Classification of damaged vegetation areas using convolutional neural network over unlabelled sentinel-2 images

Abstract

Several researches have been made in recent years which use Convolutional Neural Networks (CNN) to classify multispectral images. However, a scarcity of labelled satellite images makes it difficult to classify damaged forest areas using pre-trained CNNs. If any new destruction hits the forests, then there will be a lack of labelled satellite data and manually labeling the satellite images will be inefficient and will ultimately fail the purpose of rapid reforestation. Hence, this research aims to label the unlabelled Sentinel-2 satellite images by using a pre-trained ResNet-50 to obtain flattened feature maps from difference images of disaster struck vegetation areas followed by applying K-Means clustering algorithm on them. Thus, a complete labelled dataset is produced which is used to train two CNN models to classify the areas into undamaged and damaged classes for Model-1 and undamaged, mildly damaged and severely damaged classes for Model-2. Model-1 achieved 89.77% test accuracy and Model-2 achieved 85.69% test accuracy. Model-1 recorded 89.77% in micro f1-score, 76.99% for kappa score and 0.10 for overall error and Model-2 had 85.69%, 72.87% and 0.14 values in micro f1-score, kappa score and overall error respectively.