An Interpretable, Satellite-Based, Machine Learning Convection Detection Algorithm Created by an Explainable Boosting Machine and Refined using Domain Knowledge.

Explainable Boosting Machines (EBMs) are extensions of Generalized Additive Models that have been shown to achieve comparable performance on some applications to modern black-box machine learning (ML) techniques, while also retaining the interpretability of glass-box methods and allowing for post hoc alterations without the need to retrain. In this study, we propose a method that utilizes EBMs to detect convective regions from satellite imagery. We also demonstrate the potential of using domain knowledge to update or correct the strategies that the automatically trained EBM models have settled on. The overall goal of this work was twofold: (1) accurately detect deep convection occurring in the upper troposphere and (2) demonstrate how we can harness EBM’s interpretability to inform the model’s strategies based on both the human understanding of the physics surrounding convection and the properties and trends found within the data.

Goal (1), detecting convection in satellite imagery, is a complex task that has been approached with many methods, including neural networks, in the past. Here, we seek to simplify the approach by combining two frameworks: the use of intelligent feature engineering with mathematical methods and the use of EBMs. Specifically, we use two channels of the Advanced Baseline Imager (ABI) sensor on the Geostationary Operational Environmental Satellite 16 (GOES-16), namely channel 2, visible imagery, and channel 13, infrared imagery, then use the mathematical framework of Gray Level Co-occurrence Matrices (GLCMs) to extract texture information from the visible satellite imagery. This preprocessing step allows us to then use simpler ML methods, such as EBMs, while still taking spatial context into account. Our EBM model then focuses on the binary classification task of predicting regions as either convective or non-convective using convection flags from Multi-Radar/Multi-Sensor (MRMS) system data as ground truth.

To achieve goal (2), we explored the performance of the trained EBM to determine which criterion the model was using to make its decisions and which parts of the decision-making process might benefit from alteration. Finding that the model erred toward underprediction, we sought to increase how often the model predicted regions as convective. In doing so, the number of false positives shot up dramatically. Thus, we worked toward decreasing the number of false positives while keeping the number of true positives at a desirable level. Because of the interpretability inherent to EBMs, we were able to utilize knowledge surrounding the physics of convective systems and of patterns found within the data to inform changes to the model’s predictions by, for example, increasing the importance of cold, highly textured areas and decreasing the importance of cold, flat areas.

The result of our efforts is a fully interpretable human-machine collaboration that has been shown to generally quantitatively outperform the unaltered model and produce predictions that are qualitatively evaluated to be reasonable by domain experts.