Finding suitable locations for in-stream wetland creation/restoration: comparing suitability analysis with machine learning approach

Wetlands are critical nature-based solutions (NbS) for addressing environmental challenges, playing an important role in sediment and nutrient retention, agricultural runoff mitigation, and carbon storage, contributing to climate change adaptation. However, agricultural intensification and land conversion have drastically reduced wetland coverage globally, necessitating the precise selection of sites for restoration/creation. Depending on fieldwork and expert judgment, traditional methods often struggle to scale effectively, highlighting the need for advanced geospatial techniques.

This study compares two approaches for in-stream wetland site selection, the Analytic Hierarchy Process (AHP) and the machine learning Random Forest (RF) algorithm, within the diverse hydrological landscape of Estonia. Both methods utilized environmental variables, including slope, topographic wetness index (TWI), flow accumulation, soil organic carbon (SOC), and clay content, to evaluate their influence on hydrological and soil conditions critical for determining suitable sites for in-stream wetland creation and restoration. These variables were selected for their ability to capture the key factors that drive wetland formation and functionality. Geospatial datasets, including local and global environmental variables, were processed at 10- and 50-meter resolutions to analyze how spatial resolution influences model performance, providing high-detail insights for localized assessments and broader, regional-scale perspectives.

The AHP framework integrates expert knowledge to prioritize variables, while the RF algorithm provides a data-driven, scalable alternative. The RF model was trained using data from existing wetlands, which were identified based on geospatial datasets and intersected with stream networks, channels, ditches, and rivers to focus on areas directly connected to water flow. Training points were randomly sampled within these wetlands to represent suitable areas. In contrast, points from non-wetland areas, such as forests, shrublands, grasslands, and arable land, were sampled to represent unsuitable areas. This approach ensured that the training data captured the variability of environmental conditions influencing wetland suitability

Validation was conducted using a historical map to evaluate model accuracy and reliability across varying scales and data conditions. Results indicate that the RF algorithm outperformed AHP in predictive performance, achieving an accuracy of approximately 0.8 at broader resolutions and slightly lower accuracy at finer resolutions. This underscores the influence of spatial resolution on model performance. However, AHP underscored the importance of structured decision-making and stakeholder input, ensuring practical applicability. This research advances the integration of NbS into wetland planning, bridging traditional expertise-driven methods and machine learning innovations to enhance precision, scalability, and cost-effectiveness.