Integrating K-Fold Cross-Validation with Advanced Classification Techniques for Shrub Coverage Mapping in Fire-Prone Landscapes using Sentinel-2 Imagery

This study employs state of the art classification techniques, Support Vector Machine (SVM), Artificial Neural Network (ANN), and Random Forest (RF), to map shrub coverage in a fire-prone landscape characterized by fine-grain heterogeneous land cover and a Mediterranean climate. Sentinel-2 satellite imagery from 2020 to 2022 were labeled at the pixel level, after intersecting a high-resolution (20 cm pixel) land cover map of 2020 with the Sentinel pixel grid. The goal was to produce an algorithm able to predict shrub cover percentage at the spatial resolution of 10 m x 10 m just informed by satellite imagery. The methodology incorporates K-Fold Cross-Validation (K-fold CV) to strengthen the model structure’s generalization ability to generalize to unseen data by minimizing bias towards specific area. The RF model, identified as the most effective, achieved 73% Precision, 72% Recall, and 72% F-Score for Category 1 (0%-10% shrub coverage), 62% Precision, 70% Recall, and 65% F-Score for Category 2 (>10%-50% shrub coverage), and 89% Precision, 79% Recall, and 84% F-Score for Category 3 (>50% shrub coverage). Our method thus demonstrates significant results for shrub coverage in Category 3, crucial for fire hazard assessments and implementing prevention measures at the landscape scale.