IMPROVING THE BINARY CLASSIFICATION OF PEAT LOCALITIES FROM MULTI-SOURCE REMOTELY-SENSED DATA USING CNN

Abstract. Neural networks were explored to achieve a binary classification for determining land corresponding to peat for a study area in the boreal forest of northern Ontario, Canada. Environmental covariates were employed as predictors and obtained from multiple sources, which included multispectral imagery, LiDAR, SAR, and aeromagnetic data. A dense neural network (DNN), as well as a convolutional neural network (CNN), were each implemented. Logistic regression, support vector machine (SVM) and random forest (RF) approaches were also modelled. Neighboring pixels surrounding the soil sampling sites were incorporated as input into the CNN, that permitted training on additional information that was not exploited by other methods. Preliminary results indicate that a CNN can attain improved accuracies for peat classification, when compared against other approaches.