Introducing ARTMO's Machine-Learning Classification Algorithms Toolbox: Application to Plant-Type Detection in a...

Area of Science:

• Remote Sensing
• Machine Learning
• Ecology

Background:

• Accurate plant-type (PT) detection is crucial for sustainable land management, biodiversity, and ecosystem services.
• Sentinel-2 satellite imagery offers valuable data for mapping and monitoring PTs.
• A gap exists in user-friendly software for applying diverse machine-learning classification algorithms (MLCAs) to remote sensing data.

Purpose of the Study:

• To introduce a novel GUI software package, the MLCA toolbox within ARTMO, for systematic training, validation, and application of pixel-based MLCA models.
• To facilitate and automate the use of MLCAs for remote sensing image classification.
• To compare the performance of various MLCAs for PT mapping.

Main Methods:

• Developed and integrated a user-friendly GUI software package (MLCA toolbox) in Matlab.
• Applied 21 supervised MLCAs to Sentinel-2 imagery for mapping four PTs in Southwest Iran.
• Validated classifier performance using accuracy metrics, including overall accuracy (OA).

Main Results:

• Gaussian process classifier (GPC) achieved the highest overall accuracy (90%), significantly outperforming other MLCAs.
• Random forests (86% OA) and ensemble learning methods (83%, 82% OA) showed strong performance.
• Thirteen classifiers had OA between 70-80%, and four performed below 70%.

Conclusions:

• GPC demonstrates substantial potential for accurate land-cover classification and provides valuable probabilistic information.
• The MLCA toolbox effectively automates classifier evaluation and thematic mapping.
• Supervised classifier performance is dependent on training data, necessitating comprehensive assessment for specific applications.