Wood identification based on macroscopic images using deep and transfer learning approaches.

Area of Science:

• Forestry science
• Computer science
• Artificial intelligence

Background:

• Accurate forest type identification is crucial for ecological, economic, and social benefit assessment.
• Traditional expert observation is being augmented by technological advancements like artificial intelligence (AI).
• Deep learning offers potential for faster and more efficient forest species recognition.

Purpose of the Study:

• To adapt and evaluate various deep learning models for forest species recognition.
• To assess the performance of different deep network architectures using transfer learning.
• To identify a lightweight and efficient model for practical application in forestry.

Main Methods:

• Adaptation of pre-trained deep network models (RestNet18, GoogLeNet, VGG19, Inceptionv3, MobileNetv2, DenseNet201, InceptionResNetv2, EfficientNet, ShuffleNet) using transfer learning.
• Training and evaluation on a new forest species dataset.
• Performance assessment using metrics: accuracy, recall, precision, F1-score, specificity, and Matthews correlation coefficient.

Main Results:

• Deep network models demonstrated effectiveness in forest species recognition.
• ShuffleNet emerged as a lightweight model achieving high performance with reduced computational demands.
• Customized ShuffleNet achieved accuracy comparable to other models, highlighting its efficiency.

Conclusions:

• Deep learning models are powerful tools for advancing forest species recognition.
• The study contributes valuable insights for forest conservation and sustainable management.
• Efficient models like ShuffleNet can facilitate wider adoption of AI in forestry practices.