Cloud-Aerosol Discrimination in CATS Observations Using a Machine Learning Approach

Providing real-time feature detection and classification is important for future space-based and airborne lidar instruments. Applying machine learning and neural networks to lidar data can enable real-time data products at the sensor’s native resolution. We worked with uncalibrated observations of the atmosphere from the space-based Cloud-Aerosol Transport System (CATS) at native resolution to train a neural network to perform layer detection and cloud-aerosol discrimination (CAD). Finding layers in the atmosphere and assigning a class to those layers is in essence a semantic segmentation problem that can be approached using established machine vision tools. We applied the U-net architecture at varying levels of network depth and used a multitask learning strategy to give layer finding and layer classification equal priority during training. We found that the neural networks outperformed the CATS operational algorithm in both layer finding and CAD.

CATS operated between 2015 and 2018 aboard the International Space Station. For this study, only polarized 1064 nm laser observations were used to emphasize the kind of high quality results that can be achieved with a single lidar wavelength. We compare and contrast the performance of our trained neural networks against each other and the CATS operational algorithm. The CATS operational algorithm suffers from low spatial resolution, some misclassification problems, and a relatively slow observation-to-product pipeline (~6 hours). We show that our neural network can produce data products from raw, uncalibrated lidar observations at native resolution with better fidelity and in a fraction of the time than the traditional operational algorithms.

It is difficult to quantify which algorithm produces a better elevated data product without comparing against a known ‘ground truth.’ To address this issue, we generated simulated CATS observations using airborne Cloud Physics Lidar (CPL) data. The simulated observations were processed using the operational algorithm and our neural networks. The inferred presence of layers and CAD performed by each method were compared with the ground truth by calculating percent agreement and disagreement using confusion matrices. Each scene is dominated by the absence of a layer, so small percent improvements in agreement with ground truth are large improvements in data product fidelity.

We found that increasing the number of layers in the neural network increased agreement with the ground truth. In comparison with the operational algorithm, the shallower networks could classify at finer spatial resolution and perform mixed classification in a single layer. They were, however, less able to identify complete layers to the degree that they misclassified the presence of aerosol and cloud layers on the order of tens of percent. Increasing network depth mitigated this misclassification problem. Increasing the network depth from 1 to 6 layers increases agreement with the simulated ground truth by upwards of 5% in the CAD task in comparison with the operational algorithm. This result is achieved without spatial averaging to increase signal to noise ratios per profile.

We examined ensembling the deeper network’s layer finding with the classification abilities of the shallower networks. The ensembling strategy appears biased toward cloud identification as cloud identification agrees with the ground truth upwards of 5% more frequently than the operational algorithm, but it does not improve the detection of aerosols. Ensembled models also incorrectly over-classified clouds with respect to the operational algorithm upwards of 7% of the time.

We also investigated each algorithm’s stability with respect to artificially increased levels of solar background contamination.

We will deploy our optimized neural network model on an embedded computing platform with GPU accelerated compute capabilities and simulate performing inferences on raw data as it is being acquired. In a recent communication, it was shown that a Nvidia Jetson TX2 can perform near-real-time analysis of incoming lidar data. This would significantly reduce the time between observation of atmospheric conditions and data product availability. Near-real-time availability of lidar data products is critical for the atmospheric lidar data user community, especially for forecasting the impact of fire and volcanic events.

In this presentation, we show how we trained and optimized our neural networks to perform high quality layer detection and CAD at native sensor resolution using a single lidar wavelength. We show the experiments we conducted to compare and contrast the operational algorithm and our separate implementations of U-net with respect to network depth and ensembling, which includes how each algorithm is affected by increasing levels of solar contamination. We present how we quantified which algorithm is best for this application and show optimized neural network inferences on noisy data at high spatial and temporal resolution against the operational algorithm’s products. We also present progress on deploying our best neural network on a Jetson TX2.