Comprehensive study of machine learning classification methods applied to UAV-based management of intertidal ecosystems

Intertidal marine ecosystems are ecologically significant, with macroalgae playing a crucial role in carbon sequestration, habitat formation, and marine productivity. Traditional monitoring and management methods are labour-intensive and limited in scale, making remote sensing a valuable alternative. Unmanned Aerial Vehicles equipped with RGB cameras offer a cost-effective solution, but accurately classifying macroalgae remains challenging due to species similarity. Image textures have already been successfully used to classify RGB images. However, different terrains and classes may benefit from classifying using different image textures. This study presents a workflow for macroalgae classification using RGB imagery and Haralick Textural Features (HTFs). To assess the most important features for our particular task, feature selection techniques were used to reduce the number of HTFs from 16 to 8, optimising classification accuracy. We also evaluated four classifiers (Support Vector Machine, Random Forest, Multi-layer Perceptron (MLP), and XGBoost) across RGB bands. Additionally, we analysed the impact of superpixel segmentation using Simple Linear Iterative Clustering (SLIC) and Superpixels Extracted via Energy-Driven Sampling (SEEDs) on classification performance. Superpixel segmentation was used to group pixels with similar characteristics for calculating Haralick textural features. This method enables classifiers to predict the classes of significant image regions, particularly in complex natural environments where the boundaries of macroalgae are not always clearly defined. There was no significant difference between colour datasets, with MLP achieving the highest accuracy (0.75), followed closely by XGBoost (0.74). Both SLIC and SEEDS performed better between 250 and 350 initial superpixels, and SEEDS outperformed SLIC with a slightly higher accuracy (0.77 vs. 0.75) when using similar parameter configurations. SEEDS sped up the creation of superpixels and retained most of the initial superpixels. The workflow has been validated using a case study in Galicia (NW Iberian Peninsula), demonstrating its potential for scalable, automated management of natural environments. • Assessment of classification methods in large-scale remote sensing with UAVs. • Feature reduction of Haralick textural features. • Comparison of superpixel creation for automatic classification of orthoimages. • Comparison of machine learning models for macroalgae classification. • Case study in two intertidal areas of the NE Atlantic coast.