Remote sensing tree classification with a multilayer perceptron.

Area of Science:

• Ecology
• Computer Science
• Remote Sensing

Background:

• Remote tree classification is crucial for biodiversity and ecological sampling.
• Developing advanced informatics methods is needed to improve tree species identification.
• A community-based competition was initiated to foster innovation in tree classification.

Purpose of the Study:

• To evaluate and compare the performance of three machine learning algorithms for pixel-level tree species and genus classification.
• To assess the effectiveness of hyperspectral and LiDAR data in tree classification.
• To determine the potential of the multilayer perceptron algorithm for remote tree identification.

Main Methods:

• Pixel-level classification of tree species and genus was performed using hyperspectral and LiDAR observations.
• Three algorithms were compared: support vector machines, random forests, and multilayer perceptron.
• An ensemble approach averaging pixel-level outputs was used for crown-level species classification.

Main Results:

• The multilayer perceptron algorithm achieved high accuracy at the pixel level (92.7% for species, 95.9% for genus) on training data.
• Multilayer perceptron outperformed support vector machines and random forests in pixel-level classification.
• The ensemble MLP approach achieved an average crown-level accuracy of 68.8% for nine species on the test set.

Conclusions:

• The multilayer perceptron algorithm demonstrates significant promise for tree species classification using hyperspectral and LiDAR data.
• Neural network-based algorithms, like MLP, show superior performance compared to other classification methods in machine vision tasks.
• The findings support the use of MLP for accelerating remote tree classification and ecological sampling efforts.