Synthetic Reflections on Resource Extraction

1. Problem Statement

The paper addresses the challenge of interpreting the environmental impact of industrial-scale mining operations on a planetary scale using Artificial Intelligence. While extractivism is a fundamental driver of human civilization, it causes irreversible landscape degradation. The authors aim to:

• Translate the history of extractivism into a tangible artifact that AI systems can "ingest" and understand.
• Develop a computationally parsimonious pipeline that interprets satellite imagery of mining sites to produce succinct, accurate commentaries.
• Bridge the gap between raw geospatial data and human-understandable narratives, while also making this data accessible to other AI models via Retrieval-Augmented Generation (RAG).
• Challenge the narrative of AI as purely resource-hungry by demonstrating how older, efficient models can be repurposed for planetary computing.

2. Methodology

The authors propose a multi-stage pipeline combining satellite remote sensing, custom spectral indices, multimodal AI, and agentic reasoning.

A. Data Acquisition and Preprocessing

• Source: Sentinel-2 satellite imagery (10–60m resolution, 13 spectral bands) from the EU Copernicus program, chosen for its free, open data policy and spectral richness.
• Dataset: Approximately 200 distinct mining sites globally.
• Preprocessing:
  • Selection based on atmospheric conditions (low cloud cover) and seasonal constraints (snow-free).
  • Automated quality filtering using image statistics (contrast, sharpness, Shannon entropy).
  • Human-in-the-loop verification to ensure aesthetic and analytical acceptability.

B. Feature Engineering and Custom Indices

To enable AI models to distinguish mining from urban development (which both lack vegetation), the authors created specific spectral indices:

• Standard Indices: Normalized Difference Vegetation Index (NDVI) to detect vegetation loss; Normalized Built-Up Difference Index (NDBI) to detect built-up areas.
• Novel Contribution: Urban Dwelling and Mining Index (UDM).
  • A bespoke binary classifier trained on 10 Sentinel-2 bands.
  • Uses human annotations (positive samples for mining/urban, negative samples for non-relevant features) to train centroid-based classifiers.
  • Incorporates morphology and NDVI gating to reduce false positives (e.g., distinguishing roads or shadows from mining pits).

C. Multimodal AI Integration

• Model Selection: The team evaluated Kosmos-2, GPT-5, and Llama-4. They selected Llama-4 Maverick for its balance of performance, cost, and energy efficiency (consuming ~4-5x less power than GPT-5).
• Input Adaptation: Since Llama-4 accepts only 3-band RGB input, the authors preprocessed multi-spectral data (RGB + NDVI + UDM) into composite images to convey mining-specific features to the model.
• Prompt Engineering:
  • System Prompts: Strictly defined constraints to focus on environmental conditions, mining distribution, and urban sprawl while prohibiting vague quantifiers and hallucinations.
  • Context Injection: Each site is paired with 250-word metadata segments containing geology, history, and controversies (sourced from Wikipedia, corporate reports, and environmental agencies).

D. Evaluation Framework ("AI to Evaluate AI")

• Human Review: Visual experts verify image quality.
• Automated Text Evaluation: An older LLM (Gemini-Flash/Gemini 3) acts as a critic.
  • Uses a multi-shot instruction set to score outputs on five dimensions: environmental focus, terminology, pattern observation, constraint adherence, and conciseness.
  • Employs a per-criterion evaluation strategy (scoring each dimension in separate calls) to prevent intermediate dimensions from being skipped, ensuring structured JSON output.

E. Agentic Retrieval-Augmented Generation (Agentic RAG)

• To share findings with other AI systems without retraining them, the authors built an Agentic RAG system.
• Architecture:
  • Indexing: Text chunks (150 tokens) are enriched with metadata and embedded using sentence-transformers.
  • Three-Layer Hierarchy: Document map (metadata/headers) $\rightarrow$ Section summaries $\rightarrow$ Atomic evidence chunks.
  • Inference: The agent performs a coarse-to-fine retrieval cascade, refining queries if context is insufficient, and strictly constrains the LLM to generate responses based only on retrieved evidence to minimize hallucinations.

3. Key Contributions

1. Urban Dwelling and Mining Index (UDM): A novel, custom binary index that effectively distinguishes mining sites from urban settlements using multi-spectral data and human-in-the-loop training, outperforming standard NDBI.
2. Resource-Parsimonious Pipeline: Demonstrates that high-quality landscape interpretation can be achieved using efficient models (Llama-4) and open data (Sentinel-2) rather than relying solely on massive, energy-intensive frontier models (GPT-5).
3. Agentic RAG for Geospatial Data: A framework for connecting static satellite observations with dynamic, long-form contextual knowledge (e.g., books on Indigenous impacts), allowing external LLMs to reason across multiple mining sites and historical contexts.
4. Synthetic Reflections Artifact: A tangible "picture book" of extractivism that serves as a shared history between humans and AI, fostering a common ground for discussing planetary resource use.

4. Results

• Model Performance:
  • GPT-5: Produced the highest quality descriptions and could ingest native 13-band GeoTIFFs directly but required excessive compute time (>10 mins) and energy.
  • Llama-4: With the addition of NDVI and UDM indices and multi-shot prompting, Llama-4 achieved results comparable to GPT-5 in under 60 seconds with significantly lower energy consumption.
  • Kosmos-2: Failed to differentiate nuanced landscape features.
• Qualitative Output: The system successfully generated analytical, cool-toned descriptions of mining impacts (e.g., Kolwezi, Garzweiler, Nyurbinsky) that accurately identified vegetation loss, soil erosion, and water risks.
• Agentic RAG: Successfully enabled an external LLM (DeepSeek) to synthesize specific satellite observations with long-form academic texts regarding Indigenous land rights and environmental degradation in Australia, citing specific page numbers and sources.

5. Significance

• Ethical AI: The project reframes AI not as a competitor to humanity but as a tool to document and understand a shared, extractive history. It proposes a "human-in-the-loop" approach where human agency guides AI interpretation.
• Sustainability: By optimizing for energy efficiency and utilizing "no-longer-new" models, the paper advocates for a sustainable approach to planetary computing.
• Scalability: The pipeline offers a scalable method for NGOs and researchers to monitor global resource extraction without prohibitive costs.
• Future of AI Interaction: The Agentic RAG component suggests a future where specialized, curated datasets can be "nudged" into frontier models to improve their factual grounding and domain-specific reasoning without the need for expensive retraining.