CS-6: Autonomous Earth Observation CubeSat

• CogniSAT-6/HAMMER (CS-6) is a 6U CubeSat designed to demonstrate onboard AI by performing real-time hyperspectral analysis and dynamic targeting for Earth observation.
• It integrates a VNIR hyperspectral imager with an Intel Myriad X VPU, enabling deep learning inference for tasks like cloud detection and anomaly identification within 60–90 seconds.
• The mission achieves high accuracy in cloud (97.5%) and thermal event (99.9%) detection, showcasing reduced data volume strategies and agile retargeting capabilities.

CogniSAT-6/HAMMER (CS-6) is a 6U Earth-observation CubeSat developed as a flight demonstration of onboard inference and spacecraft autonomy for Earth science. Published descriptions emphasize two tightly coupled capabilities: first, in-orbit execution of spectral analysis algorithms and deep learning on visible and near infrared imagery; second, dynamic targeting (DT), in which sensor data is acquired, rapidly analyzed, and used to drive subsequent observation during the same orbital pass. In the reported implementation, CS-6 combines a visible and near infrared range hyperspectral instrument with neural network acceleration hardware, with the stated aim of enabling rapid responses to events, reduced data volume, and observing modes such as dynamic targeting and cross-satellite cueing (Zilberstein et al., 20 Aug 2025, Chien et al., 22 Aug 2025).

1. Platform configuration and mission context

CS-6 is described as a 6U CubeSat in sun-synchronous orbit at approximately 500 km altitude, launched in March 2024; one account further specifies launch on SpaceX Transporter-10 (Zilberstein et al., 20 Aug 2025, Chien et al., 22 Aug 2025). The mission is presented as a collaboration involving the Jet Propulsion Laboratory and Ubotica Technologies, while the DT work refers to the spacecraft as CogniSAT-6 (Ubotica/Open Cosmos). The common technical theme across these descriptions is the use of edge computation to move portions of the analytic pipeline from the ground segment onto the spacecraft.

The payload characterization in the onboard-inference report centers on a HyperScape 100 hyperspectral instrument spanning 440 nm to 884 nm in the visible and near infrared (VNIR), with 5 meters ground sample distance and support for multi/hyperspectral imaging campaigns for Earth observation (Zilberstein et al., 20 Aug 2025). The DT report describes the operational sensor arrangement somewhat differently: CS-6 is said to have a single imaging sensor with red, green, blue, and very near-infrared bands, used both for lookahead imaging and for the primary observation (Chien et al., 22 Aug 2025). This suggests an operational distinction between the full instrument capability and a reduced-band acquisition mode used to satisfy DT latency constraints.

The mission context is explicitly Earth-science oriented. The cited applications include cloud avoidance, storm hunting, search for planetary boundary layer events, plume study, thermal anomaly detection, flood assessment, vegetation and mineral mapping, land use/land cover mapping, and harmful algal bloom detection. In this framing, CS-6 is not only a sensor platform but also a testbed for autonomous selection, prioritization, and summarization of scientifically relevant observations.

2. Onboard compute architecture and edge-processing model

The central onboard compute element is the Intel Myriad X Vision Processing Unit (VPU), identified as the neural network acceleration hardware on CS-6 (Zilberstein et al., 20 Aug 2025). It is characterized as capable of high-speed computer vision, image signal processing, and neural network inference in orbit, and as specifically designed for edge AI inference and onboard computer-vision workloads (Zilberstein et al., 20 Aug 2025, Chien et al., 22 Aug 2025). In the mission architecture, the VPU is the enabling component for rapid, low-latency analysis performed on orbit rather than after downlink.

The compute model is constrained by orbital geometry and spacecraft operations. The DT report states that, at an orbital velocity of 7.5 km/s at approximately 500 km altitude, the total latency from lookahead acquisition to targeted imaging is roughly 60–90 seconds, and that acquisition, transfer, preprocessing, analysis, command generation, and slew must all complete within this window (Chien et al., 22 Aug 2025). The mission therefore relies on algorithmic efficiency, reduced data movement, and simplified acquisition modes such as decimation and band reduction.

A notable implementation detail in the deep-learning pipeline is that preprocessing is embedded as custom layers, including normalization and quantile-based band stretching, specifically to minimize CPU processing (Zilberstein et al., 20 Aug 2025). This design choice aligns with the broader edge-computing objective: keep data local to the accelerated inference pathway, reduce host-side overhead, and preserve enough time budget for retargeting and follow-up acquisition. The reported effect is not merely faster inference, but a restructuring of the end-to-end observation pipeline around low-latency decision support.

3. Spectral analysis and deep-learning algorithms

The onboard algorithm stack on CS-6 consists of two main classes: classical spectral analysis and convolutional neural networks. The spectral methods are used for mineral and vegetation mapping as well as anomaly or outlier detection, while the CNNs are used for semantic segmentation tasks in multiple Earth-science domains (Zilberstein et al., 20 Aug 2025).

The spectral analysis methods explicitly identified are Spectral Angle Mapper (SAM), Matched Filter (MF), and the Reed-Xiaoli (RX) anomaly detector. Their reported formulations are:

$$SAM(x, y) = \cos^{-1} \left( \frac{x \cdot y}{\|x\|^2 \|y\|^2} \right)$$

$$MF(x, t; \mu, \Sigma) = \frac{(t - \mu)^T \Sigma^{-1} (x - \mu)}{(t - \mu)^T \Sigma^{-1} (t - \mu)}$$

$$RX(x; \mu, \Sigma) = (x - \mu)^T \Sigma^{-1} (x - \mu)$$

In the supplied descriptions, lower SAM values indicate greater similarity, higher MF response indicates stronger target match, and high RX values identify spectral outliers (Zilberstein et al., 20 Aug 2025). The DT report additionally names spectral unmixing among the methods used for spectral signature detection, while the onboard-inference report states that spectral unmixing using deep learning is being prototyped (Chien et al., 22 Aug 2025, Zilberstein et al., 20 Aug 2025).

The primary deep-learning architecture is U-Net, tailored for efficient execution on flight hardware and accelerated by the Myriad X VPU (Zilberstein et al., 20 Aug 2025). The reported application set includes cloud detection using Haze Optimized Transform (HOT), surface water extent (SWE) using Normalized Difference Water Index (NDWI), thermal event detection for volcanoes and wildfires, land use/land cover mapping across classes such as city, forest, water, and cropland, and harmful algal bloom detection (Zilberstein et al., 20 Aug 2025). The DT work also refers more generally to deep convolutional neural networks for cloud and anomaly classification (Chien et al., 22 Aug 2025).

Training and dataset construction are reported with substantial specificity. Binary segmentation tasks are optimized with sparse categorical cross-entropy, or weighted sparse categorical cross-entropy for thermal detection to address class imbalance. Automated labeling is derived from HOT, NDWI, and external reference datasets including WorldCover and Sentinel-2, and the datasets cited for transfer learning and cross-validation include CS-6, Menut, and Planetscope imagery (Zilberstein et al., 20 Aug 2025). These choices indicate a workflow in which physically motivated indices and external cartographic products are used to bootstrap labels for supervised onboard models.

4. Dynamic targeting workflow

Dynamic targeting on CS-6 is defined as a spacecraft autonomy concept in which sensor data is acquired and rapidly analyzed and used to drive subsequent observation (Chien et al., 22 Aug 2025). In the low Earth orbit scenario described for CS-6, lookahead imagery is analyzed to detect clouds, thermal anomalies, or land use cases so as to drive higher-quality near-nadir imaging. The DT use cases named in the report include cloud avoidance, storm hunting, search for planetary boundary layer events, plume study, and beyond (Chien et al., 22 Aug 2025).

The representative DT sequence begins with acquisition of a lookahead image, obtained by pointing the imager 40–50° ahead along-track, corresponding to roughly 60–90 seconds before nadir overpass (Chien et al., 22 Aug 2025). The image is then transferred to the onboard processing unit, where preprocessing includes decimation, band selection, and simple dynamic range stretching. Onboard analysis then runs event or anomaly detection algorithms. If an event of interest is detected, a retargeting or repointing command is issued, after which CS-6 slews from lookahead to nadir view or another targeted observation angle. The report adds that slewing is initiated as soon as possible and may begin before all computations complete if the timeline is constrained (Chien et al., 22 Aug 2025).

The same sensor is used for both lookahead imaging and main observation. Because CS-6 lacks a dedicated wide-field lookahead imager, the system compensates through off-nadir pointing and operating modes such as coarser spatial resolution emulation by pixel binning or decimation, together with fast readout for larger pixel coverage and lower data volume (Chien et al., 22 Aug 2025). This is an important architectural feature: the DT capability does not depend on a separate scout sensor, but on rapid re-use of the primary imager under different acquisition modes.

The workflow extends beyond retargeting to communications and dissemination. Processed images and analyses can be downlinked through standard paths, and the reports describe the potential use of inter-satellite or low-latency communication links for rapid delivery and cross-platform tasking (Chien et al., 22 Aug 2025, Zilberstein et al., 20 Aug 2025). A plausible implication is that CS-6 treats onboard inference not only as a local decision aid, but as a possible trigger for distributed observation strategies.

5. Demonstrated results and reported performance

The onboard-inference report provides quantitative results for several semantic-segmentation tasks and for onboard execution time on the Myriad X VPU (Zilberstein et al., 20 Aug 2025).

Function	Reported metric	Reported value
Clouds	Accuracy	97.5%
Clouds	Positive IoU	0.91
Surface Water Extent (SWE)	Accuracy	87.3%
Surface Water Extent (SWE)	Positive IoU	0.71
Thermal Detection	Accuracy	99.9%
Thermal Detection	Positive IoU	0.97
CNNs (U-Net, ~4.5 MB)	Inference time per image	~0.48–0.53 s
Spectral algorithms (SAM/MF/RX, 3–177 KB)	Inference time per image	2.5–18 s

The same report states that higher-dimensionality inputs increase spectral-algorithm runtime and that execution on the VPU and CPU is near-identical, verified on ground hardware (Zilberstein et al., 20 Aug 2025). The latter result is significant because it frames the VPU deployment as a flight implementation of algorithms whose outputs were checked against a non-accelerated execution path, rather than as a qualitatively different approximation.

The applications demonstrated or explicitly targeted on CS-6 include cloud detection and screening, surface water extent mapping, thermal event detection, vegetation and mineral mapping, land use classification, and harmful algal bloom detection (Zilberstein et al., 20 Aug 2025). In DT mode, the practical emphasis is on acquiring lookahead evidence quickly enough to improve follow-up science collection during the same overpass (Chien et al., 22 Aug 2025). The scientific value therefore lies not only in classification accuracy, but also in whether the inference can be completed early enough to alter spacecraft behavior.

6. Scientific significance, constraints, and prospective extensions

The reported significance of CS-6 is framed around edge AI for Earth observation. The onboard-inference report states that CubeSats with dedicated AI hardware can autonomously analyze high volumes of hyperspectral data, enabling observing strategies such as dynamic targeting, cross-mission coordination, and real-time trigger actions (Zilberstein et al., 20 Aug 2025). It also emphasizes bandwidth optimization: only regions or summaries of interest may need to be downlinked, thereby reducing bandwidth requirements and storage pressure. Within that framing, CS-6 functions as a technology-readiness demonstration for machine-learning-based onboard inference.

Several constraints are also explicit. CS-6 was not designed for extremely fast sensor readout, and the lack of a wide-field lookahead sensor increases DT timeline pressure; consequently, pixel decimation, band reduction, and intelligent band selection are described as key tactics (Chien et al., 22 Aug 2025). Slews of 40–50° are characterized as significant, even though tests are reported to show that CS-6 has sufficient agility (Chien et al., 22 Aug 2025). Thermal event detection is also described as operating via thermal spectral proxies within VNIR limits, which constrains interpretation relative to dedicated thermal infrared sensing (Zilberstein et al., 20 Aug 2025).

These constraints address a common misconception about onboard autonomy in Earth observation: the technical challenge is not only whether a model can classify an image, but whether the full chain of sensing, transfer, preprocessing, inference, guidance, and communications can close inside a hard real-time window. Another misconception is that DT necessarily requires a dedicated lookahead payload. The CS-6 implementation instead uses a single imaging sensor for both lookahead and main observation, relying on off-nadir viewing and modified readout modes (Chien et al., 22 Aug 2025).

The forward-looking vision in both reports is a networked one. The DT paper describes future missions in which spacecraft collaboratively retask using inter-satellite communications, while the onboard-inference report states that models and methods are being ported for use on additional satellites, forming the backbone for constellations of autonomous, collaborative Earth-observing assets (Chien et al., 22 Aug 2025, Zilberstein et al., 20 Aug 2025). This suggests that CS-6 is best understood not as an isolated instrument demonstration, but as an architectural precursor for distributed, inference-driven remote sensing systems.