Unsupervised Wildfire Change Detection based on Auto-Encoders and Contrastive Learning

Each year, the severity and frequency of wildfires increase, resulting in billions of dollars in damage, and they are expected to increase due to climate change. Therefore, early wildfire spread detection can save immeasurable infrastructure, forest resources and human lives as well as reduce fire suppression costs. The aim of this study is to develop an autonomous system built on top of high-resolution multispectral satellite imagery, in an advanced deep learning architecture for detecting wildfires as well as monitoring their progression at the pixel level. Our proposal leverages unsupervised learning to remain general, and therefore it can compare performances on several free and commercial satellite solutions.

We propose two unsupervised models for feature extraction based on the current state of the art research in ML. The first method uses an auto-encoder (AE) network which learns to encode images into a compressed latent space and then reconstructs them into the original dimension. By learning how to reconstruct, a feature embedding can be obtained using the encoder part of the model. This model can be used for change detection by encoding sequential images into latent vectors and by measuring distances between these embeddings. Estimating uncertainties over predicted changes can be possibly explored using a variational version of the AE.

As a second method, we use a contrastive learning technique from recent ML research, the SimCLR model. It is trained to minimize the distance between augmentations of a single image, the so called "positives" while it is also evaluated on pairs of different images, the so called "negatives". This model learns to minimize cosine distance between positive samples. There are exciting possibilities of using contrastive learning with remote sensing imagery by having access to the geographical location of images. Following the previous approach, the distance between encoded images is also used for change detection. Unsupervised clustering can be experimented in follow-up processing of the learned embedding spaces. With the proposed methods we are able to detect changes in several study areas that have experienced fire. While these methods can be used with different satellite data sources, the originating spatial and spectral resolution influences the performances of the models.