Detection of Hazelnut Orchards with Sentinel-2 imagery and machine learning classification algorithms

Abstract. Hazelnut (Corylus avellana L.) is an economically important crop in Turkey, with Sakarya being a major cultivation region. Effective large-scale monitoring of hazelnut orchards can be achieved using remote sensing and machine learning techniques. In this study, field surveys were conducted in approximately 150 hazelnut orchards in Sakarya to provide training data. Multi-temporal Sentinel-2 imagery from six acquisition dates capturing key phenological stages was stacked for the classification of hazelnut orchards and other land use/land cover (LULC) types. Vegetation indices including NDVI, AVI, SAVI, and EVI were applied to enhance class separability. Supervised classification was performed using Random Forest (RF) and Extreme Gradient Boosting (XGBoost) algorithms, with hyperparameters optimized via RandomizedSearchCV and cross-validation. Both models achieved high performance in detecting hazelnut orchards; however, RF yielded better overall results in quantitative metrics and visual assessments. These findings demonstrate that integrating multi-temporal Sentinel-2 data, vegetation indices, and machine learning enables accurate large-scale mapping of hazelnut orchards in Sakarya.