The team’s central tool was ensemble species distribution modeling, implemented in the widely used biomod2 platform. Rather than relying on a single algorithm, the researchers calibrated multiple machine-learning models against 403 quality-controlled presence records of cacao in the region, each one verified to reduce the errors that plague many mapping exercises. The predictor variables drew on three distinct data streams: bioclimatic layers describing temperature and precipitation regimes, topographic information such as elevation and slope, and edaphic data capturing soil properties. On top of the statistical machinery, the authors applied spatial exclusion criteria that removed biophysically unsuitable zones, legal conservation constraints that protect protected areas, and regulatory filters tied to the European Union Deforestation-Free Products Regulation, known as EUDR, which will increasingly govern whether Peruvian cacao can enter European markets.

The modeling pipeline was deliberately strict about quality. Each candidate algorithm was evaluated under cross-validation, and only models exceeding demanding thresholds were retained: an area under the receiver operating characteristic curve, or AUC, greater than 0.9, and a true skill statistic, or TSS, greater than 0.7. These metrics matter because species distribution models can look impressive while hiding serious weaknesses. AUC above 0.9 indicates near-excellent discrimination between suitable and unsuitable locations, while TSS above 0.7 signals strong agreement between predicted and observed distributions after correcting for the prevalence of the species. Among the individual algorithms, decision tree-based approaches performed best, with Random Forest and XGBoost standing out as the strongest single predictors. Random Forest builds many decision trees on random subsets of the data and averages their votes, while XGBoost builds trees sequentially, each one correcting the errors of the last. Their dominance suggests that the relationships between cacao suitability and environmental gradients are nonlinear and interactive, precisely the kind of structure tree ensembles capture well.

Yet the study’s most consequential finding came not from any single model but from the consensus. When the researchers merged the retained models into an ensemble, the resulting suitability map showed greater spatial stability and overall accuracy than any individual algorithm. This is a recurring lesson in predictive ecology: individual models disagree at the edges of their training data, but a well-constructed consensus dampens those idiosyncrasies. For policymakers, that stability is what transforms a scientific map into a planning instrument. A boundary drawn on a consensus surface is far less likely to shift dramatically when new data arrive than one drawn on a single model’s output.

So where does the chocolate tree want to grow in Amazonas? Under current climatic conditions, approximately 14.8 percent of the regional territory shows high suitability for cacao. That footprint is not spread evenly across the landscape. It concentrates in inter-Andean valleys and in low- to mid-elevation zones, where temperatures remain warm and relatively stable and moisture is adequate through the year. The single most important environmental predictor, according to the model’s variable importance analysis, was thermal seasonality, the degree of variation in temperature across the year. Cacao, a tropical understory species, tolerates little in the way of cold stress or wide temperature swings, so it makes intuitive sense that the steadiness of the thermal regime, more than total rainfall or soil chemistry, defines the crop’s climatic envelope in this part of the Andes-Amazon transition.

Here, however, the study delivers its sharpest and most policy-relevant twist. The 14.8 percent figure describes climatic suitability, but climate is not permission. When the researchers overlaid territorial and regulatory constraints, the effectively available area shrank substantially. Protected conservation areas came off the map, as did zones excluded on biophysical grounds. Most striking was the impact of the EUDR scenario: because the European regulation prohibits products grown on land deforested after a defined cutoff date, large portions of the climatically suitable territory become ineligible for export-oriented cacao under the new rules. The gap between where cacao could grow and where cacao may legally be grown is, in other words, wide, and it is widening as environmental due-diligence requirements harden in major consumer markets.

This regulatory squeeze does not leave the region empty-handed. The analysis identified high-priority degraded areas that emerge as focal opportunities for productive restoration through agroforestry systems. Agroforestry, the practice of growing cacao beneath and alongside native trees, offers a way to rebuild canopy cover, restore ecological function, and produce a marketable crop simultaneously. Because these degraded lands are not subject to the same deforestation concerns as intact forest frontiers, they represent the sweet spot where farmer livelihoods, restoration goals, and EUDR compliance converge. In effect, the study reframes cacao expansion not as a question of finding new land to clear, but of finding old, damaged land to heal.

The third layer of the analysis is the one least often seen in suitability mapping: plant health. The researchers conducted a phytosanitary validation covering 2021 through 2024, comparing the modeled suitability surface against field records of the prevalence and incidence of major cacao pathogens and pests. The result was a high spatial concordance that carries an uncomfortable implication. The areas of greatest climatic suitability for cacao are also the areas where the crop’s worst enemies thrive, particularly the fungi Moniliophthora roreri, the agent of frosty pod rot, and Moniliophthora perniciosa, which causes witches’ broom disease. Both pathogens have devastated cacao production across Latin America for decades, and both flourish under the same warm, humid conditions that cacao itself prefers. The finding confirms that an optimal climatic niche for cultivation is simultaneously an optimal niche for biotic pressure, meaning that any expansion strategy must build disease management, resistant varieties, and integrated pest control into its foundations rather than treating them as afterthoughts.

Taken together, the study offers something rarer than another suitability map: an integrative framework in which climatic modeling, territorial management, and phytosanitary risk are assessed in a single pipeline. The authors position the work as a scientific input for public policy, land-use planning, and sustainable cacao production strategies across the Amazon. The practical chain of reasoning is clear. Climate models identify where the crop can grow; legal and conservation filters identify where it may grow; degraded-area analysis identifies where it should grow; and phytosanitary data identify where it will need the most protection once it does. Each layer disciplines the next, and the final product is a map of opportunity that is honest about its own limits.

For Peru, whose cacao has earned a reputation among fine-flavor chocolate makers, and for tropical agricultural science more broadly, the message is twofold. Ensemble machine-learning models, fed with multisource data and validated against ground truth, can now deliver planning-grade precision at the regional scale. But the era in which suitability could be defined by climate alone is over. Trade regulations written in Brussels, conservation commitments made in Lima, and fungi evolving in Amazonian pods all shape the same map. The future of chocolate in its birthplace will be decided not by where the tree can survive, but by where the full weight of science, law, and biology allows it to flourish.

Subject of Research: Ensemble predictive modeling of climatic, regulatory, and phytosanitary suitability for cacao cultivation in Amazonas, Peru

Article Title: Geospatial suitability of cacao in Amazonas (Peru): An ensemble predictive modeling approach (BIOMOD2) integrating multisource data, territorial management, and plant health

Article References: Zabaleta-Santisteban, J. A., Medina-Medina, A. J., Tuesta-Trauco, K. M., Rivera-Fernandez, A. S., Silva-Melendez, T. B., Grandez-Alberca, M. A., Puscan-Rojas, J., Salas López, R., Oliva-Cruz, M., Cotrina-Sanchez, A., Rojas-Briceño, N. B., Silva-López, J. O., Gómez-Fernández, D., Leiva-Espinoza, S. T., Huaman-Pilco, A. F., & Barboza, E. (2026). Geospatial suitability of cacao in Amazonas (Peru): An ensemble predictive modeling approach (BIOMOD2) integrating multisource data, territorial management, and plant health. PLOS One, 21(10), e0360376. https://doi.org/10.1371/journal.pone.0360376

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360376

Keywords: cacao, Amazonas, Peru, BIOMOD2, ensemble modeling, species distribution models, EUDR, deforestation, agroforestry, Moniliophthora roreri, witches' broom, land-use planning