1 of 9

A LLM-Driven GeoAI Pipeline for Sustainable Development Indicators: Conversational Insights

SDL Internship Program Presentation

Liu Zhao

liu.zhao@utas.edu.au

Mentor: Dr. Devika Jain

Dr. Yongze Song

2 of 9

Outline

Background
Aim and Objective
Research Objectives
Methodology
Workflow Overview
Results
Future things can do

3 of 9

Sustainable Development Goal (SDG) indicators are used to measure global progress in economic, social, and environmental domains. However, due to the vast amount of data, diverse sources, and complex analysis required, efficiently extracting and utilizing these indicators remains a challenge.

In recent years, the advancement of Large Language Models (LLMs) and Geospatial Artificial Intelligence (GeoAI) has provided new possibilities for SDG indicator analysis. Traditional geospatial analysis relies on complex queries and specialized tools, creating a high entry barrier. LLMs, however, enable natural language interactions, lowering the difficulty of data access and analysis.

This study develops an LLM-driven GeoAI pipeline that allows users to obtain spatial analysis results for SDG indicators through conversational queries. The pipeline integrates PostGIS for geospatial data processing and LLMs for natural language query interpretation, enabling intelligent querying, dynamic visualization, and in-depth insights.

This system not only enhances the interpretability of SDG data but also improves policy-making, environmental monitoring, and socio-economic assessments. Through conversational interactions, users can explore and analyze SDG indicators without requiring specialized expertise, fostering more transparent and data-driven decision-making.

Background

https://databank.worldbank.org/source/sustainable-development-goals-(sdgs)

Sustainable Development Goal (SDG) indicators are used to measure global progress in economic, social, and environmental domains. However, due to the vast amount of data, diverse sources, and complex analysis required, efficiently extracting and utilizing these indicators remains a challenge.

In recent years, the advancement of Large Language Models (LLMs) and Geospatial Artificial Intelligence (GeoAI) has provided new possibilities for SDG indicator analysis. Traditional geospatial analysis relies on complex queries and specialized tools, creating a high entry barrier. LLMs, however, enable natural language interactions, lowering the difficulty of data access and analysis.

This study develops an LLM-driven GeoAI pipeline that allows users to obtain spatial analysis results for SDG indicators through conversational queries. The pipeline integrates PostGIS for geospatial data processing and LLMs for natural language query interpretation, enabling intelligent querying, dynamic visualization, and in-depth insights.

This system not only enhances the interpretability of SDG data but also improves policy-making, environmental monitoring, and socio-economic assessments. Through conversational interactions, users can explore and analyze SDG indicators without requiring specialized expertise, fostering more transparent and data-driven decision-making.

4 of 9

Develop an LLM-driven GeoAI pipeline that enables users to explore, analyze, and visualize Sustainable Development Goal (SDG) indicators through natural language interactions, making complex geospatial data more accessible and actionable.

Design a natural language query system that translates user inputs into geospatial database queries for SDG analysis.
Implement geospatial data processing using PostGIS to support spatial queries and generate location-based insights.
Enhance data interpretability by integrating dynamic visualizations and trend analysis powered by GeoAI techniques.
Improve accessibility and usability by enabling non-experts to explore and analyze SDG indicators through conversational AI.
Validate the pipeline’s effectiveness through case studies, ensuring accuracy, efficiency, and real-world applicability.

Aim and Objective

My aim is Develop an LLM-driven GeoAI pipeline that enables users to explore, analyze, and visualize Sustainable Development Goal (SDG) indicators through natural language interactions, making complex geospatial data more accessible and actionable.

And Objectives

Is �Design a natural language query system that translates user inputs into geospatial database queries for SDG analysis.

Implement geospatial data processing using PostGIS to support spatial queries and generate location-based insights.

Enhance data interpretability by integrating dynamic visualizations and trend analysis powered by GeoAI techniques.

Improve accessibility and usability by enabling non-experts to explore and analyze SDG indicators through conversational AI.

Validate the pipeline’s effectiveness through case studies, ensuring accuracy, efficiency, and real-world applicability.

5 of 9

Methodology

GPT-4o

NL2SQL

Prompt Engineering

PostGIS

Respond

Leafmap

RAG

This study builds a LLM-driven GeoAI pipeline that allows users to explore Sustainable Development Goal (SDG) indicators using natural language queries. This system integrates GPT-4o, Retrieval-Augmented Generation (RAG), Natural Language to SQL (NL2SQL), PostGIS for geospatial processing, and Leafmap for visualization.

First, the user enters a natural language query, such as “Which countries have the highest forest coverage?”

The system then applies prompt engineering to refine the query, ensuring it is correctly understood by GPT-4o.

Once the query is processed, GPT-4o decides how to handle it:

• If the question requires text-based information, it uses RAG to retrieve relevant knowledge.

• If the query involves structured data, it converts it into an SQL query using NL2SQL.

For general SDG-related questions, RAG retrieves relevant documents and GPT-4o generates a response based on factual data.

If the query requires data analysis, GPT-4o translates the question into a structured SQL query.

The SQL query is executed in PostGIS, which processes spatial SDG data, such as population distribution, environmental changes, and economic trends.

The processed data is then visualized using Leafmap, generating interactive maps to help users better understand geospatial trends.

Why This Approach?

✅ No need for SQL or GIS expertise – anyone can explore SDG indicators easily.

✅ Combines LLMs and geospatial AI – making data analysis more intelligent and interactive.

✅ Supports both text-based and spatial data queries – allowing users to get clear insights.

6 of 9

Workflow Overview

Prompt Engineering

User inputs a natural language query related to SDG indicators.

GPT-4o Processing

GPT-4o interprets the query and determines whether it requires:

A structured data query (processed through NL2SQL for PostGIS.

Retrieval-Augmented Generation (RAG) Module

GPT-4o generates a natural language response using the retrieved knowledge.

NL2SQL Query Generation

The SQL query is optimized to fetch relevant SDG indicators from a PostGIS database.

PostGIS Geospatial Processing

The generated SQL query is executed in PostGIS, retrieving spatially relevant SDG data.

Leafmap-Based Visualization

The retrieved geospatial data is passed to Leafmap, which generates interactive maps and visualizations.

7 of 9

Results

�

The system successfully processed queries and generated accurate SQL statements.

• First, it executed “Query all data” to retrieve the entire dataset.

• Then, it ran “Query basic country information” to extract key attributes like name, continent, region, and geometry.

• Next, it processed “Query most populated countries”, filtering countries with a population over 100 million.

�

For more complex queries:

• It successfully translated “Which countries have high fertility rates but low life expectancy?” into an SQL statement.

• It also generated the correct SQL for “Which countries have high adolescent fertility rates and high child mortality?”

�

Each query was correctly converted into structured SQL, demonstrating the system’s ability to handle both basic and advanced queries effectively.

8 of 9

Results

Query Description	SQL Query
Basic
Query all data	SELECT * FROM "SDG"
Query basic country information	SELECT "name", "continent", "region", "geometry" FROM "SDG"
Medium
Query most populated countries	SELECT "name", "Population, total", "geometry" FROM "SDG" WHERE "Population, total" > 100000000 LIMIT 100
Query low fertility rate countries	SELECT "name", "Fertility rate, total (births per woman)", "geometry" FROM "SDG" WHERE "Fertility rate, total (births per woman)" < 2 LIMIT 100
Query countries with high forest coverage	SELECT "name", "Forest area (sq. km)", "geometry" FROM "SDG" WHERE "Forest area (sq. km)" > 1000000 LIMIT 100
Query countries with high net migration	SELECT "name", "Net migration", "geometry" FROM "SDG" WHERE "Net migration" > 100000 LIMIT 100
High
Which countries have high fertility rates but low life expectancy?	SELECT "name", "Fertility rate, total (births per woman)", "Life expectancy at birth, total (years)", "geometry" FROM "SDG" WHERE "Fertility rate, total (births per woman)" > 4 AND "Life expectancy at birth, total (years)" < 65 LIMIT 100
Which countries have high adolescent fertility rates and high child mortality?	SELECT "name", "Adolescent fertility rate (births per 1,000 women ages 15-19)", "Mortality rate, under-5 (per 1,000 live births)", "geometry" FROM "SDG" WHERE "Adolescent fertility rate (births per 1,000 women ages 15-19)" > 50 AND "Mortality rate, under-5 (per 1,000 live births)" > 60 LIMIT 100
Which countries have high net migration but low GDP?	SELECT "name", "Net migration", "GDP (current US$)", "geometry" FROM "SDG" WHERE "Net migration" > 100000 AND "GDP (current US$)" < 100000000000 LIMIT 100

9 of 9

Enhancing user interface and feedback mechanisms.
Comparing performance with open-source models to evaluate efficiency, accuracy, and scalability.
Integrating dialogue and query generation to create a seamless conversational data exploration experience.
Enhancing visualization capabilities with interactive charts, time-series analysis. (I just used 2020’s data in this project)

Future things can do

Future Work – Presentation Speech

�

There are several ways to improve this system in the future.

�

First, we can enhance the user interface and feedback mechanisms to make interactions smoother and more user-friendly.

�

Second, we can compare the system’s performance with open-source models to evaluate its efficiency, accuracy, and scalability. This will help us understand how well our approach performs against existing solutions.

�

Third, we plan to integrate dialogue and query generation to create a seamless conversational experience. This will allow users to ask follow-up questions and refine their queries more naturally.

�

Finally, we aim to improve visualization capabilities by adding interactive charts and time-series analysis. Right now, we only use 2020’s data, but in the future, we can include data from multiple years to analyze trends over time.

�

These improvements will make the system more powerful, interactive, and insightful for SDG data exploration.