Managing nature-based tourism with AI: Evidence from the Peruvian Amazon

From tropical forests to coastal reserves, destinations built around natural heritage are facing mounting pressure to manage tourism more precisely. The pandemic disrupted visitor flows and revealed how reactive planning leaves ecosystems vulnerable to sudden shocks and long-term overuse, especially in regions with limited institutional capacity.

A recent scientific study, titled Artificial intelligence for biodiversity and tourism governance: predictive insights from multilayer perceptron models in Amazonia, explores how artificial intelligence (AI) could reshape tourism governance, using Amazonia as a case study. By analyzing two decades of tourism data from Peru's Bagua province, the research shows how neural network forecasting can help authorities anticipate visitor demand and design recovery strategies that reduce the risk of overtourism in fragile environments.

Pandemic shock exposes governance gaps in biodiversity-rich destinations

The collapse of global tourism during the COVID-19 pandemic was particularly damaging for regions that rely on domestic travel and small-scale tourism enterprises. In Peru, national tourism arrivals fell sharply in 2020, with the impact felt most strongly in provinces such as Bagua in the Amazonas region. Known for its waterfalls, rivers, caves, indigenous cultural practices, and forest ecosystems, Bagua depends heavily on visitor inflows to sustain local incomes and employment.

Bagua is a case that reflects broader challenges across biodiversity hotspots worldwide. Tourism activity in such regions often expands faster than governance capacity, leaving local authorities without reliable tools to forecast demand, manage visitor flows, or prevent environmental degradation. The pandemic intensified these weaknesses, creating what the authors describe as an urgent need for predictive and adaptive management approaches.

The research links demand prediction directly to governance outcomes. Accurate forecasts, the authors argue, can inform decisions about infrastructure planning, visitor management, marketing strategies, and conservation thresholds. Without such foresight, recovery efforts risk reproducing pre-pandemic patterns of unmanaged growth, placing additional stress on ecosystems already vulnerable to climate change and human activity.

The analysis focuses exclusively on domestic tourism, reflecting data realities in the region. International visitor numbers remained consistently low and showed little variation across the study period, making them unsuitable for meaningful modeling. Domestic travel, by contrast, displayed strong long-term growth punctuated by abrupt disruption during the pandemic, providing a rich dataset for testing predictive approaches.

Neural networks outperform conventional forecasting in post-crisis tourism planning

The study primarily discusses the application of multilayer perceptron neural networks, a class of artificial neural networks well suited to capturing nonlinear patterns in time-series data. The researchers designed a feedforward neural network architecture tailored to monthly visitor data, enabling the model to learn from historical trends and generate short-term forecasts for future tourism demand.

Two modeling scenarios were tested. The first incorporated the full dataset from 2003 to 2023, including the sharp collapse in visitor numbers during 2020. The second excluded 2020 from the training data and instead used pre-pandemic trends to estimate what visitor numbers might have looked like in the absence of COVID-19 disruptions. This approach treated the pandemic year as a structural outlier rather than a normal fluctuation.

The results show a clear advantage for the second scenario. Forecasts generated without the 2020 outlier achieved lower error rates, indicating greater accuracy and stability. According to the authors, this finding underscores the importance of data quality and contextual judgment in tourism forecasting. Extreme crisis events, if included without adjustment, can distort predictive models and reduce their usefulness for planning.

The study highlights how neural networks can support adaptive governance in volatile environments. By learning from long-term patterns while remaining flexible to new inputs, the models allow tourism managers to anticipate periods of high demand and plan interventions accordingly. This includes adjusting promotional activities, scheduling infrastructure maintenance, and preparing conservation measures for peak visitation months.

The forecasts produced by the model indicate a generally stable and gradually increasing trend in domestic tourism demand, with seasonal peaks that require careful management. Such insights, the authors argue, are critical for destinations where even modest increases in visitor numbers can have outsized ecological impacts if unmanaged.

AI-driven forecasting reframes tourism as a sustainability governance issue

According to the authors, predictive accuracy alone is not the ultimate goal. Instead, the value of neural network forecasting lies in its ability to inform policy decisions that balance economic recovery with environmental protection.

In the context of Bagua, this means using demand forecasts to regulate visitor flows at ecologically sensitive sites, align tourism development with carrying capacity limits, and avoid the risks of overtourism that have plagued other nature-based destinations. By anticipating demand rather than reacting to it, local authorities can shift from crisis management to proactive planning.

Accurate predictions allow communities to prepare for future shocks, whether from health crises, climate events, or economic downturns. This aligns with a growing body of research that views resilience not as a return to pre-crisis norms, but as the capacity to adapt governance structures to uncertainty.

The authors also outline the limitations of their approach. The analysis relies on aggregated monthly data, which constrains the ability to link visitor forecasts directly to ecological indicators such as habitat disturbance or species stress. Daily or site-specific data would be needed to establish more precise relationships between tourism pressure and biodiversity outcomes.

The study also does not include formal statistical tests of time-series stationarity or comparisons with alternative forecasting models such as ARIMA or hybrid deep learning approaches. While the neural network performs well within the study's design, the authors acknowledge that future research should benchmark AI models against conventional methods to strengthen policy confidence.