Land use – whether it is forests, agricultural land or urban areas – has a fundamental impact on the Earth's carbon balance and thus on climate changes. CO2 emissions from land use change (ELUC) are one of the main factors contributing to climate change. While land-use planning alone is not enough to achieve net zero emissions, it can help offset emissions from sectors that are harder to decarbonise. Making the right decisions about how land is used can mitigate carbon emissions from its change.

Land-use planning requires decision-making that minimizes carbon emissions while maintaining satisfactory levels of land benefits, also known as ecosystem services, such as food supply. Optimizing this process is complex, and previous methods have had their limitations.

A new approach using neuroevolution

Recent research, conducted as part of the Resilience project, a non-profit ITU project, aims to provide decision-makers with a tool to help them understand the long-term impact of their land-use decisions on CO2 fluxes and to suggest how to optimize these decisions. The tool is based on a method called Evolutionary Surrogate-assisted Prescription (ESP).

The basic idea of ESP is to first use historical data to learn substitution model (predictor), which predicts how land-use decisions in different contexts affect carbon emissions. Substitution model learns from historical land use change data and simulations of associated emissions and carbon removals (specifically from LUH2 data and the BLUE model). This model can effectively evaluate the different options available to decision makers.

The context for decision-making includes the geographic grid cell (latitude, longitude, area), time point (year), and current land use (percentage of area for each use type). Actions they represent decisions on how to change land use, with restrictions – primary land and urban areas cannot be changed. Results include ELUC (Land use change emissions in tC/ha) and land amendment costs (percentage of area changed).

The BLUE model provides ELUC estimates based on land use changes from LUH2. Although BLUE is accurate, it is computationally intensive, and therefore its simulations were used to train the efficient substitution model. LUH2 provides land use data for various types, such as primary and secondary forests, non-forest vegetation, urban areas, crops, and pastures. However, there are limitations in the data, such as coarse resolution and the grouping of all crop types into one category.

Policy prescription using evolutionary search

Subsequently, effective land use policies emerge through evolutionary search – prescribing models (prescriptors)These models suggest actions (land use changes) for a given context with the aim of optimizing outcomes. Since the optimal actions are not known in advance, the performance of each candidate is measured using a substitution model.

The result of this process is Pareto front solutions that represent a trade-off between carbon emission reduction (ELUC) and the extent of land use change. Each point on the Pareto front is the optimal policy for a given trade-off. For more accurate predictions, a global neural network (NeuralNet) was selected as a substitution model, which demonstrated a better ability to extrapolate to large changes and capture nonlinearities in the data compared to linear regression (LinReg) and random forest (RF).

Key findings and practical applications

Evolutionary prescriptors outperformed heuristics, especially for medium-sized changes. Evolutionary prescriptors were found to have a strategy of making only a few large changes in the cells where they have the greatest impact on emissions. Although evolutionary prescriptors almost never outperform heuristics in any individual cell, averaging across all cells outperforms heuristics overall. These models particularly recommend converting cropland to secondary forest. They also learned to target larger changes in tropical, temperate, and continental regions, while changing drylands less.

The Resilience project showed that it is possible to include other constraints, such as minimizing cropland loss. Evolution was able to find solutions that effectively trade off between all three objectives (ELUC, soil change, cropland change), making the tool more practical. An interesting finding was that the evolutionary prescription model outperformed the heuristics on all three objectives by changing less cropland, but compensating for this by reducing pasture.

The importance of human expertise

Incorporating human expertise has proven to be very useful. Injecting two “seeded” models into the initial population (one for no change, the other for maximum change to secondary forest) significantly improved performance and allowed evolution to find better solutions, especially for scenarios with small changes. These seed models contributed to the ancestors of almost every prescription model on the final Pareto front.

A tool for decision-makers

This project provides proof of concept a tool for decision-makers that can help them optimize land-use decisions. An interactive demo is available online where users can view proposed land-use changes for specific sites, see predicted outcomes (emission reductions, area change), and experiment with alternative trade-offs and adjustments.

Future work includes using more detailed data, improving the accuracy of the prediction model using ensemble, estimating the uncertainty of predictions, developing transparent rules instead of neural networks, adding additional objectives (e.g. water quality, food production), and planning for longer periods.

Ultimately, machine learning can play an important role in empowering decision-makers to act on climate change issues through effective land-use planning. Spring

The study is published in the journal Environmental Data Science .

Glossary of key terms

• Terrestrial carbon balance: The net flow of carbon between the atmosphere and terrestrial ecosystems (soil and vegetation). Influenced by processes such as photosynthesis (takes up carbon) and decomposition/fires/land use changes (releases carbon).
• Emissions from Land-Use Change (ELUC): CO2 emissions (or removals, if negative) directly attributed to changes in land use patterns, such as deforestation for agriculture or reforestation.
• Negative Emission Technology (NET): Processes or technologies that remove CO2 from the atmosphere. An example in the context of land use is afforestation.
• Evolutionary Optimization: Optimization methods inspired by biological evolution, such as genetic algorithms, use processes such as mutation and crossover to gradually improve a population of candidate solutions.
• Neural Networks: Computational models inspired by the structure and function of biological neural networks. They are used in machine learning for prediction and decision-making.
• Surrogate Model: A simpler, less computationally intensive model that mimics the behavior of a more complex system. It is used to quickly evaluate candidates in optimization searches when the original system's evaluation is too slow.
• Evolutionary Surrogate-assisted Prescription (ESP): A machine learning method that combines evolutionary optimization with a surrogate model. First, a predictor (surrogate model) is learned from historical data, and then a prescriptor (decision policy) is discovered using this predictor through evolutionary optimization.
• Prescriptor: A model or policy that recommends actions (decisions) in a given context to optimize certain outcomes.
• Predictor: A model that predicts the results of given actions in a given context based on historical data. It serves as a surrogate model in the ESP method.
• Pareto front: In multi-objective optimization, it is a set of non-dominated solutions. A solution is non-dominated if it cannot be improved in one objective without worsening in at least one other objective. It represents the set of best compromises.
• LUH2 (Land-Use Harmonization dataset): A global dataset providing annual information on land use fractions, land use transitions, and key agricultural management information in a gridded format.
• BLUE (Bookkeeping of Land-Use Emissions model): A semi-industrial statistical model that calculates CO2 fluxes based on changes in biomass and soil carbon content caused by land use changes.
• RHEA (Realizing Human Expertise through AI): A framework for incorporating expert human knowledge into artificial intelligence processes, for example by distilling expertly designed solutions into an initial population of evolutionary algorithms.
• Hypervolume: A metric used in multi-objective optimization to evaluate the performance of a set of solutions (e.g., a Pareto front). It measures the volume of the objective space that is dominated by the set of solutions relative to a reference point.