Performance analysis of machine learning techniques to identify aquatic vegetation with Sentinel-2 bands Análisis del desempeño de técnicas de aprendizaje automático para identificar vegetación acuática con bandas de Sentinel-2

VENTURINI, VIRGINIA; MARCHETTI, ZULEICA Y.; WALKER, ELIZABETH; FAGIOLI, GIANFRANCO

ECOLOGÍA AUSTRAL

ASOCIACIÓN ARGENTINA DE ECOLOGÍA

Año: 2023 vol. 33 p. 743 - 756

0327-5477

Natural disasters, such as river overflows, extreme droughts, natural forest fires are more frequently observed in Argentina. Faced with these catastrophes, efficient management is essential to make quick decisions to minimize damage, which is a latent concern in local and regional governments and in the scientific community. In Argentina, the Paraná River basin represents a strategic resource in itself, as it encompasses the greatest fluvial, ecological wealth and large urban centers. However, the extreme events that characterize the dynamics of the wetlands affect the urban centers located near them. The presence of aquatic vegetation (free or rooted) masks the flooded areas, hiding the first signs of flooding, making the monitoring and rapid detection of these areas difficult. In this work, optical satellite images and machine learning models were used to classify the different land covers in wetlands of the Paraná river system. The focus was on environments where free water and aquatic marsh vegetation coexist, characteristic of the metropolitan region of the city of Santa Fe, and considering the technical limitations of decision-making agencies. Therefore, the Sentinel-2 (S2) mission images were used to train and evaluate different machine learning algorithms. All bands of S2 images were used, unifying the spatial resolution to 10 m. The results indicated that the coastal aerosol bands (B1) and two mid-infrared bands (B11 and B12) provide the most information for the identification of the samples. Moreover, the random forest method showed the best performance for the aquatic vegetation class, which was of primary interest for this work.