Assessment of smoothing and scale change effects on emergent aquatic vegetation in high spatial resolution image segmentation

Abstract. Identifying emergent aquatic vegetation (EAV) species is important to monitor environmental changes and support in decision-making. With the advent of uncrewed aerial vehicle (UAV), it has been possible to acquire high resolution images and assist in this task. At the same time, the high level of detail of these images can be considered noise for some segmentation algorithms and testing different smoothing and subsampling variations can be very relevant. Hence, the aim of this study is to analyse the performance of three segmentation algorithms (region growing, SLIC superpixel and watershed) to generate “homogeneous” regions in high resolution images, considering blur and scale change effects. To do this, images of a lake impacted by EAV were captured using a multispectral camera on board of UAV with 8 mm ground sample distance. After image processing, an orthomosaic was produced and three clippings were extracted from it to be segmented and tested empirically with variations of subsampling (1 cm, 1.6 cm, 2 cm, and 3 cm) and standard deviation smooth filter application (σ = 2, 4, and 8). The results showed that region growing and watershed algorithms are the most affected by high spatial resolution, and greatly benefits from the smoothing and subsampling applied, i.e., reducing the amount of detail, while superpixel algorithm created more consistent and uniform results, especially after smoothing, as evidenced by the quantitative evaluation based on segment entropy, characterized by kurtosis.