How MoFuSS works?

MoFuSS is a spatio-temporal model that operates on maps and evolves over time. It starts in the past (e.g., 2010), moves through the present during the calibration-validation period, and extends into the future (e.g., 2050), which we refer to as the prospective simulation period. One common point of confusion is that the past data used in MoFuSS, such as maps depicting aboveground biomass (AGB) in 2010, may have been created using the latest methodologies, even as recently as yesterday. The key takeaway: the year the data represents isn't tied to when it was produced. MoFuSS always uses the most up-to-date and robust datasets available. Starting in the past is necessary to assess whether simulations from the past to the present match real-world outcomes.

In its initial phase, MoFuSS focuses on calibrating, validating, and forecasting land use/land cover changes (LULCC), such as transitions from forest to cropland or cropland to urban areas. These changes are not driven by woodfuel demand but are instead modeled based on historical patterns and relationships with spatial variables like road proximity, slope, distance to previously deforested areas, and property rights, among others. The model assigns probabilities to transitions and simulates these changes, generating a new land cover map each year from 2010 to 2050. This stage can be also used in REDD+ projects as well, as it gives a numerical estimate of the chances a given area will suffer change (e.g. deforestation) within a time period (e.g. 2010-2050). My understanding is that Peru is using this approach, not 100% sure.

Woodfuel supply and demand patterns evolve within this dynamic landscape. This is crucial for two reasons:

1. Different land use categories provide varying amounts of wood over time (not a fixed amount—each category responds uniquely to harvesting).
2. When deforestation occurs, some of the cleared wood is allocated to woodfuel use.

Land cover classes are categorized as either degradable or non-degradable. Degradable classes, such as natural woodlands, degrade if wood is harvested faster than it regrows. This is known as non-renewable biomass in carbon credit methodologies. In MoFuSS, this isn’t a simple harvest-minus-MAI calculation—vegetation grows at different rates, and sporadic or frequent harvest events influence the curve. If the AGB stock shrinks over the simulation period, degradation and emissions occur. Non-degradable lands, on the other hand, include sources like pruning of tree outside forests, collecting deadwood or crop residues, utilizing sawmills and plantations residues, among many others. Globally, the model distinguishes between 690 land use categories when using MODIS data and 789 categories with Copernicus as the base LUC map.

How does MoFuSS connect where fuelwood and charcoal are consumed to where they are sourced?

MoFuSS connects where fuelwood and charcoal are consumed to where they are sourced using a combination of population data and spatial algorithms. Consumption data comes from WorldPop datasets combined with UN and WHO scenarios from 2000 to 2050. To model sourcing, MoFuSS uses a custom C++ code that runs on two high-performance computing (HPC) clusters, evaluating the time cost of gathering woodfuel for each pixel from the surrounding landscape. This surrounding landscape can extend up to 12 hours away. For instance, an urban area in southern Malawi may draw charcoal from Mozambique, depending on the road network, international borders, and other spatial features affecting accessibility. HPC clusters are essential due to the trillions of parallel computations involved. While we hope to eventually train AI models to predict charcoal producing hotspots, current data is insufficient and some funded-research would be needed (Hi there funders! My name is Adrian ??).

The geospatial analysis described so far is conducted multiple times with slight variations in input parameters, a process known as Monte Carlo. The final results synthesize all the Monte Carlo realizations, each one with minor differences in input data. The current version of MoFuSS also accepts up to three input maps for land cover and AGB, producing results that account for both parameter variations and differences in data sources.

The model will finally compare how much AGB were lost and gain between a Business as Usual and a demand reduction scenario. The difference in these pixel-based losses and gains between both scenarios represent the estimated avoided emissions by the intervention, without the need of fNRB or aggregating by admin units. As I mentioned in a previous post: “…while all the rest keep varying everywhere and all the time, but consistently across both scenarios.”?

There are several other uses of MoFuSS, like highlighting where interventions will have the highest impacts in NRB reductions. Another source of confusion is that depending on where each pixel is, a reduction in demand might have a high or low impact on NRB. We can’t tell a priori, and this is why we built MoFuSS in the first place.

Finally, the global 1km model, developed between March and August of 2023, and improved during the first semester of 2024, comes with challenges related to its coarse resolution mostly. Comparing simulations to observed dynamics at this coarse scale is complex. So far, we've found that the model generally aligns with reality but sometimes greatly overestimates impacts (including fNRB) in some countries. The reasons for this were outlined in a previous document: https://www.mofuss.unam.mx/mofuss-ds/files/MoFuSS_visualization_tool_EXPANDED.pdf

As an example, the animation below (one of many outputs from running MoFuSS at 100m resolution) depicts two villages (black dots) within a forest dominated by “degradable” land in the Yucatán Peninsula, Mexico. To simulate such intense degradation, we had to increase the per capita consumption of woodfuel by 100x. MoFuSS employs a cellular automaton to model harvest events at local scales, with a tunable degree of stochasticity. In this case, the settings were adjusted towards a more deterministic approach, reducing randomness to maximize degradation. This is why the incursions into distant forests are minimal in this example.