lgnd: ChatGPT for the Earth - AI for Planetary Health

The Abundance of Earth Observation Data

Our planet is currently generating a vast quantity of data pertaining to its own characteristics. Daily, orbiting satellites collect approximately 100 terabytes of image data.

Challenges in Data Interpretation

However, extracting meaningful insights from this data isn't always straightforward. Even seemingly simple inquiries can present significant analytical hurdles. Consider the question of critical economic importance to California: What is the current number of fire breaks within the state capable of halting wildfire progression, and how has this number evolved since the previous fire season?

Traditional Methods and Their Limitations

“Historically, this task relied on manual image analysis by human observers,” explains Nathaniel Manning, co-founder and CEO of LGND, in a discussion with TechCrunch. “This approach, however, possesses inherent scalability limitations.” Recent advancements in neural networks have offered some improvement, enabling machine learning specialists and data scientists to develop algorithms capable of identifying fire breaks within satellite imagery.

The Cost of Traditional Machine Learning

“The development of such a dataset typically requires an investment in the range of a few hundred thousand dollars – potentially exceeding that amount – and is limited to addressing that specific task,” he stated.

LGND's Approach to Efficiency

LGND aims to substantially reduce these costs, potentially by an order of magnitude or more.

Augmenting, Not Replacing, Human Expertise

“Our intention is not to supplant individuals currently performing these analyses,” clarifies Bruno Sánchez-Andrade Nuño, LGND’s co-founder and chief scientist. “Rather, we seek to enhance their efficiency, increasing it tenfold or even a hundredfold.”

Recent Funding and Investors

LGND has recently secured $9 million in seed funding, led by Javelin Venture Partners, as exclusively reported to TechCrunch. Additional investors include AENU, Clocktower Ventures, Coalition Operators, MCJ, Overture, Ridgeline, and Space Capital. The round also saw participation from several angel investors, such as John Hanke, founder of Keyhole, Karim Atiyeh, co-founder of Ramp, and Suzanne DiBianca, an executive at Salesforce.

Core Technology: Geographic Vector Embeddings

The startup’s primary offering centers around vector embeddings of geographic data. Currently, most geographic information is represented either as pixels or traditional vectors (points, lines, and polygons). While these formats are versatile and easily distributed, interpreting the underlying information often demands specialized spatial knowledge or considerable computational resources.

Simplifying Spatial Data Analysis

Geographic embeddings provide a condensed representation of spatial data, facilitating the identification of relationships between different locations on Earth.

The Power of Upfront Computation

“Embeddings perform the majority of undifferentiated computation upfront,” Nuño explains. “They serve as concise, universal summaries that encapsulate 90% of the computational effort typically required.”

Applying Embeddings to Fire Break Detection

Consider the case of fire breaks. These can manifest as roads, rivers, or lakes. Each will appear uniquely on a map, yet they share common attributes. For instance, the pixels composing an image of a fire break will lack vegetation. Furthermore, a fire break must maintain a minimum width, often correlated with the surrounding vegetation height. Embeddings streamline the process of locating areas on a map that meet these criteria.

LGND's Product Offerings

LGND offers both an enterprise application designed to assist large organizations in answering complex spatial data questions, and an API for users with more specialized requirements.

New Possibilities in Geospatial Querying

Manning envisions LGND’s embeddings enabling companies to formulate entirely new types of geospatial queries.

An Example: The AI Travel Agent

He illustrates this with the example of an AI-powered travel agent. Users might request a short-term rental with three bedrooms near excellent snorkeling opportunities. “However, they might also specify a preference for a white sand beach, minimal seaweed in February, and, crucially, the absence of construction within a one-kilometer radius of the property at the time of booking,” he adds.

The Inefficiency of Traditional Geospatial Models

Constructing traditional geospatial models to address these queries would be a time-intensive process, even for a single request, let alone a multitude of them.

Market Potential and Future Vision

If LGND succeeds in making such a tool widely accessible, even to professionals who routinely work with geospatial data, it could capture a significant share of a market currently valued at approximately $400 billion.

A Standard for Geospatial Data

“Our ambition is to become the foundational standard for this type of data,” Manning concludes.