Eyes on the Earth: Documenting Resource Extraction with AI

by

Author: Denis Avetisyan

A new system uses artificial intelligence to analyze satellite imagery of mining sites worldwide, creating a lasting record of environmental impact.

This review details a multimodal AI system leveraging retrieval-augmented generation and Earth observation to interpret resource extraction activities globally.

The increasing scale of resource extraction demands new methods for comprehensive environmental monitoring, yet traditional analysis often lags behind the pace of change. This paper, ‘Synthetic Reflections on Resource Extraction’, details a novel pipeline leveraging multimodal large language models and retrieval-augmented generation to interpret satellite imagery of global mining sites. The resulting system produces succinct commentaries documenting the impact of industrial activity, effectively creating a planetary-scale record of human-environment interaction. Could this approach not only illuminate past practices, but also inform more sustainable resource management strategies for the future?

The Planetary Footprint: Extractivism and Earth’s Transformation

The pursuit of resources through extractivism – encompassing mining, forestry, and fossil fuel recovery – has fundamentally shaped Earth’s surface, leaving alterations visible from local ecosystems to the planetary scale. While essential for societal development, these processes inherently disrupt landscapes, initiating cycles of deforestation, soil erosion, and habitat fragmentation. Beyond the immediate footprint of extraction, extensive infrastructure – roads, pipelines, and processing facilities – further amplifies environmental consequences, often persisting long after resource depletion. This legacy isn’t merely topographical; chemical alterations from processing, such as acid mine drainage or oil spills, contaminate vital resources and pose long-term risks to biodiversity and human health, demonstrating that the demand for materials inevitably results in profound and lasting modifications to the planet’s natural systems.

A comprehensive understanding of the environmental consequences stemming from resource extraction is paramount for forging genuinely sustainable practices and effective resource management strategies. The impacts of extractivism – ranging from deforestation and habitat loss to water contamination and soil degradation – are not isolated incidents but rather interconnected alterations to planetary systems. Detailed assessment allows for the identification of critical thresholds beyond which ecosystem function is irrevocably compromised, enabling proactive mitigation efforts and responsible planning. Furthermore, quantifying the scale of these impacts is essential for developing accurate environmental accounting systems, informing policy decisions, and ultimately, transitioning towards a circular economy that minimizes future ecological disruption and ensures long-term resource availability for generations to come.

Assessing the environmental consequences of resource extraction presents a significant challenge, as conventional monitoring techniques often fall short when applied to the sheer scale of modern operations. Current efforts, while tracking over 200 mining sites globally, struggle to deliver the detailed, real-time data needed for effective environmental management. These limitations stem from the logistical difficulties of consistently surveying expansive and often remote landscapes, coupled with the time-consuming nature of traditional data collection and analysis. Consequently, impacts like deforestation, water contamination, and habitat loss may go undetected or be assessed only after substantial damage has occurred, hindering proactive mitigation strategies and sustainable resource governance.

Seeing the Unseen: Foundations of Remote Sensing

The Sentinel-2 mission, operated by the European Space Agency, delivers a continuous stream of high-resolution multispectral imagery covering the Earth’s land surface. This data is acquired by two polar-orbiting satellites, providing revisit times of up to 5 days. The mission’s core capabilities stem from its 13 spectral bands, ranging from visible and near-infrared to shortwave infrared wavelengths, with spatial resolutions varying from 10 meters for visible and near-infrared bands to 60 meters for coarser resolution bands like thermal infrared. This combination of spectral detail and spatial resolution enables detailed monitoring of vegetation, land cover change, water quality, and various other environmental parameters. Data is freely available, supporting a wide range of applications in environmental monitoring, agriculture, and disaster management.

Vegetation indices, such as the Normalized Difference Vegetation Index (NDVI), are calculated from the visible and near-infrared portions of the electromagnetic spectrum to quantify vegetation greenness and biomass; stressed vegetation typically exhibits lower NDVI values due to reduced chlorophyll content. The Ferrous Metal Index specifically targets the spectral characteristics of iron oxides, enabling the identification of areas with exposed ferrous minerals associated with mineral extraction activities. These indices function by leveraging the differing reflectance properties of various surface materials, allowing for automated detection through remote sensing data analysis; data is often presented as a raster image where pixel values correspond to the index value at that location.

Effective analysis of remote sensing data, such as that from Sentinel-2, necessitates more than just data acquisition; it demands specialized tools and methodologies. These include atmospheric correction algorithms to mitigate distortions caused by atmospheric effects, geometric correction to ensure accurate spatial positioning, and image classification techniques – both supervised and unsupervised – to categorize pixel values into meaningful land cover classes. Furthermore, advanced techniques like time-series analysis are crucial for monitoring changes over time, while machine learning algorithms are increasingly employed for automated feature extraction and pattern recognition. Data calibration and validation procedures are also essential to ensure the accuracy and reliability of derived information, often requiring ground truth data for comparison and refinement of analytical models.

Intelligence from Above: AI and Landscape Interpretation

Multimodal Large Language Models (LLMs), including Llama-4 and GPT-5, represent a departure from traditional landscape data analysis techniques by integrating multiple data types directly into a single processing framework. These models are not limited to textual or numerical inputs; they can natively process imagery, such as satellite or aerial photography, alongside derived geospatial indices. This capability allows for a holistic assessment of landscape features without requiring separate pre-processing steps for each data source. By simultaneously considering visual and analytical data, LLMs can identify complex patterns and relationships often missed by conventional methods, offering improved accuracy and efficiency in landscape interpretation and monitoring.

Multimodal Large Language Models process landscape data by integrating imagery with spectral indices, such as the UDM Index. The UDM Index is derived from the Normalized Built-Up Difference Index (NBDI) and is specifically designed to improve the identification of built environments and mining operations within landscapes. This combined input allows the models to move beyond solely visual analysis, incorporating quantitative data about land cover composition. By analyzing both pixel values in imagery and the calculated UDM Index values, the models can more accurately classify landscape features and extract relevant information regarding human impact and land use patterns.

The UDM Index is a specialized metric designed to improve the identification of built-up environments and mining locations within landscape analysis. Calculated from spectral data, it builds upon the Normalized Built-Up Difference Index (NBDI) by incorporating additional spectral bands and algorithmic refinements to better differentiate impervious surfaces and disturbed land characteristic of these features. This results in a more sensitive and accurate detection capability compared to traditional NBDI-based methods, particularly in regions with complex land cover or subtle variations in built-up materials. The UDM Index serves as a complementary data layer, enhancing the robustness of landscape interpretation when integrated with other remote sensing datasets and analytical techniques.

Llama-4’s performance in landscape interpretation is significantly improved through the implementation of targeted System Prompts and Retrieval-Augmented Generation (RAG). System Prompts guide the model’s behavior and focus its analysis, while RAG enables the incorporation of external knowledge sources during inference. This combination expands the model’s effective context window to approximately 10 million tokens, allowing it to process and correlate significantly larger datasets of imagery and landscape indices – such as the UDM Index – than previous iterations. This expanded context is critical for understanding complex spatial relationships and accurately identifying features within large-scale landscapes.

From Data to Insight: Intelligent Automation and Global Impact

Agentic Retrieval-Augmented Generation, or RAG, represents a significant advancement over traditional RAG systems by introducing the capacity for complex, multi-step reasoning. Rather than simply retrieving information based on a single query, agentic RAG empowers the system to formulate a plan, break down a problem into smaller, manageable steps, and iteratively retrieve and synthesize information to arrive at a comprehensive understanding. This process isn’t simply about gathering data; it incorporates a crucial element of self-evaluation, where the system assesses the relevance and quality of retrieved information, refining its approach and ensuring a more robust and accurate outcome. By enabling this iterative reasoning and self-critique, agentic RAG moves beyond simple information retrieval towards a form of autonomous investigation and insightful analysis.

Landscape interpretation benefits from a sophisticated information retrieval process powered by Sentence Transformer Models. These models don’t simply search for keywords; instead, they analyze the semantic meaning of both the query and the available data – satellite imagery descriptions, geological reports, and even historical records – to identify genuinely relevant information. By converting text into numerical vectors that represent meaning, the model can assess the similarity between different pieces of information, even if they don’t share common words. This allows for the extraction of nuanced details crucial for understanding complex landscapes, surpassing the limitations of traditional keyword-based searches and enabling more accurate, context-aware analysis. The result is a targeted retrieval of information, fostering deeper insights into environmental changes and resource management.

The reliability of automated insights hinges on integrating human expertise through a Human-in-the-Loop paradigm. This approach doesn’t simply accept the outputs of artificial intelligence; instead, it actively incorporates human review and validation at critical stages. By subjecting AI-generated conclusions to scrutiny, potential inaccuracies or biases are identified and corrected, ensuring a higher degree of confidence in the final analysis. This collaborative process moves beyond mere automation, fostering a synergistic relationship where the strengths of both AI – rapid data processing – and human intelligence – nuanced understanding and contextual awareness – combine to deliver robust and trustworthy results. Ultimately, this validation step is not merely about error correction; it’s about building trust and ensuring responsible application of increasingly powerful automated systems.

The convergence of artificial intelligence and human discernment presents a transformative approach to global resource management, notably exemplified by detailed analyses of mining operations worldwide. This synergistic model doesn’t simply automate data collection; it leverages AI to identify relevant information, then integrates human expertise for validation and nuanced interpretation. The result is a system capable of not only detecting changes in land use and infrastructure at mining sites, but also providing contextual commentary on the implications of those changes. This collaborative intelligence offers a more comprehensive and reliable understanding of planetary resource dynamics, enabling proactive monitoring, informed decision-making, and ultimately, more sustainable practices for a rapidly evolving world.

The pursuit of documenting resource extraction via AI-powered landscape interpretation, as detailed in this work, echoes a fundamental principle of efficient understanding. The system strives to distill vast quantities of Earth observation data – satellite imagery, reports, and more – into a concise, accessible record. This aligns with the thought of Henri Poincaré: “It is better to have a few habits than many theories.” The presented system isn’t burdened by expansive, untested models; instead, it prioritizes a focused, practical approach to data retrieval and generation, creating a shared history through curated information. The elegance lies in what the system omits – superfluous complexity, allowing clarity to emerge from the essential patterns of environmental change.

What Remains?

The endeavor detailed here, the parsing of planetary wounds through algorithmic eyes, inevitably arrives at a familiar impasse. The system functions – it documents, correlates, and generates a shared, if asymmetrical, history. Yet, the true measure isn’t the volume of imagery processed, but the signal lost in the noise. Each pixel interpreted is also a pixel not understood. The ambition to comprehensively map resource extraction’s impact founders on the irreducible complexity of ‘impact’ itself – a concept laden with human values, and thus, fundamentally resistant to pure computation.

Future work will undoubtedly refine the multimodal integration, chase higher resolution, and expand the geographic scope. But a more pertinent question lingers: what constitutes ‘enough’? The pursuit of complete documentation feels less like scientific progress and more like a digital hoarding. Perhaps the value lies not in building a perfect record, but in designing a system that gracefully acknowledges its own limitations, highlighting what remains hidden rather than pretending to reveal everything.

The challenge, then, isn’t to build a more intelligent AI, but a more humble one. An instrument that understands its role is not to solve the problem of resource extraction, but to present a clearer, albeit incomplete, reflection of it. The elegance, ultimately, will be found not in what is added, but in what is left unsaid.

Original article: https://arxiv.org/pdf/2602.09299.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-02-11 16:19