Hyperspectral tree crown classification using the multiple instance adaptive cosine estimator.

Area of Science:

• Remote Sensing
• Ecology
• Computer Vision

Background:

• Tree species classification from hyperspectral imagery is difficult due to spectral similarities and variability.
• Accurate species identification is crucial for ecological monitoring and forest management.

Purpose of the Study:

• To develop and evaluate a novel algorithm for robust tree species classification using hyperspectral data.
• To address challenges of spectral similarity and label uncertainty in hyperspectral image analysis.

Main Methods:

• The Multiple Instance Adaptive Cosine Estimator (MI-ACE) algorithm was employed.
• A one-vs-one classification strategy with majority voting was used for multi-class classification.
• The method does not require parameter tuning, enhancing ease of application.

Main Results:

• Species-level rank-1 classification accuracy reached 86.4% on testing data.
• Cross-entropy on testing data was 0.9395, indicating good model performance.
• The MI-ACE approach demonstrated effectiveness in classifying both common and rare tree species.

Conclusions:

• The MI-ACE algorithm provides an effective and user-friendly solution for tree species classification from hyperspectral imagery.
• The method shows promise for ecological applications requiring detailed forest inventory and monitoring.
• Accurate classification of rare species highlights the robustness of the proposed approach.