GSA Resource Security
Satellite-derived resource footprint auditing for Malaysia's data centre sector
Live visual library · Methodology · References · Known limitations · Sinar Project
GSA Resource Security is an independent, satellite-derived audit of the physical resource footprint of data centre facilities in Malaysia: estimated power demand, energy consumption, water use, carbon emissions, and land-use/deforestation impact, cross-referenced against the regulatory and legislative record. It was built because operators routinely withhold facility-level resource figures under non-disclosure agreements and national-security framing, leaving no independent, verifiable account of what this sector actually costs the country.
Developed by Sinar Project with support from the Global South Alliance Datafication and Democracy Fund, as part of the Resource Security: Data Center Resource Extraction and Autonomy in the Global South project.
Status: research prototype produced for a fixed-term subgrant (March–September 2026). Not maintained as production software; see Known Limitations before relying on any figure.
- Satellite building detection: computer-vision pipeline (OpenCV) that independently measures each facility's physical footprint from satellite imagery, rather than relying on operator-claimed capacity.
- Confidence-scored, abstaining detector: colour-space masking, dual-direction Otsu thresholding, and multi-factor contour filtering, with an explicit confidence score and no forced detection where nothing reliable is found.
- Deduplication: collapses co-located multi-tenant entries into physical-building records, with explicit handling for genuinely distinct multi-building campuses sharing one imprecise geocode.
- Resource-extraction estimation: converts footprint area into estimated power, annual energy, carbon emissions, and water use (optimised and regional water-usage-effectiveness scenarios), keeping operator-disclosed and modelled facilities strictly separate.
- Independent QA re-check: a second, separate verification pass (geometric re-check, coordinate-region plausibility check) run across the full dataset, not just the original validation sample; findings are applied as a scripted pipeline step, not a manual edit.
- Regulatory and legislative corpus analysis: environmental-review policy mapping across every Malaysian jurisdiction with a measured facility, plus a systematic keyword search and triage of the federal and Johor State Assembly Hansard record (2008–2026).
- Land-cover change rollout: facilities clustered into geographic corridors and checked against Hansen Global Forest Change data to quantify deforestation associated with development corridors.
- Public visual library: an interactive map, searchable/filterable gallery of every measured facility's satellite capture, and per-facility resource estimates, published as a static site (see Live visual library).
Python, OpenCV, NumPy · Google Places API / Google Static Maps API (imagery acquisition) · Global Forest Watch / Hansen Global Forest Change (land-cover data) · Leaflet.js (visual library map) · vanilla HTML/CSS/JS (static site, no build step)
src/ pipeline scripts (see Pipeline Order below)
docs/ public GitHub Pages site: interactive map, visual library, per-facility estimates
docs/reports/ companion deliverable reports (methodology, valuation, policy audit, briefs), see References
logs/ running workflow/methodology log
This repository contains the pipeline code, a derived public dataset extract (docs/data/manifest.json and the satellite captures under docs/images / docs/thumbs, covering the 71 measured buildings and their resource estimates), and the companion deliverable reports (docs/reports/, see References). It does not contain the underlying raw facility coordinate database, the full satellite imagery archive, or the policy/Hansard document corpus itself, which are maintained separately by the project team on request.
Requires Python 3.10+, opencv-python, numpy, and a Google Maps Platform API key (Places API + Static Maps API enabled) for imagery acquisition steps.
git clone https://github.com/Sinar/GSA_ResourceSecurity.git
cd GSA_ResourceSecurity
pip install opencv-python numpysatellite_cv_pipeline.py / improved_cv.py
→ validate_bboxes.py
→ dedup_buildings.py
→ qa_recheck.py (review findings manually)
→ apply_qa_corrections.py
→ resource_estimates_v3.py
Each script's role is documented in its module comments. qa_recheck.py is reporting/discovery only, findings require manual visual confirmation before apply_qa_corrections.py applies them.
The live site (source in docs/) provides:
- An interactive map of all 71 independently measured facilities, colour-coded by inclusion status.
- A searchable, filterable, sortable gallery of each facility's satellite capture with its detected bounding box.
- A detail view per facility: footprint area, estimated power/energy/carbon/water/waste heat, detection confidence, QA flags, and any independent-verification notes.
To run it locally: cd docs && python3 -m http.server 8000, then open localhost:8000.
The facility list was compiled and satellite imagery acquired during Phase 2 of the project workflow (March-June 2026); CV modelling, resource estimation, and the independent QA re-check ran across that same window, with the land-cover change rollout and Hansard corpus search completed by August 2026. Figures reflect the state of public disclosure as of that period, not real-time; see Known Limitations below.
- Imagery acquisition: candidate facility addresses/coordinates resolved via the Google Places API, satellite imagery pulled via the Google Static Maps API.
- Building detection: a confidence-scored, abstaining OpenCV pipeline (colour-space masking, dual-direction Otsu thresholding, multi-factor contour filtering) measures each facility's physical footprint; it declines to output a box where nothing reliable is found rather than forcing a detection. 124 candidate facilities were processed.
- Validation and deduplication: detected boxes are checked against source imagery, then co-located multi-tenant entries are collapsed into physical-building records (with an explicit, documented exception for genuinely distinct multi-building campuses sharing one imprecise geocode, e.g. AirTrunk JHB2/JHB3/JHB4). This reduced the candidate set to 71 physical buildings.
- Independent QA re-check: a second, separate verification pass (geometric re-check, coordinate-region plausibility check) run across the full 71-building dataset, not just the original validation sample. Found 11 buildings (15.5%) with an issue not caught by the pipeline's original automated flags, including three facilities whose coordinates resolve outside Malaysia.
- Resource estimation: footprint area converted to estimated power, annual energy, carbon emissions, and water use, with operator-disclosed and modelled facilities kept strictly separate rather than blended into one aggregate (see Valuation Benchmarking below for why).
- Regulatory and legislative corpus analysis: environmental-review policy mapping across every Malaysian jurisdiction with a measured facility, plus a systematic keyword search of the federal and Johor State Assembly Hansard record (2008–2026), and a comparative desk review of Indonesia's environmental-review architecture.
- Land-cover change rollout: facilities grouped by single-linkage geographic clustering (8 km threshold) into development corridors, each checked as a Hansen Global Forest Change area of interest via Global Forest Watch to quantify deforestation associated with the corridor.
Full formulas, coefficients, and caveats are in the companion reports listed under Documentation.
The economic valuation paper benchmarks the pipeline's modelled power estimates against operator-disclosed capacity for the subset of facilities (7 of 71) where a specific MW figure is publicly disclosed, rather than against a single published industry coefficient treated as ground truth. The published reference coefficient used for the initial comparison is 1.5 kW/m² power density at PUE 1.4, a figure drawn from general industry guidance rather than a facility-specific measurement. Applying that coefficient to the disclosed-capacity facilities' CV-detected footprint area and comparing the implied density against each facility's actual disclosed density produced a spread of more than two orders of magnitude (approximately 0.36 kW/m² for a large colocation facility to approximately 40 kW/m² for a high-density liquid-cooled AI facility fragment). The paper argues this spread is not measurement noise but evidence of vintage capital heterogeneity (Solow's putty-clay framework) across a portfolio built over a decade or more, and situates the finding within the natural resource and industrial ecology aggregation-bias literature and the UN SEEA Central Framework for natural capital accounting (see References).
Other figures benchmarked in the paper: Malaysia's pledged data centre investment (RM280 billion since 2021, RM131 billion/47% realised as of May 2026, cited from public investment-tracking reporting); Peninsular Malaysia's estimated annual grid generation, against which the portfolio's modelled ~31,023.5 GWh/yr (~20.0% of grid generation) is compared; and a 360 MW Nvidia-backed campus in Batam, Indonesia, used as a comparative single-facility benchmark against the entire 71-building Malaysian dataset.
These are the technical methodology, resource-estimation formulas, regulatory/legislative corpus analysis, land-cover change rollout, and full caveats, maintained alongside the underlying data corpus. They are grant deliverables under the project's funding agreement, published as static files under docs/reports/ in this repository.
- Full Project Methodology - Satellite, Regulatory Corpus, and Land-Cover Change
- Economic Valuation Paper Draft - Vintage Capital Heterogeneity and Hidden Natural Capital Loss
- GSA Master Facility Database - Methodology Note and Data Dictionary
- QA Sample Report - Positional Accuracy and CV Extraction Spot-Check
- Policy Audit Report - Malaysia Data Centre Governance with Indonesia Comparison
- Comparative Regulatory Desk Review - Indonesia vs Malaysia Data Centre Governance
- Resource Security - Public Policy Report
- Policy Brief 1 - Closing the Cumulative-Impact Blind Spot (regulator-facing)
- Policy Brief 2 - Checking a Data Centre's Footprint (journalist-facing)
- Preliminary Results Report - Expanded Background and Assessment
- Malaysia Environmental Quality Act 1974 and Environmental Quality (Prescribed Activities) (Environmental Impact Assessment) Order 2015 — Department of Environment Malaysia; full EIA Order 2015 text via DOE eSWIS
- Indonesia Government Regulation No. 22 of 2021 (AMDAL / UKL-UPL / SPPL environmental-approval framework) — Ministry of Environment and Forestry (KLHK)
- UN System of Environmental-Economic Accounting (SEEA) Central Framework — UN Statistics Division
- Hansen et al. Global Forest Change dataset — University of Maryland / Google Earth Engine
- Global Forest Watch — World Resources Institute, area-of-interest analysis tool used for the land-cover change rollout
- Global South Alliance Datafication and Democracy Fund — funding agreement governing this project; contact the project team for details
The technical methodology, resource-estimation formulas, regulatory/legislative corpus analysis, land-cover change rollout, and full caveats are documented in the set of companion reports listed under References, published as static files under docs/reports/ and maintained alongside the underlying data corpus.
- Geocoding fallback failure modes exist for facilities whose operators don't disclose physical addresses (parent-company registered office; town-level centroid). Three instances found a facility's coordinate resolving to a different country entirely: one to Metro Manila, Philippines, and two (a single corridor) to Marina Bay, Singapore, the latter discovered during a downstream land-cover change analysis rather than by
qa_recheck.pyitself. - No single power-density coefficient is defensible across the full facility portfolio (observed range spans over 100x between low-density colocation and high-density AI/liquid-cooled facilities); resource estimates therefore report disclosed-capacity facilities separately from modelled facilities rather than blending into one aggregate.
- An independent verification pass across the full 71-building dataset, run across three separate discovery routes (automated re-check, manual visual sample, and the land-cover change rollout), found 11 buildings (15.5%) with an issue not caught by the pipeline's own original automated flags. The dataset should be described as reviewed and substantially reliable, not fully verified.
- This is a fixed-term research output, not a continuously maintained monitoring system; figures reflect the state of publicly available information as of mid-2026 and will drift out of date as facilities are built, disclosed, or decommissioned.
This is a research subgrant output, not a supported product. For questions about the methodology or data, contact the project team via Sinar Project.
- Adhura H. Farouk, PhD Candidate in Economics, International Islamic University Malaysia (IIUM). Principal Investigator.
- Hadirah Husna Huzaidi, Final-year B.Econs, International Islamic University Malaysia (IIUM). Research Assistant.
Produced with support from the Global South Alliance Datafication and Democracy Fund and Sinar Project.
Not yet specified. Coordinate with Sinar Project before reuse or redistribution.