Deep Convolutional Neural Network for Multi-Label Classification of Satellite Images of the Amazon Forest Fires
Most fires in Amazon tropical rainforest are man-made, which is because of many activities such as deforestation, burning trees for agriculture, slash and burn methods, etc. Dry environment, lightning strikes and volcanic eruptions are some other causes of forest fires, which are triggered by the nature. As the fires in the forests are difficult to predict in advance, satellite images can be used for the classification and prediction of them. In this research, the primary objective is to classify the satellite images using deep convolutional neural network, which has fire along with the other pre-trained classes such as agriculture and slash-and-burn. This research aims to obtain results, which can identify the origin of the fires. Planet dataset developed by Kaggle has been used for this research, which has images with different climatic conditions, such as cloudy and haze, which might make it tougher for the satellites to identify the fires in the forest. We have primarily used the VGG16 model and then compared the accuracy with the Resnet model.
Keywords: Convolutional neural network, forest fires, VGG, Resnet, multi-label classification, image augmentation, transfer learning.