[Paper Review] Challenges in data-based geospatial modeling for environmental research and practice

Diana Koldasbayeva, Polina Tregubova+4 more|arXiv (Cornell University)|Nov 18, 2023

Atmospheric and Environmental Gas Dynamics4 citations

TL;DR

This paper identifies and addresses critical challenges in data-based geospatial modeling for environmental research, including data imbalance, spatial autocorrelation, uncertainty estimation, and model generalization. It proposes a comprehensive review of techniques and tools—such as semi-supervised learning, multimodal data integration, and self-supervised deep learning—to improve model accuracy, reproducibility, and real-world deployment in climate and ecosystem monitoring.

ABSTRACT

With the rise of electronic data, particularly Earth observation data, data-based geospatial modelling using machine learning (ML) has gained popularity in environmental research. Accurate geospatial predictions are vital for domain research based on ecosystem monitoring and quality assessment and for policy-making and action planning, considering effective management of natural resources. The accuracy and computation speed of ML has generally proved efficient. However, many questions have yet to be addressed to obtain precise and reproducible results suitable for further use in both research and practice. A better understanding of the ML concepts applicable to geospatial problems enhances the development of data science tools providing transparent information crucial for making decisions on global challenges such as biosphere degradation and climate change. This survey reviews common nuances in geospatial modelling, such as imbalanced data, spatial autocorrelation, prediction errors, model generalisation, domain specificity, and uncertainty estimation. We provide an overview of techniques and popular programming tools to overcome or account for the challenges. We also discuss prospects for geospatial Artificial Intelligence in environmental applications.

Motivation & Objective

• To identify and systematize persistent challenges in data-based geospatial modeling, including data imbalance, spatial autocorrelation, and uncertainty estimation.
• To evaluate existing techniques and programming tools that mitigate these challenges in environmental applications.
• To examine the role of new data sources—such as Earth observation, climate data, and social media—in enhancing model performance and robustness.
• To explore the prospects of self-supervised and multimodal deep learning models for scalable, transparent, and reliable geospatial predictions.
• To emphasize the importance of model deployment, monitoring, and maintenance in real-world environmental decision-making systems.

Proposed method

• Systematic review of 100+ studies and tools in geospatial machine learning, focusing on data quality, model interpretability, and uncertainty quantification.
• Analysis of techniques such as data augmentation, transfer learning, and ensemble methods to address class imbalance and improve generalization.
• Evaluation of spatial autocorrelation correction methods, including spatial filtering and covariance-based modeling in geostatistical frameworks.
• Survey of emerging deep learning architectures—such as Transformers and vision transformers—applied to satellite and radar data for high-resolution mapping.
• Investigation of semi-supervised and self-supervised learning pipelines using large-scale, weakly labeled datasets (e.g., JFT-3B, SEVIR, LVD-142M) to reduce labeling costs.
• Discussion of multimodal integration strategies combining satellite imagery, climate data, and social media to enhance model robustness and real-time responsiveness.

Figure 1: Examples of geospatial mapping performed for different tasks of environmental monitoring and assessment a) maps of forest disturbance regimes of Europe [ 19 ] ; b) land cover and mapping of losses for different types of forest in Indonesia [ 20 ] ; c) maps of soil organic carbon (SOC) frac

Experimental results

Research questions

• RQ1 How do data imbalance and spatial autocorrelation affect the reliability and generalization of geospatial machine learning models in environmental applications?
• RQ2 What techniques are most effective for uncertainty estimation and model interpretability in geospatial predictions?
• RQ3 How can semi-supervised and self-supervised learning approaches improve model performance with limited labeled geospatial data?
• RQ4 What role do multimodal data integration and advanced deep learning architectures play in enhancing geospatial model robustness?
• RQ5 What are the key challenges and best practices for deploying and maintaining geospatial models in operational environmental monitoring systems?

Key findings

• Data quality, quantity, and diversity are critical for reliable geospatial modeling, with large-scale, curated datasets like JFT-3B and SEVIR significantly improving model performance.
• Semi-supervised and self-supervised learning methods show strong potential for reducing labeling costs while maintaining high accuracy in geospatial applications.
• Multimodal integration of satellite, climate, and social media data enhances model robustness and enables real-time environmental monitoring and response.
• Spatial autocorrelation and data imbalance remain major challenges, but techniques like spatial filtering and class reweighting can significantly improve model generalization.
• Model deployment and maintenance are often underestimated; continuous monitoring for concept drift and data distribution shifts is essential for long-term operational reliability.
• Self-supervised pretraining on large-scale geospatial data is emerging as a key enabler for next-generation environmental AI, mirroring success in NLP and computer vision.

Figure 2: General workflow for the tasks including geospatial modelling process and common issues relevant for each stage.