Inspiration
Forest wildfires are a major environmental threat, stripping the world of carbon absorption while also releasing all the previously stored carbon by the trees into the atmosphere. As such, quantifying the carbon stored within forests is critical in risk-assessments for wildfires as well as conservation efforts for sustainable management of forest ecosystems. Carbon stock can be estimated through above ground biomass (ABG) measurements. The most direct forms of measuring ABG include laborious and destructive methods, requiring cutting down and drying of trees to assess the dry biomass of each sample.
As such, current methods of quantifying carbon stock within forests rely on remote sensing techniques to capture data on forest structure to predict ABG. Of the data collected on forest structure, the forest canopy height is consistently seen to be the key to an accurate assessment of forest biomass.
Currently the most accurate and advanced remote sensing technique of capturing canopy height relies on Detection and Ranging (LiDAR) to acquire high-precision three-dimensional information of forests. The current method for creating a canopy height model (CHM) involves subtracting the digital elevation model (DEM) representing the elevation of the bare earth's surface from the digital surface model (DSM) representing the elevation of the earth including the environment's features, in this case the tops of trees.
However, collecting LiDAR data relies on laser scanners mounted on aircrafts flying across the survey regions, making data acquisition expensive, time-consuming \cite{jakubowski_using_2012}. With the growing environmental threat caused by forest wildfires, there is an increasing need for an inexpensive yet accurate means of regularly measuring carbon stock in forests to help with forest conservation efforts and to aid governments to efficiently triage potential fire threats.
What it does
We built a machine learning model that predicts canopy height using only 2D satellite images, as opposed to the current field standard of using expensive LiDAR data acquisition techniques.
How we built it
To approach this problem, we first aim to build fully connected neural networks (FC-NNs) as our baseline model, as others have also attempted. Additionally, we hope to leverage transfer learning and plan to utilize ResNet, AlexNet, and VGGNet to see if pre-trained weights provide a stronger prediction performance. Finally, given that our outputs are a CHM with the same dimensions as output input images, we liken this problem to a style-transfer type problem, where we're transferring from the satellite imagery into a CHM. Thus, we are building a Generative Adversarial Network model to test if generative models perform better for this task. Our FC-NN yielded the following results: 0.30 prediction accuracy on the test set, and an RMSE of 19.6 meters per CHM. In comparison, a simple K-nearest-neighbor model used as a very rudimentary baseline, with hyperparameter of K = 5 (optimized over K = 1...7) had a prediction accuracy of 0.21 when quantizing the outputs to the nearest integer.
What's next for Carbon Quantification: ML to Estimate Canopy Height
In the future, we plan to improve accuracies by using a GAN and exploring transfer learning.
Log in or sign up for Devpost to join the conversation.