Speaker
Description
Forest parameter estimation via satellite remote sensing techniques offers substantial advantages compared to more traditional methods. Global coverage and fast revisit times have enabled constant monitoring of forest conditions, leading to robust datasets of remote sensing imagery.
Machine Learning has become an important tool when analyzing these increasingly large datasets, and even more recently, Deep Learning models have been shown to excel at dealing with even the most complex datasets, be it large, multi-channel images, or even time-series images.
The standard setup when training a Machine Learning model to predict forest parameters relies on data from a single satellite mission, often yielding good results. However, methods which combine two or more satellite missions have more recently been explored, hoping to capture the best properties from each sensor type, while covering for each other's flaws.
Specific Deep Learning architectures are also able to exploit the spatial correlations of pixels in images, and even temporal correlations in time-series images. This allows the model to learn more information from the same dataset when compared to classical Machine Learning models, or even shallow neural networks.
Our work focuses on estimating forest canopy height by training Deep Learning models using a variety of different remote sensing data sources. Our custom-built dataset consists of 12,000 256x256 forest patches over different sites of boreal forests in Quebec, Canada. The canopy height labels are provided by the government of Quebec's airborne lidar-derived canopy height model, while the features come from Copernicus Sentinel-1 and Sentinel-2 images, as well as weather data from the European Center for Medium-Range Weather Forecasts (ECMWF).
Our model architecture is based on a U-Net, which allows us to feed in an image stack, and get a canopy height prediction for each pixel in the entire image. Like other convolutional neural networks, it is able to leverage the spatial structure in an image. We've additionally modified it to be able to leverage the temporal structure, as well as while also being able to handle data from different sensors and data sources regardless of their native image resolution.
Our model vastly outperforms classical ML methods like Random Forests, as well as shallow neural networks, and even the regular U-Net architecture when training and testing on our same dataset. This demonstrates that the spatiotemporal correlations have a lot of relevant information that simpler models are not able to capture. Similarly, our full model trained on Sentinel-1, Sentinel-2, and weather data outperforms simpler versions where we omit any of the sensor data modalities.