Seasonal land use and land cover mapping in South American agricultural watersheds using multisource remote sensing: the case of cuenca Laguna Merín, Uruguay

Abstract

Recent advancements in Earth Observation sensors, improved accessibility to imagery and the development of corresponding processing tools have significantly empowered researchers to extract insights from Multisource Remote Sensing. This study aims to use these technologies for mapping summer and winter Land Use/Land Cover features in Cuenca de la Laguna Merín, Uruguay, while comparing the performance of Random Forests, Support Vector Machines, and Gradient-Boosting Tree classifiers. The materials include Sentinel-2, Sentinel-1 and Shuttle Radar Topography Mission imagery, Google Earth Engine, training and validation datasets and quoted classifiers. The methods involve creating a multisource database, conducting feature importance analysis, developing models, supervised classification and performing accuracy assessments. Results indicate a low significance of microwave inputs relative to optical features. Short-wave infrared bands and transformations such as the Normalised Vegetation Index, Land Surface Water Index and Enhanced Vegetation Index demonstrate the highest importance. Accuracy assessments indicate that performance in mapping various classes is optimal, particularly for rice paddies, which play a vital role in the country’s economy and highlight significant environmental concerns. However, challenges persist in reducing confusion between classes, particularly regarding natural vegetation features versus seasonally flooded vegetation, as well as post-agricultural fields/bare land and herbaceous areas. Random Forests and Gradient-Boosting Trees exhibited superior performance compared to Support Vector Machines. Future research should explore approaches such as Deep Learning and pixel-based and object-based classification integration to address the identified challenges. These initiatives should consider various data combinations, including additional indices and texture metrics derived from the Grey-Level Co-Occurrence Matrix.