Evaluating Decision Tree-Based Machine Learning Models as Alternatives to Neural Networks in Retrieving Atmospheric VOC Columns from Thermal Infrared Satellite Measurements

Volatile organic compounds (VOCs) are atmospheric trace gases that impact air quality and human health. Isoprene is one of the most important atmospheric VOCs, and is predominantly emitted by terrestrial plants. Unfortunately, it remains challenging to quantify the impact of isoprene and other VOCs on climate and air quality due to significant uncertainties in their global concentrations and variability. Prior work in our research group has converted space-based thermal infrared measurements into isoprene column concentrations using artificial neural networks (ANNs), a machine learning model. This study investigates the suitability of two other machine learning algorithms, random forests (RF) and gradient-boosted tree (GBT) models, for estimating isoprene concentrations from satellite observations. The RF and GBT models are compared to the ANN model based on their predictive performance and computational efficiency. We find that the RF model provides similar performance to the ANN model, with an RMSE of 6.4 × 10¹⁴ molec cm^-2 and R² of 0.93. However, the RF model is computationally more efficient, taking 7 minutes to train compared to 7.5 hours for ANN. Results for GBT will also be presented, along with recommendations for future machine-learning based satellite retrievals of atmospheric VOCs.