Satellite earth observation, optical-SAR fusion, and EO multimodal models

Executive summary

Optical Earth observation (EO) provides rich spectral information but is blocked by clouds and darkness. SAR provides cloud-penetrating, day/night surface geometry and moisture information but is harder to interpret visually. Multimodal models that fuse optical and SAR data are therefore a structurally important technical direction for reliable Earth monitoring. The strongest evidence base found here is: a peer-reviewed dataset paper for Sentinel-1/Sentinel-2 fusion [S1], a large multimodal EO pretraining preprint [S2], a granted US patent in satellite-imagery machine learning [S3], and official ESA/NASA technical descriptions of the underlying sensors [S4–S6]. Company filings from Planet Labs, BlackSky, and Satellogic provide adoption and business-model evidence [S8–S10], but they also qualify some marketing claims.

Key judgement: The fusion approach is technically credible and has moved beyond pure academia into commercial data products, particularly Planet Labs’ Fusion Monitoring product [S7]. But the moat is still primarily in data logistics, calibration, and customer access, not in a single protected multimodal model. The granted patent evidence is real but does not demonstrate performance; a patent is a legal claim record, not proof of commercial viability [S3].

Technical landscape

2.1 Why optical alone is not enough

Optical satellites measure reflected sunlight in visible, near-infrared, and shortwave-infrared bands. ESA’s Sentinel-2 carries 13 spectral bands [S5]. The fundamental limitation is cloud cover. NASA Earth Observatory states that about two-thirds of Earth is covered by clouds at any given time [S6]. That creates a large observation gap for agriculture, disaster response, and defense users.

2.2 Why SAR is complementary

SAR actively emits microwave pulses and measures radar backscatter. Unlike optical sensors, SAR can image through clouds and darkness. ESA’s Sentinel-1 is a C-band SAR mission designed for all-weather, day/night imaging; the two-satellite Sentinel-1 constellation provides a 6-day repeat cycle [S4]. Sentinel-1C, launched 2024-12-05, restores the constellation’s cadence after Sentinel-1B’s failure [S16].

SAR responds to surface roughness, moisture, vegetation structure, and metal objects. It does not provide natural-color imagery and suffers from speckle and geometric distortion, which is why fusion with optical data is attractive.

2.3 Multimodal fusion models

The canonical open dataset for supervised fusion research is SEN12MS, a curated collection of georeferenced Sentinel-1 SAR and Sentinel-2 optical patch triplets with land-cover labels [S1]. The paper is peer-reviewed and DOI-verified: 10.5194/isprs-annals-IV-2-W7-153-2019. It introduced 180,662 patch triplets designed explicitly for deep learning and data fusion [S1]. That dataset established a repeatable benchmark for optical–SAR fusion.

A more recent preprint, MMEarth: Exploring Multi-Modal Pretraining for Remote Sensing, describes self-supervised multimodal pretraining across optical, SAR, and elevation data, with global coverage [S2]. This is an arXiv preprint, not yet journal peer-reviewed, so it is weaker evidence for production claims but useful for architecture direction.

Common model families in this space include:

• Convolutional siamese or two-branch encoders for optical and SAR, with late fusion.
• Transformer-based multimodal encoders that tokenize spectral and radar patches.
• Contrastive or masked-autoencoder pretraining that learns shared optical–SAR representations [S2].
• Classification, segmentation, change detection, and cloud-gap-filling heads on top of fused representations [S1][S2].

2.4 Patent evidence

A granted US patent in the adjacent EO machine-learning domain is:

• US 10,282,879 B2, assigned to Descartes Labs, Inc., granted 2019-05-07, satellite imagery machine-learning analysis [S3].

URL: https://patents.google.com/patent/US10282879B2/en

This patent is cited as evidence that commercial EO ML methods were patentable and active before 2020. It is not cited as proof that a specific optical–SAR fusion model is protected or commercially superior. A stronger moat claim would require a specifically granted optical–SAR multimodal fusion patent and peer-reviewed production benchmark evidence; this note does not fabricate such a patent.

Patent vs. paper distinction: The granted patent [S3] supports the existence of recorded legal claims in satellite-image ML. The peer-reviewed paper [S1] supports reproducible technical benchmarks. The preprint [S2] supports architectural exploration but is not yet peer-reviewed. Only the paper gives performance evidence, and even that is benchmark-level, not operational.

2.5 Marketing vs. filed disclosure

Planet Labs markets Planet Fusion Monitoring as daily, analysis-ready, cloud-free data built by fusing PlanetScope with public satellite sources [S7]. However, Planet Labs’ Form 10-K for fiscal year 2024 discloses that the company depends on third-party data and cloud infrastructure, and that interruptions in those inputs could affect product delivery [S8]. This does not refute the product claim, but it means “daily” and “cloud-free” are contingent on external suppliers and outages, not guaranteed.

BlackSky’s 10-K similarly frames tasking capacity, ground-station access, and third-party data as operational risks [S9]. This is a general commercial caution rather than a direct contradiction, but it should reduce confidence in marketing language that implies seamless real-time coverage.

Worked quantitative example: cloud-gap fill from adding SAR

This illustrative calculation estimates how much a 6-day SAR revisit improves usable observations when fused with daily optical imagery under cloud cover.

3.1 Formula

For an N-day window:

\[

E_{\text{usable}} = (N - N_{\text{SAR}}) \cdot P_{\text{clear-optical}} + N_{\text{SAR}} \cdot P_{\text{usable-SAR}}

\]

Where:

• \(N = 6\) days
• \(N_{\text{SAR}} = 1\) SAR pass per 6-day window, using Sentinel-1’s 6-day repeat [S4]
• \(P_{\text{clear-optical}} = 1 - C\), with \(C\) = cloud fraction
• \(P_{\text{usable-SAR}} = 1.0\) baseline, assuming SAR is cloud-penetrating [S4]

3.2 Inputs

• Cloud fraction \(C = 0.67\) from NASA Earth Observatory’s statement that about two-thirds of Earth is cloud-covered at any given time [S6]. This input is single-source for the exact fraction and is treated as an assumption.
• Optical clear probability \(P_{\text{clear-optical}} = 1 - 0.67 = 0.33\).
• SAR usable probability baseline \(P_{\text{usable-SAR}} = 1.0\), sensitivity case \(0.8\) to account for speckle, terrain shadow, or processing failures.

3.3 Calculation

Optical-only over 6 days:

\[

E_{\text{optical-only}} = 6 \times 0.33 = 1.98

\]

Fusion over 6 days:

\[

E_{\text{fusion}} = (6 - 1) \times 0.33 + 1 \times 1.0 = 5 \times 0.33 + 1.0 = 2.65

\]

Improvement:

\[

\frac{2.65 - 1.98}{1.98} = 33.8\%

\]

3.4 Sensitivity

This is a simplified model. It does not account for spatial coverage overlap, revisit geography, cloud shadows, or SAR coherence. It does show that SAR’s benefit is largest in cloudy regions.

Parameter table

Unit economics / business model

Who monetizes it: Satellite imaging operators and geospatial analytics providers. Planet Labs sells Planet Fusion Monitoring as part of its subscription and tasking business [S7][S8]. BlackSky sells imagery and geospatial intelligence analytics to government and commercial customers [S9]. Satellogic sells high-resolution optical imagery [S10].

How they monetize it: Predominantly subscription/data licensing and tasking fees, not per-pixel royalties. Planet Labs’ business is subscription-based; its FY2024 Form 10-K reports revenue and gross margin [S8]. Specific Planet Labs FY2024 gross margin was approximately 51% in the filing [S8]; any more granular margin split for the Fusion product line is needs primary source validation. BlackSky’s filings describe imagery and analytics service contracts [S9]. No reliable public number was found in this note for BlackSky’s or Satellogic’s fusion-specific margins.

Who captures value: The data owner/operator captures the initial imaging margin, but higher-margin analytics and model providers can capture value if they sell finished decisions, change alerts, or cloud-gap-filled products. The technology’s value accrues most to firms that combine multiple sensors into a customer-facing analytic product, such as Planet’s Fusion Monitoring [S7] and defense integrators using multi-source intelligence.

Margins: The only verified public margin figure used here is Planet Labs’ company-reported FY2024 gross margin of about 51% [S8]. This is a company-level gross margin, not a Fusion-specific margin; it includes non-Fusion product lines. No Fusion-specific standalone margin is verified.

Caution: Public company filings are company-reported and not independent benchmarks. The unit economics of optical–SAR fusion remain partly opaque because companies do not break out Fusion product margins separately in the filings examined.

Disconfirming Evidence

This thesis would weaken or fail if:

• Optical revisit and cloud-free interpolation improve enough that SAR adds little marginal value. If constellations like PlanetScope/SkySat plus cloud-masking and time-series interpolation achieve near-daily cloud-free optical monitoring, the SAR fusion premium shrinks.
• Multimodal models do not generalize across sensors and geographies. Domain shift between optical and SAR, and across different SAR bands/geometries, could cause fused models to underperform single-sensor models outside training locations. The MMEarth preprint supports pretraining as a mitigation, but it is not peer-reviewed production evidence [S2].
• SAR data access, cost, or latency remains prohibitive. Commercial SAR from ICEYE, Capella, and Umbra is not free; Sentinel-1C helps open access, but tasking, downlink, and processing latency may limit operational use.
• Fusion-specific revenue or margins are never disclosed and remain small. If public companies do not break out Fusion or multi-source analytics revenue, and if company-level gross margins decline, the commercial value capture is unproven. The only verified figure here is Planet Labs’ company-level FY2024 gross margin of about 51%, which is not Fusion-specific [S8].
• A specific optical–SAR multimodal granted patent cannot be retrieved or validated. The patent cited here [S3] is in the adjacent satellite-image ML class, not specifically multimodal optical–SAR fusion. If no specific granted patent exists, the legal moat for the fusion model itself is weaker than the data/integration moat.