Dense neural network outperforms other machine learning models for scaling-up lichen cover maps in Eastern Canada

Area of Science:

• Remote Sensing
• Ecology
• Machine Learning

Background:

• Accurate lichen mapping is crucial for caribou management and land conservation.
• Previous studies utilized random forest, dense neural network, and convolutional neural network models for lichen coverage estimation.
• A comparative analysis of these models' performance in lichen mapping was lacking.

Purpose of the Study:

• To evaluate and rank the performance of random forest, dense neural network, and convolutional neural network models for predicting lichen percent coverage using Sentinel-2 imagery.
• To identify the most suitable machine learning model for generating accurate regional lichen maps.

Main Methods:

• Trained three machine learning models (random forest, dense neural network, convolutional neural network) on 10-m resolution lichen coverage maps derived from drone surveys.
• Evaluated model performance using mean absolute error and R2 metrics on Sentinel-2 imagery from Québec and Labrador, Canada.
• Generated a regional lichen map using the best-performing model.

Main Results:

• The dense neural network model demonstrated superior accuracy with a mean absolute error of 5.2% and an R2 of 0.76.
• Random forest and convolutional neural network models showed comparable performance, with mean absolute errors of 5.5% and 5.3%, respectively.
• The dense neural network achieved a 5.9% performance gain over other models, despite higher computational requirements.

Conclusions:

• The dense neural network is the most suitable model for accurate lichen mapping from satellite imagery, offering significant performance improvements.
• The study advances the methodology for creating precise lichen maps essential for caribou conservation and sustainable land management.
• Model uncertainty was observed for land cover types not represented in the training data, highlighting areas for future research.