Elizabeth Duffy is Product Marketing Director at EarthDaily; trained geospatial analyst, bridging the gap between EO and AI.
In the fast-changing field of artificial intelligence (AI), the importance of high-quality data can’t be overstated. AI models are only as good as the data they’re trained on, and so poor data can mean inaccurate predictions, unreliable outcomes and missed opportunities. Earth observation (EO) data is becoming increasingly critical across various sectors, with AI and machine learning (ML) enhancing its applications with many companies positioning this data as part of their “space strategy.” The combined revenue of AI use cases for image recognition, algorithmic trading strategies, and localization and mapping is $20 billion with a substantial portion supported with Earth observation data. As organizations advance their AI strategies and look to leverage the vast amounts of EO data that’s being generated, understanding quality becomes increasingly essential.
Today’s Challenges Leveraging EO Data For AI
The proliferation of low Earth orbit (LEO) satellites has brought a drastic increase in the volume and availability of EO data. It, however, creates its own challenges:
Data Quality Inconsistencies: The problem of varying acquisition times and angles hampers the trustworthiness of computer vision models. The ever-growing pool of EO data lacks consistency in spectral reflectance, which in turn creates major issues for data scientists.
Limited Spectral Diversity: Many current EO datasets lack the spectral diversity needed for nuanced insights, which reduces AI’s ability to detect subtle variations in land use, vegetation health and environmental changes.
Sensor Noise: Typically measured with signal-to-noise ratio (SNR), this parameter quantifies how much the signal has been corrupted by noise. A higher SNR indicates better image quality and SNR can be improved with digital denoising. While denoising techniques can improve data, they’re not a magical solution and inherently insert assumptions into the data. In short, there’s no replacement for original signal quality.
Poor data quality can have serious consequences on AI. A 2023 report published by Gartner estimates that operating costs may be lowered by up to 20% for businesses using data quality tools. Additionally, IBM claims that $3.1 trillion of America’s GDP is lost by U.S. companies annually as a result of poor-quality data.
How Quality Of EO Data Impacts AI
Investing in high-quality EO data has profound implications for AI model performance:
More Accuracy, Less Bias: With improved data, prediction models rely less on biases and yield better predictions.
Better Explainability Of Models: Comprehensive and consistent non-faulty data enable explanation-based modeling of AIs without compromising accuracy.
Practical Examples Of AI Built On EO
Agriculture And Commodities: AI can be used to estimate harvest quantities and monitor vegetation. For example, one company used an image recognition algorithm to determine tree height and type from 13 trillion pixels from satellite images in under two minutes, a task that would take seven years for a human analyst to complete. Put more tangibly, for a commonly used dataset like Sentinel-2, this would be roughly the equivalent of analyzing 90% of all the Earth’s landmass in the time it takes to heat up a cold cup of coffee in the microwave.
Disaster Response: AI-based EO information can be applied to supervise and detect natural disasters, and this will reduce the response time significantly.
Insurance And Financial Services: EO-AI approach allows insurers to better calculate risk and limit exposure to climatic-related risks.
Discerning Quality EO Data For AI
There are several parameters to look for in regard to EO data quality:
Global Coverage, Every Day: There are several new space constellations that cover the globe every day, at least with tasking missions, providing near real-time data coverage. However, it’s important to note that what may be advertised as daily global coverage is often only achievable through tasking requests and may not be readily available. Most revisit claims also don’t account for cloud coverage, which can make an optical image unusable. Though cloud coverage is difficult to account for, one metric EO users can look to is imaging time or when the data is acquired. As a general rule, images taken later in the day are more likely to be affected by cloud cover or haze.
Greater Dimensionality: With more spectral bands, AI-driven models are better able to find the finer changes in land use as well as shifts in the environment. As a general rule of thumb, it’s better to have greater dimensionality and winnow down with a technique like PCA than to have insufficient dimensions to properly capture key relationships in the data.
Uniformity In Spectral Reflectance: Satellites that acquire images at the same hour and angle are inherently less noisy and therefore the computer vision models’ accuracies improve. This approach delivers data with the highest possible direct comparability and the minimum of confounding effects from things like shadows. Beyond consistency in acquisition times, there are other standard signal quality measures such as the SNR.
Currently, there are some limitations in the EO market in what’s available. That said, there are new constellations launching regularly, and there are several government missions that mostly meet these metrics of quality.
The Value Of Maintaining Data Integrity: A Long-Term Investment
A course of action or product built on poor data integrity can expose a business to avoidable and extensive risks and financial costs. The global data volume stood at 120 zettabytes in 2023, and it’s predicted to nearly double to 221 zettabytes by 2026; therefore, robust AI techniques will need to be deployed so as to ensure optimal resource use.
In conclusion, reliable, sufficient and processed EO data is the cornerstone of capable, modern AI systems. Organizations that undertake risk and prioritize the quality of their data enhance their decision-making, minimize expenses and improve their position in the market. AI and EO are a perfect technological marriage, and it’s clear that organizations that invest in high-quality data will unlock new opportunities for innovation and growth.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
