Dense Neural Network Outperforms Other Machine Learning Models for Scaling-up Lichen Cover Maps in Eastern Canada

Abstract

Lichen mapping is vital for caribou management plans and sustainable land conservation. Previous studies have used Random Forest, dense neural network, and convolutional neural network (CNN) models for mapping lichen coverage with remote sensing data. However, to date, it is not clear how these models rank in the performance of this task. In this study, these machine learning models were evaluated on their ability to predict lichen percent coverage in Sentinel-2 imagery covering Québec and Labrador, NL. The models were trained on 10-m resolution lichen coverage (%) maps created from 20 drone surveys collected in July 2019 and 2022. The maps were divided into quadrant blocks and then split into train, validation, and test datasets. The quadrant-blocking approach exposed the models to a variety of different landscapes and reduced spatial autocorrelation between the training sites. All three models performed similarly when evaluated on the test set. However, the dense neural network achieved a higher accuracy than the other two, with a reported Mean Absolute Error (MAE) of 5.2% and an R2 of 0.76. By comparison, the Random Forest model returned an MAE of 5.5% (R2: 0.74) and the CNN had an MAE of 5.3% (R2: 0.74). The models were also evaluated on their ability to predict lichen coverage (%) for larger quadrant blocks consisting of, on average, 400 Sentinel-2 pixels. The Random Forest and dense neural network had an R2 of 0.93, while the CNN had an R2 of 0.90. The MAE in this assessment for the dense neural network, Random Forest, and CNN were 2.1%, 2.3%, and 3.1% respectively. A regional lichen map was created using the dense neural network and a Sentinel-2 image mosaic. Model predictions have larger errors for land covers that the model was not exposed to in training, such as mines and deep lakes. While the dense neural network requires more computational effort to train than a Random Forest model, the 5.9% performance gain in the test pixel comparison and 9.1% performance gain in the quadrant block comparison renders it the most suitable for lichen mapping. This study represents progress toward determining the appropriate methodology for generating accurate lichen maps from satellite imagery for caribou conservation and sustainable land management.