Geo-Big Data approaches for land use and land cover mapping: insights from the 2021–2022 agricultural year in Uruguay

Abstract

Land use and land cover mapping is a key tool for understanding how territorial configuration influences biotic, abiotic, and anthropic components. In this regard, Geo Big Data technologies enable the agile and accurate generation of cartographic products. This study proposes two solutions for mapping land use and land cover in Uruguay for the 2021–2022 agricultural year. The inputs include Sentinel-1 and Sentinel-2 imagery, Google Earth Engine, GEEMAP, Scikit-learn, and the Random Forest and Support Vector Machines algorithms. The methodology highlights the creation of a multitemporal dataset, hyperparameters tuning, and supervised classification. As a result, two maps were generated: S1S2RF_uy and S1S2SVM_uy. Both products exhibited elevated levels of accuracy, although S1S2RF_uy performed slightly better, with an overall accuracy of 83 % and a kappa coefficient of 0.81, compared to 81 % and 0.78 for S1S2SVM_uy. At the class level, Random Forest showed a greater ability to classify agricultural covers, while Support Vector Machines were more effective in identifying artificial surfaces such as urban fabric. The findings confirm that hyperparameter tuning is essential for optimal classifier performance. Based on the reported accuracy statistics, it is also demonstrated that freely accessible Geo Big Data resources are well-suited for the efficient production of national-scale cartography at medium-to-high spatial resolution. Future research should prioritise regional focus and extend timeframes beyond the traditional agricultural year.