Assessing feature extraction, selection, and classification combinations for crop mapping using Sentinel-2 time series: A case study in northern Italy

Rural areas need constant monitoring to ensure sustainable farming and respond to environmental and climatic impacts. Over the last few decades, remote sensing data have been extensively used in agricultural monitoring, allowing cost-effective and efficient crop management. Selecting the suitable data combinations for crop mapping while reducing dimensionality and redundancy to speed up processing remains a challenge. This study address the challenges by testing and assessing the efficiency of various combinations of feature extraction, feature selection, and feature classification methods. We used Sentinel-2 time series data, which focused on spectral features and derived vegetation indices. Particularly, the red-edge indices which are critical for crop discrimination. To select the optimal data for classifiers, we have tested two feature selection techniques: Random Forest and Principal Component Analysis, for both spectral bands and vegetational indices. Then, we have employed three machine learning algorithms: Extreme Gradient Boost (XGB), Random Forest (RF), and Support Vector Machine (SVM) along with one deep learning approach, Pixel-Set Encoders and Temporal Self-Attention (PSETAE), to evaluate the datasets. The results suggest that the most effective feature set is the Sentinel-2 spectral bands selected by the Random Forest feature selection method. The results obtained from the previous step achieve the highest overall accuracy with the XGB classifier and outperform the RF, SVM, and PSETAE classifiers. Further, quantitative analysis of overall classification accuracies showed that Random Forest is the second-best performing classifier, and PSETAE classifier produced the lowest results for all data models.