Deforestation monitoring for climate change analysis using deep learning techniques on satellite imagery

Abstract

In the field of climate change analysis, a huge amount of information, derived from various sources and in various formats needs to be analyzed daily, for the production of accurate insights and predictions. Such a task, is heavily reliant on precise measurements and visual information. Thus, the instruments that are tasked with capturing this information are of high importance. Since climate change analysis is a wide field, the focus will be narrowed towards the detection of possible factors contributing to the phenomenon of deforestation. Furthermore, the type of information that will be processed is visual, making satellite imagery an ideal choice.

In this thesis, we first tackle the task of creating a pipeline for preprocessing said satellite imagery. The preprocessing step includes the possible transformations that will be performed on the images as well as the optimal set of bands with regards to the performance of the given model. Next, we will perform multi-label classification in an attempt to describe the content of the images in terms of the factors that contribute to the deforestation, using a set of tags. Taking into consideration the limited available resources, we employ EfficientNet, a lightweight Convolutional Neural Network which was found to achieve state-of-the-art results in image multi-label classification. Subsequently, as a baseline model we use VGG16 and as an experimental model we deploy the Vision Transformer, which seeks to integrate the Transformer layer that is widely used in natural language processing, into the field of computer vision. Furthermore, for variety purposes we finish our experiments by implementing the ResNet, DenseNet and MobileNet architectures. The results that are achieved are very promising, showcasing that there is high value in the visual information available with regards to the task of deforestation detection.