AI4Land: Scalable Deep Learning for Global High-Resolution Land Use Reconstruction
Amirpasha Mozaffari, Marina Castaño, Stefano Materia, Etienne Tourigny, Oscar Molina-Sedano, Jordi Varela-Agrelo, Dario Garcia-Gasulla, Miguel Castrillo Melguizo
Abstract
Uncertainty in the terrestrial carbon cycle remains a major constraint in climate projections, partly driven by the uncertainties affecting the land surface representation and variability in Earth system models. To address this limitation, we present a data-driven framework AI4Land, for generating high-resolution historical reconstructions and future projections of key land surface variables. The framework follows a two-phase approach using a U-Net architecture. In the first phase, which is the focus of this work, it reconstructs annual land use and land cover by integrating coarse-resolution scenario data with static geophysical features. In a planned second phase, the resulting high-resolution maps will be used to predict dynamic biophysical variables, particularly leaf area index, at finer temporal scales. Trained on Earth observation data, the models learn to reproduce spatially explicit and physically consistent land surface patterns, extending temporal coverage to periods lacking direct observations. AI4Land was developed and trained on MareNostrum5, demonstrating how GPU-accelerated HPC infrastructure enables global-scale climate AI pipelines. The final product is a suite of open-source emulators designed for real-time coupling with digital twin platforms, such as those developed under the Destination Earth initiative. By delivering realistic and evolving land surface conditions on demand, this work aims to reduce critical uncertainties and improve the predictive power of next-generation climate simulations.
AI Impact Assessments
(1 models)Scientific Impact Assessment: AI4Land
1. Core Contribution
AI4Land presents a deep learning framework for downscaling coarse-resolution land use/land cover (LU/LC) data (~28 km, from LUH2) to high resolution (~1 km) by learning the statistical mapping to HILDA+ satellite-era observations, then extrapolating to periods without ground truth (1850–1899 historical, and 2020–2100 future projections under SSP scenarios). The approach uses a standard U-Net for semantic segmentation, fusing coarse dynamic LU forcing with static geophysical features (topography, soil) and a partially masked autoregressive prior from adjacent years. The framework was trained on MareNostrum5 using distributed data parallelism.
The core novelty is not in the model architecture (a vanilla U-Net) but rather in the systems-level integration: assembling heterogeneous Earth observation datasets, preprocessing them into analysis-ready format, and deploying the pipeline at global scale on HPC infrastructure to produce a specific dataset product spanning 1850–2100 at 1 km resolution. The authors explicitly acknowledge this, framing the contribution as a demonstration that such workflows can be operationally deployed.
2. Methodological Rigor
Strengths in evaluation design: The spatial evaluation strategy using grid-based partitioning combined with Farthest Point Sampling and temporal splitting (train: 1960–2000, test: 2001–2015) is well-designed to prevent spatial autocorrelation leakage—a common pitfall in geospatial ML.
Concerns:
3. Potential Impact
The intended use case—providing dynamic, high-resolution land surface boundaries for Earth system models and digital twin platforms like Destination Earth—is genuinely important. If the dataset quality is sufficient, it could fill a real gap in climate modeling workflows. The commitment to open-source release of data, models, and code adds value.
However, the actual scientific impact depends heavily on the second phase (LAI and dynamic biophysical variables), which is not yet developed. The LU/LC maps alone, while useful, are an intermediate product. The paper's impact is currently limited by being largely a dataset generation exercise with a standard architecture.
The coupling with digital twin platforms is mentioned but not demonstrated. Without showing downstream impact on climate model simulations (e.g., does using AI4Land LU maps actually reduce carbon cycle uncertainty compared to using raw LUH2?), the claimed benefits remain aspirational.
4. Timeliness & Relevance
The paper addresses a real and recognized gap. The mismatch between the temporal coverage of LUH2 (850–2100, coarse) and HILDA+ (1960–2019, fine) is well-documented, and bridging it is valuable for CMIP-class experiments. The connection to Destination Earth and EuroHPC infrastructure is timely. However, related work by Chen et al. (2020) already provides 0.05° global projections for 2015–2100, and other regional efforts exist. The specific value-add of 1 km resolution for the full 1850–2100 period is real but incremental.
5. Strengths & Limitations
Key Strengths:
Notable Limitations:
Overall Assessment
AI4Land is a competent engineering effort that assembles known components (U-Net, DDP training, standard Earth observation datasets) into a useful pipeline. Its primary value is as a dataset contribution and an operational demonstration rather than a methodological advance. The scientific impact is currently modest—the framework delivers reasonable but not state-of-the-art segmentation performance and lacks the downstream validation needed to confirm its claimed benefits for climate modeling. The most impactful aspects (phase 2, digital twin coupling, uncertainty quantification) are deferred to future work.
Generated Jun 11, 2026
Comparison History (21)
Paper 2 likely has higher scientific impact due to its direct, large-scale real-world application to climate modeling and carbon-cycle uncertainty, a highly timely and societally critical area. AI4Land’s outputs (global high-resolution reconstructions, open-source emulators, HPC-enabled pipeline, planned coupling to digital twins) can be broadly used by Earth system science, remote sensing, ecology, and policy communities, increasing breadth of impact. Paper 1 is methodologically innovative for discrete diffusion fine-tuning, but its impact is more specialized within ML sequence generation and may face faster turnover in the field.
AI4Land addresses a critical gap in climate science—uncertainty in terrestrial carbon cycle projections—with a scalable, practical framework for high-resolution land use reconstruction. Its integration with Earth system models, digital twin platforms (Destination Earth), and open-source emulators gives it broad real-world applicability across climate science, environmental policy, and remote sensing. Paper 1, while presenting an interesting RL formulation for neural model editing, is more exploratory and incremental, demonstrating modest improvements on relatively narrow tasks (bias mitigation and unlearning) without fundamentally advancing either RL or model editing.
Paper 1 targets a major, timely bottleneck in climate science—land-surface uncertainty affecting carbon-cycle and Earth system projections—using scalable deep learning aimed at integration with digital twins (Destination Earth). Its outputs (global high-resolution reconstructions/projections and open-source emulators for real-time coupling) have broad cross-field impact (climate modeling, remote sensing, ecology, HPC/ML) and strong real-world policy relevance. Paper 2 is methodologically rigorous and valuable for standardization in WHAR, but its impact is narrower and more incremental (benchmarking/efficiency tradeoffs) compared to the potentially transformative implications for climate prediction.
Paper 1 addresses a critical, global-scale scientific challenge—reducing uncertainties in climate projections and the terrestrial carbon cycle. Its integration of deep learning with Earth system modeling and HPC digital twins offers profound, cross-disciplinary scientific impact. In contrast, Paper 2 presents a highly valuable but narrowly focused engineering optimization for LLM inference systems, which has strong technical utility but less fundamental scientific breadth.
Paper 2 introduces a fundamental methodological advancement in generative modeling by embedding strict polyhedral constraints into flow matching without projection steps. This provides a safe, mathematically rigorous, and efficient solution applicable to a broad range of safety-critical physical systems, robotics, and control domains. In contrast, while Paper 1 addresses the highly important issue of climate modeling, it primarily applies existing U-Net architectures to Earth observation data. Paper 2's core algorithmic innovation offers broader cross-disciplinary impact and sets a new standard for safe, constrained generative models.
Paper 2 (APPO) is more methodologically novel and broadly applicable: it introduces fine-grained branching and credit assignment for LLM agents, a timely, fast-moving area with immediate adoption potential across many agent/tool-use settings. It reports systematic evaluation on 13 benchmarks with consistent gains over strong baselines, suggesting solid rigor and generality. Paper 1 targets an important climate application and offers engineering value (HPC, open-source emulators), but its core method (U-Net super-resolution/fusion) is less conceptually new and impact may be narrower to Earth-system modeling workflows.
Paper 2 likely has higher impact due to strong real-world relevance (reducing land-surface uncertainty in climate projections), clear pathway to operational deployment (open-source emulators, digital twin coupling, HPC scalability), and broad cross-field reach (climate science, remote sensing, Earth system modeling, AI/HPC). Its timeliness aligns with major initiatives (Destination Earth) and could influence modeling workflows widely. Paper 1 is methodologically novel for operator learning and interpretability, but its demonstrated scope is narrower (few PDE tasks) and near-term applications are less immediate, limiting expected impact breadth.
Paper 2 has higher likely scientific impact due to its direct relevance to climate modeling and terrestrial carbon-cycle uncertainty, a high-priority global challenge with immediate real-world applications. It proposes a scalable, HPC-enabled pipeline for global high-resolution land-use reconstruction and open-source emulators intended for coupling with digital-twin Earth system platforms, enabling broad downstream use across climate science, ecology, remote sensing, and policy. Paper 1 is methodologically novel and rigorous for causal interventions in CBMs, but its impact is more specialized to interpretable ML rather than a large cross-domain, societally urgent application area.
Paper 1 offers a scalable, open-source deep learning framework for high-resolution land-use reconstruction with direct relevance to reducing uncertainties in Earth system models and enabling digital-twin climate simulations, yielding substantial real-world and cross-disciplinary impact (climate science, remote sensing, HPC, AI). It is timely given Destination Earth–style initiatives and proposes a clear pipeline with extensibility to biophysical variables. Paper 2 is valuable methodological critique improving evaluation rigor in masked diffusion LMs, but its contributions are narrower, mainly incremental/diagnostic, and likely to affect a smaller application domain than climate modeling infrastructure.
Paper 1 addresses a critical bottleneck in global climate modeling—terrestrial carbon cycle uncertainties—with a highly scalable, open-source AI framework. Its global scope, utilization of HPC, and potential integration with Earth digital twins give it a broader scientific and societal impact compared to Paper 2's localized, albeit highly efficient, flood prediction method.