Comparative analysis of different algorithms for VAS station land cover classification with limited training points

• High-accuracy land cover classification using Sentinel-2 imagery at VAS, Spain. • ResCNN model outperforms others with 96% accuracy in land cover classification. • Effective use of limited training data for multi-temporal land cover mapping. • Summer data selection improves land cover classification accuracy. • Validation of machine/deep learning models for remote sensing at the VAS site. The Valencian Anchor Station (VAS) (Spain) is an outstanding site operating as a central location for calibrating and validating numerous remote sensing instruments and products. Hence, an accurate characterization of its land cover is required. This research conducts a land cover classification within the VAS station (10 × 10 km 2 ) and its surrounding area (30 × 30 km 2 ) for 2021 using multi-temporal imagery from Sentinel-2 Multispectral Instrument (MSI). Several aspects of land cover classification have been evaluated, including i) the feature selection, ii) the temporal resolution of time series (i.e., monthly, seasonal), iii) the performance of six Machine Learning algorithms (i.e., CART, GTB, k-NN, NB, RF, and SVM, alongside three deep learning models (FC-NN, MLP-ED, and ResCNN) and iv) the optimization of classifier tuning parameters. Furthermore, the study assesses the impact of reducing sample sizes on classifying similar areas, extending the classification to three buffers (1 km, 5 km and 10 km) without increasing reference data. ResCNN emerged as the best-performing model, yielding superior classification metrics (96 % overall accuracy and 95 % kappa score) in July, coinciding with the peak vineyard phenology. Producer’s and user’s accuracy values generally exceeded 90 % for most land cover classes, with some exceptions in more challenging categories such as artificial surfaces and non-irrigated arable land, which showed lower accuracies due to inter-class similarity. Overall, the findings underscore the robust performance of all models in land cover classification, demonstrating the feasibility of achieving high-quality classification with a robust methodology and limited training data.