Speaking Different Languages 

Communication can be hard. When you’re deep in a niche, it can be difficult to remember that the same word can mean many things. 

Many years ago, I was introduced to the concept of Transdisciplinary approaches to research.

In the world of geospatial consultation we find ourselves working within a new domain of expertise each project and I often find myself thinking about this approach. Even the difference between those in geospatial, data science, and programming can make it difficult to get on the same page. 

A Vector Says What?

Throughout my career, research, and life, this theme comes up often with one word in particular, “vector”, a polyseme. Meaning it is a single word that has multiple, related meanings. 

Across Sparkgeo, the word vector can take on different meanings depending on the domain, context, and person you’re speaking to. The most common use of the word “vector” is typically in a spatial or physics-based context, where a direction and distance are implied. A vector in geospatial can mean “anything that isn’t a raster,” a shape with nodes or edges that is stored as a “geometry”. While this definition fits most of our use cases, when interacting with clients or specialized PhDs in a specific field, we need to be mindful of what words may mean. 

Here are some examples of how it can be used across domains.

Natural Language Processing

In the context of natural language processing, a vector is a list of numbers used to represent a word or phrase or token in a way that captures its relationship to other words or phrases. Vectors are typically derived using techniques such as word embeddings, which map each word to a high-dimensional vector based on its co-occurrence with other words in a large corpus of text. You can think of vectors as a way of clustering similar words. Embeddings are also used for images, audio, and other types of data beyond just text. Embeddings are a foundational component of large language models (LLMs). These form the building blocks of a multi-modal model. 

Programming Language

In the programming language, “R”, a vector c() is a common data structure that can be used to create a vector or combine multiple vectors into one. In Python, a vector can mean a one-dimensional array in NumPy. 

Design

In design, a vector is a graphic made from mathematical paths (points, lines, and curves) instead of pixels. This means it can be scaled to any size without losing quality. Vectors are commonly used for icons, logos, and UI elements because they are clean, editable, and reusable across different screen sizes.

Sector-Specific

For sectors like public health or epidemiology, a “vector” describes how a pathogen is transmitted between people or animals (e.g. tick, mosquito). In security, an “attack vector” is the pathway of unauthorized access to target systems to commit a cyberattack. Common attack vectors include social engineering attacks, insider threats, and supply chain compromises.

Now imagine trying to have a coherent conversation with a data scientist, a designer, a security expert, a project manager, and a geospatial person about building a new geospatial foundational model.

The Rise of GeoAI

Inherently, GeoAI combines many modalities and types of data and therefore a transdisciplinary approach is required and ensures smooth communication. 

For example, we use Geospatial Foundational Models built on word embeddings and spatial vectors, or a mix of both. Talking about these concepts is difficult, because we are mixing a number of specialized domains into one (e.g. geospatial, remote sensing, street view imagery, natural language processing, data science, deep learning, computer vision, and agriculture). For most classically trained geospatial folks, bridging this gap can be overwhelming. At Sparkgeo, we are lucky to have a wide range of domain expertise, experience and specialist knowledge within our team.

Beyond the number of specialized domains that encompass GeoAI, there is also a large paradigm shift in the way we think about machine learning and designing machine learning systems. Instead of creating bespoke models for an individual computer vision task, now we can leverage pre-trained Foundational Models for a particular task, and hopefully, benefit from those learned representations, which lowers model training costs. Knowing when to use which approach and when, is what we are currently exploring at Sparkgeo.