Prediction of changes in war-induced population and CO₂ emissions in Ukraine using social media

The study addresses the urgent need to monitor rapid, war-induced changes in population distribution and anthropogenic CO₂ emissions, highlighted by the large-scale displacement and infrastructure damage following Russia’s invasion of Ukraine on February 24, 2022. Traditional data collection (field and telephone surveys) is slow and often infeasible during conflicts, while standard emission inventories are annual and lagging, and satellite-based estimation faces signal-to-variability challenges. Social media usage among refugees and the availability of geo-tagged data offer a way to nowcast population shifts. Existing global emission datasets (GRACED, ODIAC, EDGAR) either lag reality or may not reflect abrupt, war-driven changes, especially where local data are scarce or population shifts are not accounted for. This work combines geo-tagged Twitter data with baseline GRACED emissions and LandScan 2021 population to estimate monthly, fine-grained (0.1° × 0.1°) changes in population and sectoral CO₂ emissions in Ukraine. Contributions: (1) Estimation of monthly war-induced population displacement and sectoral CO₂ emission changes at 0.1° × 0.1° resolution; (2) Provision of supplementary, social-media-driven information on emission changes that may be overlooked in global datasets.

The paper situates its approach within prior evidence that refugees use smartphones and social media during migration, enabling population and displacement estimation from platforms such as Twitter and Facebook. Prior studies have used Facebook’s advertising platform to estimate population changes in Ukraine during the war, but spatial coverage gaps exist (e.g., Luhanska, Donetska before March 11, 2022, Crimea, Sevastopol), and fine-grained spatial patterns remain unclear. Regarding emissions, official inventories (EDGAR, national statistics) are annual with substantial lag; ODIAC and EDGAR lack near-real-time capability, while GRACED offers near-real-time estimates but may miss abrupt, conflict-driven shifts and sectoral nuances tied to population displacement. The study leverages these insights to integrate social media signals with baseline emissions and population data to fill temporal and spatial monitoring gaps.

Data sources: (1) Geo-tagged Twitter data in Ukraine from 2022-01-01 to 2023-02-28 via the Twitter streaming API (119,046 geo-tagged tweets). Each tweet includes user ID, latitude/longitude, and timestamp. Preprocessing included monthly grouping, deduplication of same-user same-location posts, and monthly counting per 0.1° × 0.1° grid to proxy the quantity of Twitter active users. (2) GRACED anthropogenic CO₂ emissions (0.1° × 0.1°, daily) for residential consumption, ground transport, and industry sectors from 2021-01-01 to 2023-02-28; daily values aggregated to monthly averages. (3) LandScan Global 2021 population counts (~1 km), downsampled to 0.1° × 0.1° by summing counts as the baseline population. Population change estimation: Compute, for each grid and month, the change ratio of geo-tagged tweet counts relative to January 2022: CR(grid,m,y) = T(grid,m,y) / T(grid,January,2022). Estimate monthly population change relative to January 2022 as PC(grid,m,y) = CR(grid,m,y) × Pop(grid,2021). CO₂ emission change estimation: First compute baseline per-capita emissions in 2021 for each sector and grid: EPC_res(grid,m,2021) = E_res(grid,m,2021) / Pop(grid,2021); similarly for EPC_tran and EPC_ind. Then estimate monthly emission changes during the war (Feb 2022–Feb 2023) relative to January 2022 using: EC_res(grid,m,y) = PC(grid,m,y) × EPC_res(grid,m,2021) × TA_res(grid,m); EC_tran(grid,m,y) = PC(grid,m,y) × EPC_tran(grid,m,2021); EC_ind(grid,m,y) = PC(grid,m,y) × EPC_ind(grid,m,2021). Assumptions: population displacement decreases residential and industry emissions, while movement-induced congestion increases ground transport emissions. A temperature adjustment factor for residential emissions accounts for seasonal heating variability: TA_res(grid,m) = E_res(grid,Jan,2021) / E_res(grid,m,2021). Validation: Compare estimated monthly sectoral CO₂ emission changes with GRACED-derived changes during the war via linear regression, evaluating R² and slope for Ukraine and selected oblasts. Spatial analyses were conducted at 0.1° × 0.1° resolution, with additional proximity checks to conflict events (ACLED) for plausibility of population change patterns.

- Population displacement: After one year (Feb 2023 vs Jan 2022), 245 0.1° × 0.1° grid cells show population decline; about 60% decline by more than 10,000 people. Kiev City has the largest decline (≈464,000). In total, over 11 million people are displaced from their baseline grid cells. Regional declines: Kiev Oblast (including Kiev City) −2.09 million (≈44% of baseline), Donetsk Oblast −0.97 million (≈24%), Lviv Oblast −0.67 million (≈27%). - Sectoral CO₂ changes (Feb 2023 vs Jan 2022, Ukraine totals): Residential consumption decreased by 336.21 kt CO₂; ground transport increased by 58.83 kt CO₂; industry decreased by 483.3 kt CO₂. Selected oblasts: • Residential: Kiev −52.93 kt; Donetsk −28.10 kt; Lviv −18.17 kt. • Ground transport: Kiev +8.79 kt; Donetsk +6.94 kt; Lviv +3.13 kt. • Industry: Donetsk −179.86 kt; Kiev −16.38 kt; Lviv −9.57 kt. - Monthly dynamics: During the initial invasion (Feb–Apr 2022), population displacement increased by 1.79 million; ground transport emissions rose by 11.72 kt; residential and industry emissions fell by 60.75 kt and 20.43 kt, respectively. From May–Sep 2022, changes stabilized; average changes during this period were population −6.43 million, residential −207.35 kt, ground transport +42.57 kt, industry −166.4 kt. From Oct 2022–Feb 2023, new displacement (>7.56 million) followed infrastructure strikes; average sectoral changes: residential −207.35 kt, ground transport +42.57 kt, industry −166.4 kt. - Validation against GRACED: Significant linear relationships between estimated and reference sectoral changes. Residential (R² range by month 0.57–0.93); example Feb 2023 regressions: Ukraine R²=0.93; Lviv 0.92; Kiev 0.97; Donetsk 0.97. Ground transport (R² 0.41–0.90); e.g., Ukraine 0.90; Lviv 0.89; Kiev 0.92; Donetsk 0.89. Industry (R² 0.74–0.99); e.g., Ukraine 0.99; Lviv 0.87; Kiev 0.96; Donetsk 0.99.

The findings show that social media-derived signals can effectively capture rapid, fine-grained population displacement patterns during conflict, aligning with external references and correlating with proximity to political violence events. Integration with baseline per-capita sectoral emissions enables near-real-time estimation of CO₂ responses: declines in residential and industry emissions track with outmigration and disrupted activity, while ground transport emissions increase with movement and congestion, concentrated around capitals and administrative centers. Comparisons with GRACED reveal strong agreement overall but also highlight limitations of reference datasets in capturing war-induced shifts: residential emissions in GRACED assume stable annual totals and may underrepresent population-driven changes; ground transport spatial distribution in GRACED is assumed relatively stable, potentially missing conflict-related congestion; industry estimates rely on proxy data for Ukraine, limiting accuracy. The method thus complements existing inventories by incorporating dynamic population signals. Cross-validation with Facebook-based studies, IOM surveys, and national demographic assessments further supports plausibility. The approach can inform humanitarian targeting, infrastructure planning, and climate policy during crises by providing timely, spatially resolved indicators of both population and emission changes.

The study presents a fine-grained, near-real-time method to monitor war-induced changes in population and anthropogenic CO₂ emissions by combining geo-tagged Twitter data with baseline GRACED emissions and LandScan population data. Despite the lack of official displacement statistics, estimated population changes align with independent studies and institutional data. Estimated monthly sectoral CO₂ changes agree well with GRACED, with monthly R² ranges of 0.57–0.93 (residential), 0.41–0.90 (ground transport), and 0.74–0.99 (industry). Social media data provide valuable supplementary information on conflict-driven emission dynamics that may be overlooked by traditional datasets. The method is transferable to other conflict zones or disaster contexts and can support public health, reconstruction, and sustainability planning through rapid, spatially explicit insights.

- Social media bias and sparsity: Twitter users are not a representative sample; only 1–2% of tweets are geo-tagged, leading to demographic and spatial biases. - Assumption-driven estimation: Population changes are proxied by changes in geo-tagged tweet activity; sectoral emission changes assume proportionality to population change and, for residential, rely on a temperature adjustment based on 2021 seasonality. - Reference dataset constraints: GRACED may not fully capture conflict-induced changes (e.g., fixed annual totals for residential, stable spatial patterns for transport, proxy-based industry data for Ukraine), affecting validation. - Data gaps and coverage: Facebook marketing data are unavailable in some regions; local industrial statistics were not available for Ukraine, necessitating reliance on other countries’ data in GRACED. These factors may affect generalizability and introduce uncertainty in spatial and temporal estimates.