Machine learning and InVEST assess future carbon storage in Prague region

In a development that could reshape how cities worldwide plan for climate mitigation, researchers have unveiled a sophisticated artificial intelligence framework capable of predicting how land-use decisions made today will determine the carbon storage capacity of a major European metropolitan area over the next quarter century. The study, focused on the Prague Metropolitan Area in the Czech Republic, combines explainable machine learning, cellular automata modeling, and ecosystem services assessment to simulate three divergent futures for one of Central Europe’s most dynamically changing urban landscapes. The findings arrive at a critical juncture, as the Czech land-use, land-use change, and forestry sector has already flipped from a net carbon sink to a net carbon emitter, releasing approximately 15 megatonnes of CO2-equivalent in 2021 alone. Under the European Union Climate Law, the country is legally obligated to restore its land-sector carbon sink to at least negative 827 kilotonnes of CO2-equivalent per year by 2030, a target that hinges directly on how land is allocated between agriculture, forest, urban development, and natural vegetation in the coming decades. The research team, led by Khalil Gholamnia with contributions from Omid Ghorbanzadeh, Thomas Blaschke, and Lucie Kupková, developed what they describe as a unified, Python-based pipeline that bridges a longstanding gap in environmental modeling: the disconnect between predicting where land-use change will occur, understanding why it occurs, and quantifying what it means for regional carbon budgets.

The framework’s architecture rests on three interlocking computational pillars. At its foundation lies a Markov chain analysis that quantifies historical land-use transition probabilities from observed satellite-derived land-cover maps spanning 2018 to 2021, generating a mathematical matrix where each cell represents the likelihood of one land class converting to another over time. This Markov-derived demand projection establishes how much of each land type should exist by 2050 under different assumptions. The second pillar employs a Random Forest classifier, an ensemble machine learning algorithm consisting of 400 decision trees, each trained to recognize the biophysical, climatic, and socioeconomic conditions under which specific land transitions are likely to occur. Unlike conventional CA-Markov models that rely on static heuristic suitability rules, the Random Forest approach captures nonlinear relationships between driving factors and land-change outcomes, effectively learning from 14 distinct predictor variables including elevation, slope, aspect, temperature, precipitation, population density, and Euclidean distances to roads, railways, commercial zones, industrial areas, hotels, villages, and suburban settlements. The third pillar is a cellular automata allocation engine that spatially distributes the projected land-use changes across the landscape, constrained by neighborhood interactions, transition resistance factors, and scenario-specific policy rules. The composite transition potential for each pixel is calculated as the product of Random Forest probability, a transition resistance multiplier, and a neighborhood influence term weighted by a tunable coefficient, producing a spatially explicit probability surface that guides where each hectare of change materializes.

What distinguishes this study from prior land-change modeling efforts is its insistence on transparency through the integration of Shapley Additive Explanations, or SHAP, an explainable artificial intelligence technique rooted in cooperative game theory. SHAP quantifies the marginal contribution of each driving factor to every individual prediction, revealing not only which variables matter most but also the direction and nonlinear threshold effects of their influence on land-use transitions. For instance, the method can reveal whether the probability of cropland converting to built-up area increases sharply once distance to a major road falls below a certain threshold, or whether population density exerts an accelerating rather than linear effect on urbanization probability. This interpretability layer transforms the modeling framework from a black-box predictor into a diagnostic instrument capable of informing policy debates. The researchers argue that most existing studies either emphasize prediction accuracy without explaining the mechanistic drivers of land transitions, or assess carbon storage without explicitly linking those transitions to future carbon dynamics, a gap that becomes particularly consequential in heterogeneous metropolitan regions where urban growth, agricultural abandonment, and vegetation recovery occur simultaneously and create competing carbon outcomes across relatively small geographic areas.

The study area itself presents an unusually rich test bed for the methodology. The Prague Metropolitan Area encompasses approximately 190,000 hectares of exceptionally diverse landscape within a compact geographic extent, where the Vltava River carves through rolling hills and river valleys, creating a mosaic of urban fabric, cropland, mixed forest, grassland, and wetlands. Between 2001 and 2021, the metropolitan population grew by roughly 22 percent, accompanied by well-documented conversion of agricultural land to urban development in peri-urban zones. This combination of rapid suburbanization, persistent agricultural activity, and forest regeneration within a single bounded region creates precisely the kind of spatially complex carbon trade-offs that the researchers sought to capture. Land-cover data were derived from the Esri Land Cover product, generated from Sentinel-2 satellite imagery at 10-meter resolution using a deep learning classification framework, with temporal snapshots analyzed for 2018, 2021, and 2024. The original nine-class global product was reduced to seven dominant classes relevant to the study area, with all layers reprojected to WGS 84/UTM Zone 33N and resampled to a consistent 10-meter grid. Driving variables were assembled from the ASTER Global Digital Elevation Model Version 3, WorldClim Version 2.1 climate normals, OpenStreetMap infrastructure vectors, and population density surfaces derived from Esri demographic datasets and Charles University spatial databases. Multicollinearity among predictors was assessed using Pearson correlation analysis and Variance Inflation Factor calculations on a random sample of 120,000 valid pixels, with only the elevation-temperature pair exceeding the conventional correlation threshold at r equals negative 0.93, a relationship retained because both variables represent distinct environmental processes and because Random Forest’s random feature selection mechanism renders it inherently robust to collinearity.

Three contrasting scenarios were constructed to bracket the plausible range of future land-use trajectories through 2050. The Business-as-Usual scenario projects continuation of historical 2018-to-2021 transition dynamics without policy intervention, deriving land demand directly from the observed Markov transition matrix. The Urban Growth Scenario accelerates built-up expansion by promoting conversion of cropland, grassland, and bare land to urban fabric near existing settlements, road corridors, and suburban zones, while simultaneously reducing ecological protection coefficients. The Ecological Optimization Scenario inverts these assumptions, promoting transitions toward tree cover, grassland, water bodies, and flooded vegetation while strongly restricting conversion of ecologically valuable classes to built-up land through enhanced protection coefficients and greater resistance to urban expansion. Under the Business-as-Usual projection, tree-covered areas increase from 44,927 hectares in 2021 to approximately 55,647 hectares by 2050, while cropland declines from 78,483 hectares to 65,000 hectares, reflecting continued land conversion pressures. The model was calibrated against observed 2018-to-2021 transitions and validated by simulating the 2024 land-cover map and comparing predictions against the actual Sentinel-2-derived classification, providing an empirical check on spatial accuracy before scenario projections were generated for the 2024-to-2050 period.

Carbon consequences were quantified by coupling each simulated land-use map to the InVEST Carbon Storage and Sequestration model, which estimates ecosystem carbon stocks across four pools: aboveground biomass, belowground biomass, soil organic carbon, and dead organic matter. Carbon density coefficients were assigned to each land class following IPCC Tier 1 guidelines for temperate zones, with trees storing 120 megagrams of carbon per hectare aboveground, 30 belowground, 90 in soil, and 2 in dead organic matter, for a total of 242 megagrams per hectare. Cropland, by comparison, stores only 97 megagrams per hectare across all four pools combined, while grassland stores 102 and bare land stores just 4. The resulting carbon storage maps reveal the spatial distribution of carbon across the metropolitan landscape under each scenario, allowing researchers to calculate the net carbon change between the 2021 baseline and each 2050 projection. A first-order sensitivity analysis was conducted by simultaneously increasing and decreasing all carbon density coefficients by 10 percent to evaluate how uncertainty in these parameters propagates into scenario-level carbon estimates, confirming that the relative ordering of scenarios remained robust across the tested coefficient range.

The implications extend well beyond Prague’s administrative boundaries. Central European metropolitan regions remain significantly underrepresented in machine learning-based land-use and carbon modeling literature, which has concentrated heavily on rapidly urbanizing regions in Asia and Africa. The Prague application demonstrates that the framework can handle the particular complexity of European peri-urban landscapes, where historical land tenure patterns, EU agricultural subsidies, and post-socialist land restitution create transition dynamics that differ substantially from those in developing-world megacities. The Czech Republic’s legal commitment under the EU Climate Law to restore its land-sector carbon sink provides an immediate policy context for the scenario results, transforming abstract model outputs into concrete decision-support information for planners weighing ecological restoration against development pressure. The researchers emphasize that the scenarios represent exploratory planning pathways rather than deterministic forecasts, designed to illuminate the consequences of alternative policy choices rather than to predict a single inevitable future.

The broader significance of this work lies in its demonstration that explainable artificial intelligence can be operationalized within environmental modeling pipelines at metropolitan scale, producing results that are simultaneously spatially precise, mechanistically interpretable, and directly relevant to carbon accounting frameworks used in international climate policy. As cities worldwide grapple with the dual mandates of accommodating population growth and meeting net-zero emissions targets, tools that can quantify the carbon consequences of specific land-use decisions in advance become increasingly indispensable. The Prague framework, released as open-source Python code, offers a transferable template for other metropolitan regions seeking to understand how the landscapes they shape today will determine their climate resilience tomorrow.