Evaluation of Classification Accuracy of Spectral Indices and Machine Learning Algorithms for Greenhouse Detection Using the Google Earth Engine Platform

Accurate mapping of greenhouse areas is critically important for enhancing agricultural productivity and mitigating environmental impacts. Recently, remote sensing technologies have emerged as powerful tools for detailed and accurate detection of greenhouse areas and land use. The main objective of the study, which is to evaluate the effectiveness of spectral indices and Machine Learning (ML) algorithms in detecting greenhouse areas. This study employs Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Trees (CART) ML algorithms to classify Harmonized Sentinel-2 MSI (Multispectral Instrument) satellite imagery using the Google Earth Engine (GEE) platform. In addition, indices such as the Normalized Difference Vegetation Index (NDVI), Plastic Greenhouse Index (PGI), Retrogressive Plastic Greenhouse Index (RPGI), Plastic Mulched Landcover Index (PMLI), and Greenhouse Vegetable Land Extraction Index (Vi) were calculated and incorporated as bands for classification. In the study, a total of 7 data sets were created using various ML algorithms and indices. The highest overall accuracy (OA) and kappa (Κ) values were obtained as 88.10% and 0.804, respectively, in the classification using the PGI and RF algorithm. To test the significance of the accuracy assessment, the McNemar test was applied. The most significant relationship was observed in comparisons using the PGI and RF classifier, where the calculated statistic was greater than the critical x² value(x²=3.84 at 95% confidence interval).