💡Capabilities & Datasets
⭐Get started
📋Overview
Global Nature Watch (GNW) Horizon offers a set of core capabilities designed to help you explore, analyze and make sense of geospatial data to help monitor nature.
With GNW Horizon, you can ask questions in plain language to:
Discover and learn about peer-reviewed data from Land & Carbon Lab and Global Nature Watch (formerly Global Forest Watch)
Analyze data for your areas of interest and get insights to help make more sense of the data
Get answers that are tailored to your context as a starting point for reports, policies or project decisions
Save and curate your findings - build dashboards to organise insights over time, return to earlier work, and share what you have found.
When you ask a question or provide a prompt (see Prompt guidance), GNW Horizon's AI assistant processes your intent using a system of agents (see Models) to understand your request and set the context - data layer, area, date - for the response.
As you explore, you can save both individual insights and full conversations on the platform. Your saved areas, layers and conversations stay organized so you can return to earlier work, build a portfolio of places and questions and pick up right where you left off.
Because GNW Horizon is in preview, its capabilities will continue to evolve. It is designed to be open, collaborative and transparent so anyone can build on it. Over time, you can expect new features to be added, existing ones to improve and some to be retired as we learn more about how best to help you work with nature data.
ℹ️How GNW Horizon Processes Your Request
GNW Horizon combines a conversational AI interface with structured geospatial data systems. While you interact with GNW Horizon in natural language, your request is processed through a deterministic, step-by-step analytical workflow to ensure scientific accuracy and reproducibility.
At a high level, GNW Horizon uses:
A large language model (LLM) to interpret your request and determine intent
Metadata-aware retrieval (Retrieval-Augmented Generation, or RAG) to understand available datasets and boundaries
Deterministic data services to fetch and aggregate verified geospatial data
Once a prompt is entered, GNW Horizon follows a sequential workflow.
How AI and Deterministic Systems Work Together
GNW Horizon combines two types of systems:
AI (Language Models)
Understand your question
Select datasets and parameters
Explain trends and suggest follow-up questions
Deterministic Data Services
Retrieve peer-reviewed data from reliable sources (e.g., Global Nature Watch and Land & Carbon Lab)
Perform geospatial calculations, drawing on the Data API and using CodeAct
Generate charts and visual outputs using CodeAct
Return the same results every time for the same inputs
AI helps interpret and explain. Deterministic systems perform the scientific analysis.
Step 1: Understanding Your Request
GNW Horizon's language model reads your question and identifies three key components:
Where — the geographic area you're asking about
What — the type of environmental data you need
When — the time period of interest
If any of these are missing or unclear, GNW Horizon will ask you to clarify before proceeding.
Step 2: Identify the Location (Where)
GNW Horizon first determines the specific geographic area you are asking about.
It matches your place name to a predefined Area of Interest (AOI) in its spatial database
If a place name is ambiguous (for example, multiple places share the same name), GNW Horizon will ask you to clarify
GNW Horizon analyzes one defined area at a time
You can also click directly on the map to select an area. When active datasets are loaded, a prompt will appear in the chat inviting you to run an analysis on that area immediately.
No analysis begins until a specific location is confirmed.
AI interprets your place name. A rule-based system matches it to a predefined boundary.
Step 3: Dataset & Time Selection (What and When)
Next, Horizon determines which dataset best answers your question.
It interprets keywords such as “fire,” “deforestation” or “land cover change”
It selects the appropriate data layer
It confirms or requests a time range (for example, “2018–2023” or “last three months”)
When your question could be answered by more than one dataset - for example, if you ask about fires and several relevant datasets are available - the assistant may present options with a short explanation of each, rather than selecting one automatically. This helps you choose the dataset that best fits your question before the analysis runs.
If the time period is missing or unclear, Horizon will ask for clarification before proceeding.
AI selects the relevant dataset and parameters based on your request.
In some cases, the assistant may perform Step 3 in isolation - selecting a dataset or navigating to a location - without proceeding to data retrieval or analysis. This happens when your request is specific and self-contained, such as asking the assistant to add a layer to the map or move to a particular area. When a full analysis is needed, the assistant continues through Steps 4 and 5.
Step 4: Data Retrieval (Deterministic Query)
Once the location, dataset and time range are confirmed, Horizon retrieves the relevant data from its underlying data sources.
The data returned reflects:
The selected geographic boundary
The specified time range
The chosen dataset
If any required information is missing, this step does not proceed.
This step is fully deterministic: data is retrieved directly from structured geospatial services.
Step 5: Analysis & Insight Generation
Horizon then analyzes the retrieved data to produce user-friendly outputs. This includes:
Creating a chart or simple visualization
Summarizing key trends in plain language
Suggesting follow-up questions for deeper exploration
Providing access to the underlying code used to produce visualizations
The goal is to make complex environmental data understandable and actionable without requiring technical expertise.
Deterministic systems provide underlying value for charts. AI: generates executable code when creating charts (rather than manipulating data directly) using CodeAct and explains and summarizes the results.
When your question is broad or contextual, the assistant may draw on WRI and LCL Insights articles to provide additional context. Where sources are used, they appear as inline citation chips in the response - hover over one for a preview card, or scan the full source list at the end of the message. Each citation links directly to the original article.
Status Updates & Transparency
Throughout the process, GNW Horizon may provide short updates (for example, confirming the selected location or dataset). These updates are intended to keep you informed and ensure accuracy before analysis continues.
⚙️Capabilities & Limitations
This section summarizes what GNW Horizon does and does not support across locations, datasets, analysis types and outputs.
📍Areas of Interest (AOI)
GNW Horizon can analyze data only for predefined geographic boundaries. The table below outlines which types of Areas of Interest (AOIs) are supported and which are not.
Countries (admin 0)
✅
Predefined boundaries only
States / Provinces (admin 1)
✅
Varies by country
Districts / Municipalities (admin 2)
✅
Where available
Protected Areas (WDPA)
✅
Limited to groups of <25 WDPA areas
Key Biodiversity Areas (KBA)
✅
Limited to groups <25 KBA areas
Indigenous & Community Lands
✅
Limited to groups of <25 areas
Custom Polygons
✅
Can be uploaded (GeoJSON, up to 1 MB) or drawn on the map
Global / Continental Areas
❌
e.g., Africa, Europe
Ecosystems or Regions
❌
e.g., Amazon Basin, Latin America
Buffers
❌
e.g., 10km around a defined area
🗺️Datasets
GNW Horizon draws on a trusted set of geospatial datasets from world-leading researchers to provide timely and actionable insights on land use, forest change and nature-based solutions. These datasets include deforestation alerts, historical tree cover loss, land cover maps and more.
We are committed to transparency in how data is sourced and used. Most of the data powering GNW Horizon originates from or is also available on Land & Carbon Lab and Global Nature Watch (formerly Global Forest Watch), where you can explore the full range of methodologies, data providers and technical documentation. All data made accessible through GNW Horizon has been peer reviewed.
The tables below summarizes the data themes currently supported and their general characteristics.
Ecosystem Monitoring
2000 baseline, 30m resolution, global | Source: UMD
data can be filtered by % canopy cover, Forest Carbon Stocks
2000-2020, 5-year intervals, 30m resolution, global | Source: UMD
data can be filtered by % canopy cover, Forest Carbon Removals
2001-2025, annual, 30m resolution, 10% - 75% canopy density in 5% increments, global | Source: UMD
data can be filtered by % canopy cover, Primary Forest, Forest Greenhouse Gas Emissions, Intact Forest Landscapes, and Tree Cover Loss due to Fires data
2001-2025, ~1km resolution | Source: WRI/Google DeepMind
data can be filtered by % canopy cover, Forest Greenhouse Gas Emissions
2001-2025, annual, 30m, global |Source: Hansen/UMD/GLAD
data can be filtered by % canopy cover, Primary Forest, Forest Greenhouse Gas Emissions, and Intact Forest Landscapes data
Land Cover Change Monitoring
2023-present, updated weekly, 30m resolution, global | Source: UMD
alerts can be filtered by any one of the following datasets at a time: drivers (LDACS), SNTB Natural Lands Map, Global Natural/Semi-Natural Grassland Extent (2022), Global Land Cover (2024)
2015-2024, annual, 30m resolution | Source: UMD
analysis of land cover change is only available for 2015 to 2024 (not annually), and land cover analysis is only available for 2024
2024-present, updated quarterly, 30m resolution | Source: WRI/UMD
GHG Monitoring
average over 2001-2024, 30m resolution | Source: WRI
can be analyzed but isn't available to view on the map as a separate data layer
2001-2024, annual, 30m resolution | Source: WRI
can be analyzed but isn't available to view on the map as a separate data layer
Agricultural Monitoring
2020–2024, tabular data at national/subnational levels Source: LCL
Covers 42 agricultural crops. Only available for administrative boundaries (not KBAs, protected areas, etc.) *Analysis only
2000-2024, annual, 30m resolution | Source: LCL/OpenGeoHub/LAPIG/IIASA/iDiv
data only available up to 2022
📈Analyses
GNW Horizon performs structured, descriptive analyses based on the selected location, dataset and time range. The table below summarizes the types of analyses supported.
Area-based summaries
Straightforward query targeting one area or dataset
“How much area was affected by conversion alerts in Peru?”
✅
Temporal analysis
Refers to specific years or date ranges
“How much forest was lost in Brazil between 2001 and 2020?”
✅
Trend analysis over time
Analyzes trends based on data analysis over time
“Is deforestation going up or down in Brazil?”
✅
Within selected AOI
Before / after comparison
Analyzes data before and after a specified time
“Compare deforestation in Kalimantan before and after the start of Nusantara in 2022.”
✅
Requires defined time range
Cross-location comparison
Compares two or more named areas
“Which had more cropland in 2015, Nigeria or Ghana?”
✅
Nested-area comparison
Compares sub-areas within a larger one
“Which state in India lost the most forest cover?”
✅
Dataset class comparison
Compares classes of data in a given location or analyzes a specific class in the data
“What percentage of disturbances in Occitanie were due to crop management in 2023?”
✅
Ranking
Ranks subareas within a larger area
“What states in Brazil had the most tree cover loss in 2024?”
✅
Slow, and limited to 1,000 areas
Cross-country analysis
Ranks countries globally
“Where is agricultural land found in the greatest area globally?”
✅
Up to 1,000 areas
Predictive modeling
❌
Not supported
Forecasting
❌
Not supported
📊Outputs
GNW generates standardized visual and narrative outputs designed to make data accessible and interpretable. The table below outlines the supported output types.
Time series charts
✅
Primary visualization
Bar / comparison charts
✅
Simple comparisons
Written summaries
✅
Plain-language insights
Raw data downloads
✅
“View how this was generated” gives access to raw, chart data and the code used to generate the visualization
Custom maps and data visualizations
❌
Not supported
Dashboards
✅
Save insights from your conversations to a persistent dashboard. Add charts, maps, satellite imagery, text blocks, and curated insights. Accessible at /dashboards.
Satellite imagery
✅
Ask the assistant to retrieve a recent, cloud-free Sentinel-2 view of your area. Imagery appears on the map or can be added to a dashboard.
An example dashboard combining AI-assisted analysis, satellite imagery, and WRI research citations.
Last updated