidence geometry. There is no downlink, no airborne collector, and no real imagery. Heuristic tables use five seeds; PPO evaluation in the training loop uses three. Only shared-policy PPO is trained.
Reproducibility: package eosar_tip/. Heuristic tables and figures:
python -m eosar_tip.run_experiments --seeds 11 22 33 44 55 --steps 300
PPO:
python -m eosar_tip.train --preset mixed --iters 16
Smoke test:
python -m eosar_tip.smoke
Outputs under eosar_tip_results/.

8. Conclusion

Under partial cloud cover, EO-only tasking loses most successful collects once opacity exceeds the optical threshold. A dual-sensor policy that prefers EO when clear and SAR when cloudy recovers on the order of $+30$ successful collects per episode and full AOI coverage on MIXED in this simulation; the discrete tip channel is secondary to simply using both sensors. An ablation shows the discrete tip channel contributes little compared with simply using both sensors. Shared-policy PPO with local observations learns this structure and can exceed the hand-coded phenomenology heuristic on MIXED. The contribution is a controlled demonstration of cloud-aware EO–SAR collection policy, not a flight system or an image-fusion method.

Data Availability Statement

Code for the EO–SAR tip-and-cue environment is provided as the isolated package eosar_tip (Python 3.10+, PettingZoo, Ray RLlib, PyTorch). Public release via GitHub is planned upon acceptance; contact the author for details.

Conflicts of Interest

This work was supported by the University of Alaska Fairbanks Geophysical Institute. The author has no conflicts of interest to report.

References

  1. He, L.; Liu, X.; Xing, L.; Chen, Y. Cloud Avoidance Scheduling Algorithm for Agile Optical Satellites. In Bio-Inspired Computing – Theories and Applications; Springer: Berlin/Heidelberg, Germany, 2015; pp. 161–172. [Google Scholar]
  2. Wang, et al. Multi-satellite scheduling method based on a historical cloud coverage at regional scale. Prepr. J. Rec. To Be Confirmed From Publ. 2025. [Google Scholar] [CrossRef]
  3. Cudzilo, B.; Foley, K.C.; Smith, C. The Ability of a Small Satellite Constellation to Tip and Cue Other Commercial Assets. In Proceedings of the AIAA/USU Conference on Small Satellites, Logan, UT, USA, 2012. [Google Scholar]
  4. Kim, S.J.; Kim, M.; Choi, H.-L. Tip-and-cue mission scheduling in heterogeneous satellite constellations using mixed-integer linear programming. Acta Astronaut. 2026, 246, 921–940. [Google Scholar] [CrossRef]
  5. Weissman, G.; Ivry, A.; Cohen, I. An Automated Tip-and-Cue Framework for Optimized Satellite Tasking and Visual Intelligence. arXiv 2025, arXiv:2512.09670. [Google Scholar]
  6. Rabin, Z.; Scherreik, M.; Lewis, B.; Davis, J. Learning to fuse with side information for multi-sensor ATR. Proc. SPIE Def. Commer. Sens. 2024, Vol. 13032, 130320J. [Google Scholar]
  7. Hadj-Salah, A.; Verdier, R.; Caron, C.; Picard, M.; Capelle, M. Schedule Earth Observation satellites with Deep Reinforcement Learning. arXiv 2019, arXiv:1911.05696. [Google Scholar]
  8. Hady, M.A.; Hu, S.; Pratama, M.; Cao, J.; Kowalczyk, R. Multi-Agent Reinforcement Learning for Autonomous Multi-Satellite Earth Observation: A Realistic Case Study. arXiv 2025, arXiv:2506.15207. [Google Scholar]
  9. Holder, J.; Jaques, N.; Mesbahi, M. Multi Agent Reinforcement Learning for Sequential Satellite Assignment Problems. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025; Vol. 39. [Google Scholar]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.