🤖 machine learning

AI4Land: Scalable Deep Learning for Global High-Resolution Land Use Reconstruction

Original authors: Amirpasha Mozaffari, Marina Castaño, Stefano Materia, Etienne Tourigny, Oscar Molina-Sedano, Jordi Varela-Agrelo, Dario Garcia-Gasulla, Miguel Castrillo Melguizo, Mario Acosta, Amanda Duarte

Published 2026-06-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amirpasha Mozaffari, Marina Castaño, Stefano Materia, Etienne Tourigny, Oscar Molina-Sedano, Jordi Varela-Agrelo, Dario Garcia-Gasulla, Miguel Castrillo Melguizo, Mario Acosta, Amanda Duarte

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Earth as a giant, complex puzzle. Scientists trying to predict future climate changes need to know exactly what pieces are on the board: where the forests are, where the cities are, and where the farms lie. However, they are currently missing a lot of the puzzle pieces, especially for the past (before satellites existed) and the future (what hasn't happened yet).

The paper introduces AI4Land, a new tool designed to fill in these missing pieces with incredible detail. Here is how it works, broken down into simple concepts:

1. The Problem: The "Blurry Map" vs. The "Missing History"

Scientists have two main types of maps, but both have flaws:

The "Blurry Map" (LUH2): This covers a huge timeline (from 1850 to 2100), but it's very low resolution. It's like looking at a photo of a forest from a plane; you can see a green blob, but you can't tell if it's a pine tree or an oak tree, or exactly where the edge of the forest meets a city.
The "High-Def Snapshot" (HILDA+): This is a super-clear, detailed map taken by satellites, but it only exists for a short time (the "satellite era"). It's like having a 4K photo of a city today, but no photos of what that city looked like 100 years ago or what it might look like in 50 years.

The Goal: AI4Land wants to take the "Blurry Map" and use it to create a "High-Def Movie" that covers the past, present, and future, all at a resolution of 1 kilometer (about 0.6 miles).

2. The Solution: The "Smart Architect" (The AI Model)

The researchers built a digital architect called AI4Land. Think of it as a master painter who has been trained to look at a blurry sketch and instantly know how to paint the fine details.

The Inputs: The AI looks at the "blurry" land-use data (the big picture) and mixes it with "static" features that never change, like the shape of the mountains, the slope of the hills, and the type of soil. It also looks at the land use from the year before or after to understand how things change over time.
The Training: The AI was trained on a massive supercomputer called MareNostrum5 (located in Barcelona). Imagine a classroom where the AI is shown millions of examples of "blurry maps" paired with their corresponding "high-definition" satellite photos. Over time, it learned the rules: "Oh, if the blurry map shows a mix of crops and the soil is sandy, it's probably a farm. If the blurry map shows a city block and the soil is clay, it's probably a city."
The Architecture: The specific type of AI used is called a U-Net. You can think of this as a funnel that takes in a wide, blurry image, squeezes it down to understand the core patterns, and then expands it back out to draw a sharp, detailed picture.

3. The Powerhouse: The Supercomputer

Creating these maps for the entire Earth is a massive job. To do it, the team used MareNostrum5, a supercomputer equipped with powerful graphics cards (GPUs).

Scaling Up: They tested how fast the computer could work. They started with one computer node and scaled up to eight. The results were like a well-oiled machine: adding more computers made the work finish almost perfectly faster, with almost no time wasted on the computers talking to each other. This proved that the system is "scalable"—it can handle huge tasks without breaking a sweat.

4. The Results: A Complete Timeline

The AI successfully generated a continuous, high-resolution map of land use for the years 1850 to 2100.

The Past: It filled in the gaps for the years before satellites existed (1850–1899) by using the blurry data and the AI's "painting" skills.
The Future: It created three different versions of the future (up to 2100) based on different climate scenarios (like different storylines for how humanity might develop).
Accuracy: The AI got about 94.7% of the land types right. It was very good at identifying forests, water, and grasslands. However, it struggled a bit more with cities (urban areas). This is like a painter who is great at painting trees but finds it hard to paint the tiny, complex details of a city skyline because there were fewer examples of cities in the training data.

5. What's Next?

The paper describes this as Phase 1. The team has already built the foundation.

Phase 2 (Planned): They plan to use these detailed land maps to predict how plants grow (specifically something called "Leaf Area Index," which is a measure of how much green leaf is on a tree).
Improvements: They plan to teach the AI more about cities by adding data about population density and climate zones, hoping to fix the "city painting" problem.
Open Source: All the code, data, and models are being released to the public so other scientists can use them to build better climate models.

Summary

In short, AI4Land is a digital time machine for geography. It uses a supercomputer and a smart AI to turn low-quality, long-term maps into high-definition, detailed movies of the Earth's surface, helping scientists understand how our planet's land has changed and how it might change in the future.

Technical Summary: AI4Land – Scalable Deep Learning for Global High-Resolution Land Use Reconstruction

1. Problem Statement

Uncertainty in the terrestrial carbon cycle remains a critical constraint in climate projections, largely driven by the coarse resolution and lack of dynamic representation of land surface processes in Earth system models. Existing global land-use (LU) and land-cover (LC) datasets face a fundamental trade-off: high-resolution datasets (e.g., HILDA+) are limited to the satellite era, while long-term historical and future projections (e.g., LUH2) exist only at coarse resolutions (~28 km). This resolution gap introduces substantial biases in simulated carbon sequestration, soil moisture, and surface energy fluxes. Current machine learning downscaling methods have been restricted to specific regions or single countries, lacking a globally scalable framework that bridges the gap between coarse scenario data and high-resolution spatial patterns across centennial timescales (1850–2100).

2. Methodology

The authors present AI4Land, a data-driven framework designed to generate high-resolution (1 km) historical reconstructions and future projections of land use and land cover. The framework operates in two phases; this work focuses on the first phase: reconstructing annual LU/LC maps.

2.1 Data and Preprocessing

Inputs: The model integrates coarse-resolution dynamic scenario data (LUH2, 0.25° grid) with static geophysical features. Inputs include 12 fractional LU variables, topographic features (elevation, slope, aspect), high-resolution soil characteristics (clay, sand, organic content), and an autoregressive prior from adjacent time steps ( $t-1$ or $t+1$ ).
Target: High-resolution ground truth labels derived from the HILDA+ dataset. The original 13-class scheme was consolidated into 8 classes (Forest, Cropland, Pasture, Urban, Grass/shrubland, Other land, Water, and a consolidated category) to mitigate class imbalance.
Preprocessing: A standardized pipeline using Climate Data Operators (CDO) remaps all inputs to a 1 km WGS84 grid. Soil data undergoes a non-linear depth-weighted average to prioritize root-zone interactions. Data is stored in ARCO Zarr format for efficient parallel I/O.

2.2 Model Architecture

The framework employs a U-Net architecture adapted for semantic segmentation:

Structure: A standard U-Net with 35 base channels, featuring an encoder with four max-pooling and double-convolution blocks and a symmetric decoder with upsampling and skip connections.
Input Handling: Accepts a multichannel stack of 14 LUH2 variables (across 2 time steps), 6 static features, and 1 HILDA prior.
Training Strategy: To prevent the model from trivially copying the high-resolution prior, 60% of the autoregressive prior is randomly masked during training. This forces the model to learn the mapping from coarse LUH2 inputs and static features.
Separation: Two distinct models are trained: one for historical data (using LUH2h) and one for future projections (using LUH2f).

2.3 Scalable Training and Inference

Infrastructure: Training was conducted on MareNostrum5 (Barcelona Supercomputing Center) using Distributed Data Parallelism (DDP) via the Hugging Face Accelerate library. The system scaled from 1 node (4 NVIDIA H100 GPUs) to 8 nodes (32 H100 GPUs).
Scalability: The pipeline achieved near-linear scaling, retaining 97–98% of ideal throughput at 8 nodes.
Inference: Global maps are generated using a sliding window approach over overlapping 512×512 patches. Predictions are merged using a Gaussian weighting function to ensure seamless transitions and avoid blocky artifacts.

2.4 Evaluation Strategy

To eliminate spatial autocorrelation leakage, the authors employed Grid-Based Partitioning combined with Farthest Point Sampling (FPS). The spatial domain was divided into 30×30 bins, with entire cells assigned exclusively to training, validation, or testing sets. A temporal split was also applied (1960–2000 for training, 2001–2015 for testing).

3. Key Results

Performance Metrics: The model achieved a mean Intersection over Union (mIoU) of 0.805 and an overall classification accuracy of 94.67% on the test set.
Class-Specific Performance: Performance varied significantly by class. Major land covers like Forest (mIoU 0.889), Water (mIoU 0.991), and Cropland (mIoU 0.815) were predicted with high fidelity. However, minority classes, particularly Urban areas, showed lower performance (mIoU 0.463), highlighting sensitivity to class imbalance.
Convergence: Multi-node training (8 nodes) demonstrated superior convergence, achieving a training loss of ~0.29 after 10 epochs compared to ~0.40 for the single-node baseline, attributed to larger effective batch sizes providing better gradient estimates.
Global Output: The pipeline successfully generated a comprehensive dataset spanning 1850–2100, covering the pre-observation era, the observational record, and three future climate scenarios (SSP2-4.5, SSP3-7.0, SSP4-6.0). Validation for the year 2014 showed a global accuracy of 94.88% and an mIoU of 0.8569.

4. Significance and Contributions

The paper claims the following significance:

First Global Scalable Framework: AI4Land is presented as the first model to bridge the resolution gap for land use data at a global scale, spanning 1850 to 2100, rather than being limited to specific regions or the satellite era.
HPC-Enabled AI: The work demonstrates how GPU-accelerated High-Performance Computing (HPC) infrastructure (specifically EuroHPC/MareNostrum5) enables the training of complex deep learning pipelines for global climate applications, achieving near-linear scalability.
Digital Twin Readiness: The resulting open-source emulators are designed for real-time coupling with digital twin platforms (e.g., Destination Earth), providing realistic, evolving land surface conditions on demand.
Uncertainty Reduction: By providing temporally continuous and spatially detailed land boundary conditions, the framework aims to reduce critical uncertainties in the terrestrial carbon cycle and improve the predictive power of next-generation climate simulations regarding carbon, water, and energy fluxes.

5. Future Work and Limitations

The authors acknowledge that the model's performance is inherently impacted by the quality and biases of the source data (HILDA+ and LUH2). They note that urban and minority class classification requires improvement. Planned future work includes:

Incorporating additional predictors such as Köppen-Geiger climatic zones and population density data to improve urban classification.
Integrating a Recurrent U-Net to better enforce temporal consistency across centennial projections and address exposure bias via scheduled sampling.
Investigating generative approaches (Flow Matching U-Nets with Vision Transformer backbones) to capture long-range spatial dependencies and quantify uncertainty.
Exploring the use of foundation models via adapters and fine-tuning.
Extending the framework to a second phase to predict dynamic biophysical variables, specifically Leaf Area Index (LAI).

All datasets, models, code, and pretrained weights are committed to being released as open-source to support reproducibility and innovation in Earth system modeling.