Climate AI Annotation: Satellite Imagery, Carbon Accounting, and Emissions Detection

Why Climate AI Is an Annotation-Intensive Domain

The climate AI market is growing faster than most AI verticals because the demand is regulatory-driven rather than discretionary. The EU Corporate Sustainability Reporting Directive (CSRD), the SEC climate disclosure rule, the Australian ASIC climate risk reporting framework, and voluntary carbon market integrity requirements all create demand for verified, spatially explicit measurements of emissions, land use change, and biodiversity impact. AI models that produce these measurements require training data annotated to science-grade standards — not the kind of quality that general-purpose annotation platforms can deliver.

The numbers are large. The global environmental monitoring AI market was valued at USD $2.1 billion in 2024 and is projected to reach $7.8 billion by 2030 (MarketsandMarkets, 2024), growing at a 24.4% CAGR. A 2023 World Resources Institute study found that satellite-based deforestation monitoring programmes that use AI now cover approximately 1.4 billion hectares of tropical forest — generating annotation requirements at a scale that manual interpretation alone cannot sustain. The quality bottleneck is not the satellite data itself; it is the specialist annotation that turns raw imagery into model-ready training labels.

Our geospatial annotation service covers the four primary climate AI annotation domains:

Satellite Imagery Annotation: The Multi-Temporal Challenge

The distinctive annotation challenge in climate AI is the multi-temporal dimension. Unlike standard image annotation where each image is annotated independently, climate AI annotation typically involves comparing two or more images of the same location at different time points to detect change. This requires annotators to apply consistent label definitions across images acquired under different atmospheric conditions, seasonal states, and with different sensor geometries.

Deforestation change detection illustrates the problem clearly. A pixel classified as “intact forest” in a January dry-season Sentinel-2 composite must be compared to the same pixel in a December wet-season composite a year later. The forest canopy reflectance in the January dry season is different from the December wet season — a consistent forest appears to have changed simply because of phenological variation. Without harmonised time series processing and annotator protocols that account for seasonal reflectance variation, a change detection model trained on this data will generate systematic false positives in seasonally variable forest types.

The solution used in production climate AI annotation datasets is to annotate image pairs rather than individual images, using cloud-masked and atmospherically corrected surface reflectance composites with explicitly matched seasonal windows. Annotators work with multi-band false-colour composites (typically SWIR1-NIR-Red for vegetation change detection) rather than natural colour imagery, because near-infrared and shortwave infrared bands provide more robust deforestation signals than visible bands under different atmospheric conditions.

Case Study: Deforestation Monitoring AI for a Voluntary Carbon Market Verifier

A voluntary carbon market verification organisation operating REDD+ (Reducing Emissions from Deforestation and Forest Degradation) projects across Southeast Asia was building a deforestation monitoring AI to replace costly annual verification visits. Their existing approach — visual interpretation of Sentinel-2 imagery by a single remote sensing analyst — was identifying approximately 68% of deforestation events in their project areas at a cost of 6–8 weeks of analyst time per annual monitoring cycle.

The initial AI model, trained on 1,200 annotated Sentinel-2 image pairs produced by the in-house analyst, was achieving a deforestation detection recall of 0.61 and a precision of 0.54 on a held-out test set — below the 0.85 recall threshold required to meet Verra VCS (Verified Carbon Standard) monitoring requirements. Error analysis identified three systematic failure modes: false negatives on selective logging events (where individual trees are removed without visible canopy gap from 10 m resolution imagery), false positives in seasonal wetland areas that alternated between inundated and vegetated surface in wet and dry seasons, and inconsistent annotation of agroforestry systems where the boundary between forest and agricultural use is inherently ambiguous.

Our approach: We assembled a team of four annotators with tropical forest remote sensing backgrounds and rebuilt the annotation protocol from the PRODES and Global Forest Watch deforestation definition frameworks. The selective logging failure mode was addressed by adding a “degraded” class distinct from the binary “deforested/intact” scheme — a class that could only be assigned when canopy cover had visibly decreased in the NIR-Red-Green false colour composite without a complete clearing event. Seasonal wetland false positives were resolved by incorporating a third time point (dry-season wet-year baseline composite) as a mandatory reference layer for all wetland-adjacent annotations, enabling annotators to rule out inundation-cycle changes. Agroforestry ambiguity was resolved by a decision tree in the annotation protocol: tree cover >30% canopy closure in the SWIR1-NIR-Red composite = forest class; tree cover <30% with cultivated ground cover = non-forest agricultural.

The deforestation recall improvement from 0.61 to 0.88 — above the VCS 0.85 threshold — came primarily from the protocol redesign rather than annotator quality differences. The selective logging recall improvement from 0.23 to 0.71 came from the addition of the “degraded” class as a distinct training category, which gave the model a label that matched the visual phenomenon rather than forcing it into a binary scheme where degradation events were split unpredictably between “intact” and “deforested.” The seasonal wetland false positive improvement came from the three-time-point annotation protocol. The organisation proceeded to Verra VCS validation and reduced their annual monitoring cost by approximately 80%. Our geospatial annotation service handles these multi-temporal, multi-class satellite annotation requirements at project scale.

Methane Plume Annotation: The Hardest Task in Climate AI

Methane plume annotation is the most technically demanding task in climate AI annotation. GHGSat, EMIT, and TROPOMI satellites retrieve column-integrated methane concentration by analysing sunlight reflected from the Earth's surface in shortwave infrared spectral bands. The retrieved concentration anomaly (the plume signal) depends not only on the methane in the atmosphere but also on the surface albedo at 2.3 μm — meaning that dark surfaces (water bodies, dark soil, vegetation canopy) can produce apparent concentration anomalies that are retrieval artefacts rather than real methane signals.

A 2023 analysis of EMIT methane detections published in Science found that approximately 30–40% of initial “candidate plumes” identified by automated retrieval algorithms were subsequently rejected by expert human review as surface albedo artefacts, retrieval noise, or co-located non-methane absorbers. The implication for annotation is stark: naive polygon annotation of concentration anomalies by annotators without atmospheric science background will introduce a ~35% false positive rate into the training data, producing a model that confidently detects non-existent plumes at scale.

Production methane plume annotation protocols require annotators who can assess plume morphology (real plumes are elongated downwind, whereas albedo artefacts follow terrain features), check wind field consistency (does the plume orientation match the ERA5 reanalysis wind direction at the acquisition time?), and identify known surface albedo problem areas (salt flats, evaporation ponds, bare soil with high carbonate content). For large-volume annotation projects, a two-tier approach is standard: trained remote sensing annotators apply a binary plume/no-plume classification with a conservative false-positive preference, and atmospheric scientists review the positive set to filter artefacts before the data is committed to training.

ESG AI Annotation: What Corporate Sustainability Monitoring Requires

The ESG AI annotation market is distinct from scientific climate research annotation in both its data types and its quality requirements. Corporate ESG monitoring AI typically operates on a combination of satellite imagery, corporate disclosure documents, supply chain databases, and social media data — a multi-modal annotation requirement that differs from the pure imagery annotation of scientific climate AI.

Satellite-derived ESG annotation covers supply chain deforestation tracing (linking deforestation events to upstream agricultural commodities such as soy, palm oil, and cattle via spatial overlay with supply chain footprint data), facility-level emissions monitoring (annotating industrial facility activity levels from Sentinel-2 imagery — whether a facility is operating, idle, or under construction), and physical climate risk assessment (annotating flood exposure, wildfire risk boundary, and coastal erosion data for corporate asset locations).

The regulatory stakes in ESG annotation are significant. Under the EU CSRD, large companies and listed SMEs must disclose material climate risks and impacts using double materiality analysis — an assessment that increasingly uses AI models trained on annotated data to quantify physical risk exposures and supply chain deforestation footprints. Where an AI model's output is used as the basis for a CSRD disclosure, the quality of the training data underlies the accuracy of the regulatory filing. Annotation errors that inflate or undercount deforestation exposure, or misclassify facility activity levels, create material disclosure errors with regulatory and liability consequences.

For broader context on geospatial annotation requirements, our post on geospatial annotation for mapping and Earth AI covers the aerial imagery and satellite annotation workflows that underpin asset inspection, infrastructure monitoring, and urban change detection. For annotation applications in the adjacent aerospace and defence domain that also uses multi-spectral and SAR satellite data, our guide to aerospace and defence AI annotation covers EO/SAR annotation for surveillance AI. Our full geospatial annotation service handles all categories of satellite and aerial imagery annotation for environmental and climate AI applications.

Building a Climate AI Annotation Dataset: What You Need Before You Start

Climate AI annotation projects fail at the protocol stage more often than the execution stage. The annotation protocol for a deforestation detection dataset must answer at least six questions before production annotation can begin: What is the minimum deforestation event size to include (typically 0.5 ha for REDD+ compliance)? What canopy cover threshold defines “forest” (FAO definition: 10% canopy cover; PRODES definition: 75% canopy closure)? What cloud cover threshold makes an image unacceptable for annotation? What is the seasonal window for before/after image pairs? What are the boundary rules for forest/non-forest transitions at the edge of plantation crops? And how are agroforestry systems classified — as forest, agriculture, or a distinct class?

Each of these choices has downstream consequences for model performance on specific failure modes. The canopy cover threshold choice alone can change deforestation area estimates by 15–25% between PRODES and FAO definitions applied to the same imagery. If the annotation protocol does not specify the reference definition, individual annotators will make different choices at ambiguous sites, producing inconsistent training labels that degrade model performance precisely at the forest/non-forest boundary — the decision the model is being built to automate.

Annotator calibration is equally critical. Remote sensing analysts can disagree substantially on deforestation boundaries in heterogeneous forest types — inter-annotator agreement on deforestation events in logged-over selectively harvested forest is typically lower (Kappa ~0.65–0.75) than in complete clearing events (Kappa ~0.88–0.95). A calibration set of 100–150 pre-agreed reference image pairs, annotated to consensus by domain experts before production annotation begins, is the minimum requirement for production-grade climate AI training data.

Frequently Asked Questions