Our latest webinar in the Agile EO series highlighted our breakthrough technology that uses custom tailored generative AI algorithms to enhance near-daily 3 m PlanetScope® imagery into sharp 2 m resolution. Watch the full recording now!
Below are questions from the live event and answers from our product experts.
1. Availability, Access, and Platforms
Q: Does SuperRes imagery only generate in areas where SkySat images have been acquired?
A: SuperRes imagery is available wherever PlanetScope imagery is available. We use SkySat® imagery to train the model, but once the model is running, it bases the output off PlanetScope data. Therefore, a SkySat collection is not required for an area to generate a SuperRes image.
Q: How can you access SuperRes?
A: PlanetScope SuperRes Scenes are available in the Visual Analysis (VA) tool on Planet Insights Platform or via API directly. Instructions on how to use via API are available in notebook examples. You will need to have a Planet Insights Platform plan with processing units (PUs) to access the tool.
Q. Can I test out SuperRes imagery using Planet Sandbox data?
A: You can test out SuperRes imagery using Planet Sandbox data in the Visual Analysis (BETA) tool before placing your own orders (detailed steps outlined below).
Follow these steps to test SuperRes in the Visual Analysis BETA tool using Planet Sandbox data:
- Select the "Planet Sandbox Data - PlanetScope" option.
- Zoom into one of the sample datasets.
- When the Visualize button lights up, click on it.
- Optional: Change date.
- Click the "Draw or upload an area of interest" icon on the right side of the map.
- In the "Area of interest" box, select one of the three right icons and create an AOI over the PlanetScope image shown on the map.
- In the "Area of interest" box, click “SuperRes.”
- The SuperRes layer will appear on the left and the SuperRes image will appear on the map.
- Optional: Toggle the SuperRes Layer and Confidence Layer on and off.
Q: How do you download SuperRes data?
A: You can download the SuperRes image once an AOI is provided (drawn or uploaded via AOI tool) or the download is available after a SuperRes image is processed in the VA tool.
Q: Will I be able to access SuperRes imagery with an Education & Research account?
A: At this time, Education and Research Program accounts do not have access to Planet SuperRes imagery.
Q: Does my AOI size limit my use of SuperRes?
A: There is no restriction on AOI size for SuperRes Mosaics. For SuperRes Scenes via the Visual Analysis (UI tool) the max size is 10 km2. For larger AOIs, users can use the Batch or Processing APIs.
Q: How much does this cost?
A: On-demand SuperRes scenes are charged using Processing Units (PUs). For SuperRes Mosaics, the pricing is based on a price per square kilometer. When you order a Mosaic, you receive the Visual Mosaic as well as the confidence score layer. For specifics, please reach out to our sales team.
2. Technical Methodology & AI Architecture
Q: Is there any reason you choose ESRGAN over other super-resolution algorithms?
A: It was chosen through empirical testing.
Q: Can you mention the model applied for the perceptual loss? Is it a pretrained model or an internal model?
A: The model is pretrained VGG-19.
Q: How does the model handle items that are uncommon in the training data, such as irregular curves on buildings?
A: The model will struggle with densely clustered structures, but does well in other scenarios.
Q: How does it handle shadows?
A: We are not aware of any issues with shadows.
Q: How does the system cope with transient features such as changing timelines or moving vehicles?
A: The model does show improvements on visual transient features, such as yachts in the water. The per-pixel confidence layer is the best tool to help you evaluate the level of trust you can place in those specific generated features.
Q: Did you experiment with 1.5 m or other spatial resolutions to test the limit of how far the super-resolution method could be applied?
A: Yes, we found no visual benefits going beyond 2 meters for Scenes and 2.4 meters at the equator for Mosaics.
Q: Could you use this as a tool for lunar mapping or is it trained specifically for Earth?
A: The model was trained for Earth use cases only.
3. Visual Perception vs. Spectral Accuracy
Q: Did you mention that you prioritize visual presentation over accuracy?
A: That is correct. We prioritize visual realism over exact spectral accuracy because this is a visual-first solution optimized for human perception.
Q: What is the difference between visual perception and pixel accuracy? Is confidence the metric that measures this difference?
A: Please refer to our technical overview blog for specifics.
Q. How does the SuperRes process use ESRGAN to ensure fidelity of the source PlanetScope imagery as it is transformed into the SuperRes image?
A: To ensure high fidelity, the model was trained with strict quality controls, such as temporal consistency, geometric alignment, and radiometric normalization. ESRGAN emphasizes restoring fine details, making it particularly effective for applications requiring sharp and artifact-free images. Refer to the Planet SuperRes: A Technical Overview blog for more specifics.
Q: How does SuperRes handle generation of hallucinations versus visual enhancement? e.g. the appearance of what looks to be solar panels on that first example in the webinar?
A: Planet SuperRes includes a confidence layer to help assess the degree of trust in the generated pixel by evaluating various sources of error. Refer to the Planet SuperRes: A Technical Overview blog for more specifics.
4. Data Specifications & Product Features
Q: Would this work for forests or for looking at crops?
A: For forests, beta customers have been generally impressed by the ability to distinguish individual trees that were not easily visible in standard 3 m imagery. For crops and farmland, it is primarily used for visual monitoring.
Q: Is this available for archive imagery or is this just for imagery going forward?
A: It is available for both archive and monitoring imagery. You can take imagery from the archive, order it to your data collection, and then apply SuperRes.
Q: How far back in the archives is it possible to augment images in this way?
A: The algorithm will work on PlanetScope SuperDove imagery - first launched in March of 2020.
Q: What are the spectral bands?
A: RGB
Q: Is this imagery ortho-rectified?
A: Yes, the imagery is ortho-rectified.
Q: Can you incorporate NIR bands to the SuperRes products?
A: SuperRes does not currently include NIR bands.
Q: Can SuperRes be applied to the basic, unrectified imagery, or only ortho-rectified imagery?
A: The model was trained on ortho-rectified imagery.
Planet SuperRes Resources
- Get started: If you're new to Planet, sign up for a free trial to test out Planet SuperRes or reach out to our sales team for more info. Planet customers, please reach out to your Planet representative.
- Dive deeper: Read our product blog and documentation.
- Explore more: Watch past Agile EO webinars
- Watch the webinar recording on-demand
Have another question? Drop it below!
