Semantic segmentation of macrophytes in multispectral images: a comparative analysis between different backbones.

Abstract

In general, the uncontrolled growth of macrophyte banks is a big problem in reservoirs used for electricity production, affecting operational efficiency, water quality and ecological health. Therefore, effective management strategies such as mechanical removal, herbicide treatments and biological control methods are often necessary to mitigate these problems and maintain the functionality of power generation facilities. However, these strategies require constant monitoring of the growth of this type of vegetation. In this context, and with the advent of new remote sensing technologies designed to map the entire world using multiple spectral bands, there has been an interest in addressing the problem of segmenting aquatic vegetation through machine learning models or the use of spectral indices. This work aims at a more generalized segmentation through the consumption of spectral bands and spectral indices, making them primary information for training deep learning models, achieving Recall metrics and IoU values of 66% and 59% respectively for our best case scenario.