Driven geospatial artificial intelligence modeling for climate change impact assessment: a global perspective

Technical Summary: Driven Geospatial Artificial Intelligence Modeling for Climate Change Impact Assessment

Problem Statement
The paper addresses the critical challenge of assessing anthropogenic climate change impacts, a complex, polycausal phenomenon requiring the processing of vast, heterogeneous, and multi-modal Earth observation datasets. Traditional mechanistic climate models face computational and structural limitations in assimilating petabyte-scale data from satellite remote sensing, UAVs, and in-situ sensors. While Geospatial Artificial Intelligence (GeoAI)—the synthesis of spatial science, AI, and statistical learning—offers a transformative paradigm for extracting non-obvious patterns and enhancing prediction granularity, the field lacks a comprehensive, critical synthesis. Existing literature reviews often focus narrowly on specific technologies (e.g., remote sensing) or specific domains (e.g., hydrology) without capturing the methodological convergence of GeoAI and climate change impact assessment, nor do they sufficiently map the field's intellectual structure or persistent epistemological challenges.

Methodology
The study employs a convergent parallel mixed-methods approach, integrating quantitative scientometric analysis with a qualitative systematic review.

• Data Retrieval: Following PRISMA guidelines, the authors conducted a multi-database search (Scopus, Web of Science, IEEE Xplore, Google Scholar) for publications between January 2016 and June 2026.
• Corpus Selection: From an initial 892 records, 485 unique publications were identified. After screening for peer-reviewed status, English language, and substantive AI-geospatial integration in climate contexts, a core corpus of 152 articles was selected for scientometric analysis.
• Systematic Review Expansion: A forward and backward snowballing technique was applied to the 152 articles, identifying an additional 4 highly relevant studies, resulting in a final expanded corpus of 77 articles for in-depth qualitative content analysis.
• Analytical Tools:
  • Quantitative: Bibliometric data was analyzed using the R package Bibliometrix and Python libraries (NetworkX, scikit-learn) to map publication trends, keyword co-occurrence networks, and collaboration structures.
  • Qualitative: A structured, iterative coding framework (using NVivo and Python) categorized the 77 articles across five domains: GeoAI architectures, integration mechanisms, technological attributes, application domains, and challenges/ethics.

Key Contributions and Results

1. Scientometric Landscape:

   • Growth Trajectory: The field has experienced exponential growth (R² > 0.95), transitioning from a nascent phase (2–4 publications/year, 2016–2018) to a mature, rapidly consolidating domain, peaking at 22 publications in 2025 with projections exceeding 25 for 2026.
   • Thematic Clusters: Keyword co-occurrence analysis revealed four distinct clusters: (1) Methodological Core (Deep Learning, CNNs, LSTMs); (2) Data Sources (Remote Sensing, Sentinel, SAR); (3) Impact Domains (Sea level rise, drought, wildfire, biodiversity); and (4) Modeling Context (GIS, risk assessment).
   • Global Leadership: The United States and China are the dominant contributors in volume and collaboration networks. The UK, Germany, and Switzerland also play influential roles. Leading publication venues include Remote Sensing of Environment and Nature Climate Change.

2. Technological Synthesis:
   The review deconstructs the GeoAI technological stack applied to climate science:

   • Spatial Modeling: Convolutional Neural Networks (CNNs), particularly U-Net and DeepLab architectures, are foundational for semantic segmentation of land cover, glacier boundaries, and sea ice using multi-spectral and SAR data.
   • Temporal Dynamics: Recurrent Neural Networks (RNNs), specifically LSTMs and GRUs, are superior for forecasting climate variables (precipitation, temperature) by capturing long-term dependencies in time-series data.
   • Emerging Architectures: Transformer-based models (e.g., Vision Transformers) are being adapted for geospatial tasks to capture global context and long-range teleconnections (e.g., ENSO), while hybrid models (CNN-LSTM) combine spatial and temporal feature extraction.
   • Physics Integration: Physics-Informed Neural Networks (PINNs) are highlighted as a promising frontier, embedding physical laws (e.g., conservation equations) into loss functions to ensure physical plausibility, especially in data-sparse regions.

3. Application Domains:
   The review categorizes applications into monitoring/detection and forecasting/projection:

   • Cryosphere/Hydrosphere: Automated mapping of glacier retreat, sea ice concentration, and flood forecasting.
   • Ecosystems: Enhanced Species Distribution Models (SDMs) and monitoring of coral bleaching and algal blooms.
   • Agriculture: Crop yield prediction and early detection of pest infestations.
   • Urban/Health: Mapping urban heat islands, flood inundation, and vector-borne disease dynamics.

4. Critical Challenges:
   The study identifies six major impediments:

   • Data Issues: Heterogeneity, sensor noise, and a critical paucity of high-quality labeled data, particularly for rare events and remote regions.
   • Computational Intractability: High resource demands for training complex models on global datasets, creating barriers for low-resource institutions.
   • Generalizability: "Domain shift" issues where models trained on specific regions or sensors fail to transfer globally due to non-stationarity in climate systems.
   • The "Black Box" Problem: Lack of interpretability hinders trust among policymakers and domain scientists.
   • Ethical Concerns: Algorithmic biases in training data may exacerbate climate injustices and disproportionately affect marginalized communities.
   • Interoperability: Difficulties in integrating diverse data formats and reconciling AI outputs with physical models.

Significance and Future Directions
The paper posits that the evolution of GeoAI is crucial for constructing a resilient and equitable planetary future. It argues that the field must move beyond correlation-based analysis toward mechanistic, data-driven understanding. The authors outline a strategic research agenda to bridge current gaps:

• Explainability (XAI): Prioritizing the integration of post-hoc (e.g., SHAP, LIME) and intrinsic interpretability methods to build trust and uncover physical drivers.
• Causality: Shifting from predictive modeling to causal inference frameworks to answer "what-if" scenarios for adaptation planning.
• Foundation Models: Developing pre-trained, self-supervised models for Earth observation to improve transferability and democratize access.
• Physics-Guided AI: Advancing "theory-guided data science" to ensure models respect fundamental Earth system laws.
• Equity: Developing participatory frameworks that incorporate local knowledge and mitigate algorithmic biases to ensure climate justice.

The study concludes that addressing these challenges through interdisciplinary collaboration is essential to transform GeoAI from a powerful analytical tool into a trusted, actionable intelligence system for global climate resilience.