Leveraging Deep Learning as a New Approach to Layer Detection and Cloud–Aerosol Classification Using ICESat-2 Atmospheric Data

Abstract

NASA’s Ice, Cloud, and land Elevation Satellite (ICESat-2), designed for surface altimetry, plays a pivotal role in providing precise ice sheet elevation measurements. While its primary focus is altimetry, ICESat-2 also offers valuable atmospheric data. Current conventional processing methods for producing atmospheric data products encounter challenges, particularly in conditions with low signal or high background noise. The thresholding technique traditionally used for atmospheric feature detection in lidar data uses a threshold value to accept signals while rejecting noise, which may result in signal loss or false detection in the presence of excessive noise. Traditional approaches for improving feature detection, such as averaging, lead to a trade-off between detection resolution and accuracy. In addition, the discrimination of cloud from aerosol in the identified features is difficult given ICESat-2’s single wavelength and lack of depolarization measurement capability. To address these challenges, we demonstrate atmospheric feature detection and cloud–aerosol discrimination using deep learning-based semantic segmentation by a convolutional neural network (CNN). The key findings from our research are the effectiveness of a deep learning model for feature detection and cloud–aerosol classification in ICESat-2 atmospheric data and the model’s surprising capability to detect complex atmospheric features at a finer resolution than is currently possible with traditional processing techniques. We identify several examples where the traditional feature detection and cloud–aerosol discrimination algorithms struggle, like in scenarios with several layers of vertically stacked clouds, or in the presence of clouds embedded within aerosol, and demonstrate the ability of the CNN model to detect such features, resolving the boundaries between adjacent layers and detecting clouds hidden within aerosol layers at a fine resolution.

1. Introduction

The most pressing threat Americans currently face from climate change is the escalation of extreme weather events, encompassing hurricanes, floods, droughts, heatwaves, and wildfires. These extreme weather disasters have resulted in over USD 1 trillion in damages in the past seven years alone, and in 2022, they led to the displacement of an estimated 3.4 million Americans from their homes [1]. Improving weather and climate models to better predict extreme events is thus a high priority that will require new and improved sources of data. In addition, poor air quality claims millions of lives annually [2] with associated costs in the United States ranging from tens to hundreds of billions of dollars [3]. Aerosols are key components of air pollution and, at high concentrations, can degrade air quality and pose serious health risks, including respiratory and cardiovascular diseases [4,5]. Predicting air pollution depends on integrating chemical transport models with numerical weather prediction models to address the intricate connections between air quality and meteorology. Despite advancements in these models, they encounter difficulties due to the dynamic and nonlinear characteristics of the atmosphere, complexities in pollutant emissions, and constraints in observational data [6,7,8].

Several recent studies emphasize the importance of understanding the vertical and spatial distribution of cloud and aerosol layers in the atmosphere, as well as the vertical heterogeneity of aerosols over specific regions [9,10]. Additionally, the characterization of aerosol vertical distributions is highlighted as crucial for cloud development and precipitation, affecting weather and climate by modifying cloud properties and spatial distribution, and thus influencing precipitation patterns and climate dynamics [11,12,13,14]. Research has shown that the vertical dependency of aerosol impacts plays a role in local-scale convective processes, affecting precipitation patterns [15]. These findings underscore the critical role of accurate vertical registration in studying the interactions between aerosols, clouds, and precipitation in the atmosphere [16].

The Advanced Topographic Laser Altimeter System (ATLAS) was launched aboard NASA’s Ice, Cloud, and land Elevation Satellite (ICESat-2) in September 2018 [17,18]. ATLAS is a single-wavelength lidar system designed to perform high-resolution altimetry measurements of the earth’s surface. Though specifically designed and optimized for its altimetry mission, ICESat-2 also measures backscatter from molecules, clouds, and aerosols in the earth’s atmosphere. Atmospheric data products based on these measurements reporting the vertical and spatial distribution of cloud and aerosol layers in the atmosphere are made publicly available. The objectives of this work are twofold: first, to improve the ICESat-2 atmospheric data products in a manner that contributes to achieving accurate vertical registration of atmospheric layers, and second, to build future capability that can be incorporated into sensors that possess onboard processing power, allowing data to be input directly into aerosol, cloud, and air quality models with minimal latency.

Previous studies applying machine learning (ML) techniques for layer identification and cloud–aerosol discrimination based on airborne or spaceborne lidar atmospheric profiles have shown improvements over traditional threshold-based methods [19]. The Cloud-Aerosol Transport System (CATS), an instrument that operated on the International Space Station (ISS) from 2015 to 2017, was a backscatter lidar with photon counting detection with a high repetition rate laser operating at 532 nm and 1064 nm [20]. Convolutional neural networks (CNNs) were applied to the CATS data for layer identification and cloud–aerosol discrimination [21], giving a 30% increase in the number of detected layers at any resolution and the capability to identify 40% more atmospheric features during daytime operations at a 5 km horizontal resolution compared to the 60 km horizontal resolution often required for daytime CATS operational data products. The real-time detection of atmospheric features by CNN using an airborne backscatter lidar was recently demonstrated by McGill et al. [22]. Their results were shown to agree well with traditional processing methods while also being produced with near-zero latency and at higher horizontal resolutions compared to the traditional method.

Many airborne and spaceborne lidar instruments are equipped with multiple wavelengths and the capability to measure depolarization (e.g., Cloud-Aerosol Transport System (CATS; [20]), Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO; [23]) and Cloud Physics Lidar (CPL; [24])). Lidar sensors optimized for atmospheric measurements rely on linear depolarization measurement and/or multiple wavelengths to differentiate cloud and aerosol types. The color ratio derived from multiple wavelengths can be used to distinguish various particle types such as dust, water clouds, ice clouds, and smoke [25,26]. In contrast, ICESat-2 utilizes a single wavelength and lacks depolarization measurement, thus posing a challenge for traditional atmospheric lidar processing methods. ICESat-2 data also have added complications due to the folding of atmospheric scattering information above 15 km [27].

In this study, we show the efficacy of CNN-based deep learning models to perform atmospheric feature detection and cloud–aerosol discrimination through data-driven semantic segmentation based on ICESat-2 profiles of atmospheric backscatter. The CNN-based method is compared qualitatively and quantitatively to the traditional ICESat-2 atmospheric data products. Our results show that an ensemble CNN-based approach can determine atmospheric features at native data resolution, and accurately discern between cloud and aerosol layers even in highly complex atmospheric scenes where the traditional processing techniques can fail. This work highlights the potential of CNN-based techniques for building improved ICESat-2 atmospheric data products.

2. Methods

2.1. ICESat-2 Atmospheric Data Products

The ICESat-2 satellite contains a single instrument, the Advanced Topographic Laser Altimeter System (ATLAS). Optimized for surface altimetry, ATLAS utilizes a high repetition rate, low per pulse energy, 532 nm laser and photon counting detectors. The ICESat-2 atmospheric profiles consist of 30 m binned photon returns in a 14 km vertical column from 14 km above surface level (ASL) to the surface. The atmospheric profiles are downlinked after summing 400 shots onboard the satellite, resulting in 25 Hz profiles with 280 m along track resolution. The ICESat-2 atmosphere Level 2A (L2A) product ATLAS 04 (ATL04) contains Normalized Relative Backscatter (NRB) profiles and calculated 532 nm calibration coefficients, in addition to other ancillary data. The NRB profiles are created from the profiles of raw photon counts by performing background subtraction, range square correction, and laser energy normalization. The background subtraction and calibration procedure have been described previously [27,28]. The calibration coefficient derived from the NRB profiles is used to compute the Calibrated Attenuated Backscatter (CAB) profiles contained within the ICESat-2 atmosphere Level 3A (L3A) product ATLAS 09 (ATL09). ATL09 data are provided as granules (files) that span one complete satellite orbit. Included in the ATL09 data product are the top and bottom heights of cloud and aerosol layers detected in the data, and the classification of these identified layers as either cloud, aerosol, or indeterminate as determined by the ICESat-2 atmosphere L3A operational algorithms.

2.2. ICESat-2 Atmosphere L3A Algorithms

The Density Dimension Algorithm (DDA) is used to identify atmospheric features within the ICESat-2 data [28,29,30]. The input to the DDA is the NRB profiles arranged as an image into a two-dimensional matrix of vertical profiles collected over time along a track. For each vertical bin in each atmospheric NRB profile, a convolution operation of the neighboring data field with the kernel of an anisotropic Gaussian radial basis function in a sliding window signal averaging operation is performed, resulting in a backscatter density field. The density is calculated in two passes with different parameters. The first pass (sliding window size 7 × 7) is used to identify optically thick cloud layers. The second pass removes the signal from optically thick layers and recalculates the density with a larger sliding window (size 7 × 13) to identify optically thin layers. The identification of signal over noise is performed by an auto-adaptive statistical thresholding function that changes with the time of day and atmospheric conditions, which detects the top and bottom heights of layers from the backscatter density field. The resulting layer masks from each pass are combined to give the total atmospheric layer mask. The ATL09 cloud–aerosol discrimination algorithm uses the layer-wise mean CAB, mean altitude, and mean relative humidity (obtained from ancillary data sources) to compute a confidence value for each layer based on empirically derived, altitude-dependent layer-type likelihoods. The confidence value is used together with a threshold value $\chi$ to determine whether a layer is cloud, aerosol, or indeterminate. Palm et al. [28] discuss the accuracy and reliability of the ICESat-2 L3A ATL09 data product and note that the described layer discrimination algorithm is generally correct, though the data product does contain errors in classification.

2.3. CNN Technique

Our approach to ICESat-2 layer detection and cloud–aerosol discrimination is based on a fully convolutional neural network trained end to end for pixel-wise semantic segmentation [31]. The choice of using a fully convolutional network is motivated by its ability to perform pixel-level and predictions and to make predictions given arbitrary-sized inputs. This permits the efficient training of the network based on ICESat-2 data patches, and allows inference to be performed whole image at a time by dense feedforward computation. A standard U-Net CNN was trained for binary feature detection and cloud–aerosol discrimination [32]. A previous approach by Yorks et al. [21] has treated the combined feature detection and cloud–aerosol discrimination problem as a multi-class learning problem, where the predicted classes are “cloud”, “aerosol”, and “no layer detected”. In contrast, we structure our learning problem to be a binary, multi-task learning problem. For each vertical bin in each atmospheric profile, we predict as a first task whether that bin belongs to a layer or not, and as a second task whether that bin contains cloud or aerosol. The layer mask produced by the first task is applied to the cloud–aerosol prediction map to produce the multi-class feature detection map. In preliminary work, we found decoupling the layer detection and cloud–aerosol discrimination tasks gave significant improvements to the latter.

2.4. Dataset Preparation

To construct a dataset for training the CNN model, one month of ICESat-2 L3A ATL09 (version V006) data collected by ICESat-2 during November 2018 are obtained from the National Snow and Ice Data Center (NSIDC; [33]). One month of data are selected, as they provide a diverse, representative sampling of cloud and aerosol scenes yet are small enough to allow the CNN model to be trained relatively quickly. To construct a dataset for model validation, we select 12 individual granules, one from each month in 2019. An additional test dataset consisting of one month of ATL09 data collected by ICESat-2 during April 2023 is selected for evaluating the final model. The input to the CNN model, like the DDA algorithm, is comprised of atmospheric profiles arranged as an image. Each bin is assigned three attributes: the bin altitude, the CAB, and the relative humidity. These attributes, all provided within the ATL09 file, represent the CNN input channels. The cloud–aerosol scene for each CAB curtain profile is constructed using the layer top and bottom heights, and the associated cloud–aerosol classifications, assigned to each atmospheric profile by the ICESat-2 L3A operational algorithm. Separate binary labels for each pixel are obtained from the cloud–aerosol scene: the layer detection label (layer detected or no layer detected) and the cloud–aerosol discrimination label (cloud detected or aerosol detected). Vertical bins where no layer is detected or where the layer is indeterminate are assigned a dummy cloud–aerosol discrimination label (−100) that excludes it from the respective loss calculation at train time. The training dataset is further processed by a random cropping operation that produces 128 image patches, each 224 × 224 pixels, per data granule. A small fraction of training data are folded out as training validation data. The mean and standard deviation of each input channel are computed across the training set population. These statistics are used to apply a standardization transform to all data inputs.

2.5. Model Optimization and Evaluation

The Dice coefficient [34,35] generalized as a loss function for fuzzy binary vectors is used to optimize the CNN model [36]. The Dice loss is defined as:

$$D(\mathbf{p},\mathbf{l})=1-\frac{2{\sum }_i{p}_i{l}_i}{{\sum }_i{p}_i+l_i}$$

(1)

where $\mathbf{p}\equiv \left\{p_i\right\}$, $p_i\in [0,1]$ is the CNN output, a continuous variable representing the normalized positive class probability for the i-th pixel, and $\mathbf{l}\equiv \left\{l_i\right\}$, $l_i\in [0,1]$ are the corresponding binary ground truth labels. During training, the feature detection loss and cloud–aerosol discrimination loss for each training example are calculated separately and summed to give the total loss for optimization. The evaluation of the feature detection loss on the training validation data is performed once per training epoch in order to select the final model parameters by validation early stopping. At test time, we assess CNN cloud–aerosol discrimination performance by constructing the confusion matrix and calculate several accuracy metrics such as the precision and the recall:

$$\mathrm{precision}=\frac{TP}{TP+FP}$$

(2)

$$\mathrm{recall}=\frac{TP}{TP+FN}$$

(3)

where $TP$, $FP$, $FN$, and $TN$ are the true positive, false positive, false negative, and true negative classifications, respectively. The precision measures the quality of class prediction as the ratio between the total number of true positive predictions and the sum of the true positive and false positive predictions. The recall measures the ability of the model to identify all instances of a particular class within a dataset as the ratio between the total number of true positive predictions and the sum of the true positive and false negative predictions. We calculate the precision and recall for each data class (cloud, aerosol, and no layer detected) individually.

2.6. Implementation Details

All models are built and trained using the Pytorch deep learning library [37]. The CNN model is trained with the Adam optimizer [38] using a minibatch size of 16 examples and a learning rate of 2 × ${10}^{-4}$, for a total of 25 epochs. We use 32 feature maps at the initial U-Net encoding step; the number of feature maps is doubled after each encoding step, and halved after each decoding step. We refer to the number of encoding steps used in the encoder part of the U-Net architecture as the number of layers or CNN depth, which is varied between one and six in this study. We rely on several open-source Python libraries for metric evaluation and data visualization [39,40,41].

3. Results

3.1. Layer Detection

After training, we assess the ability of the CNN model to perform layer detection by processing selected ICESat-2 ATL09 data granules collected during 2019 contained in our validation set. Figure 1 shows an example of a processed scene from 1 January 2019. Panel A in Figure 1 shows the ICESat-2 532 nm profiles of calibrated attenuated backscatter on which layer detection is performed. Panel B shows the layer finding results from standard ICESat-2 data processing (i.e., using the DDA algorithm described in Section 2.2). Panel C shows the layer finding results from CNN data processing. It is evident from Figure 1 that the CNN algorithm is performing well when compared qualitatively to the traditional layer detection technique.

Figure 2 shows the confusion matrix, comparing the classifications made by the CNN method with the traditional layer detection technique, using the data shown in Figure 1. Figure 2 shows good agreement, with over 98% of bins categorized as “No Layer” by the traditional technique being similarly identified by the CNN method, and 88% of bins categorized as “Layer” present by the traditional technique being similarly identified by the CNN method. The off-diagonal entries of the confusion matrix show disagreement, and we find that the CNN method classifies 1% of the bins predicted by the traditional method as “No Layer” as “Layer”, and nearly 12% of bins identified by the traditional method as “Layer” are instead classified as “No Layer” by the CNN. We report that disagreement on layer detection between the traditional method and the CNN-based method typically arises near layer boundaries, where the CNN predicts some layers to have a somewhat smaller vertical extent, rather than the CNN-based method ignoring or missing complete layers. We find similar results when analyzing the CNN layer detection method results across the remaining data granules in the validation set.

3.2. Cloud–Aerosol Discrimination

We assess the ability of the CNN model to perform cloud–aerosol discrimination by processing selected ICESat-2 ATL09 data granules collected during 2019 contained in our validation set. Figure 3 shows the identical scene from 1 January 2019 shown in Figure 1 now processed for cloud–aerosol discrimination. Panel A in Figure 3 shows the ICESat-2 532 nm profiles of calibrated attenuated backscatter on which layer detection and cloud–aerosol discrimination are performed. Panel B shows the layer finding results from standard ICESat-2 data processing (i.e., using the DDA algorithm together with the empirical confidence-based classification described in Section 2.2). Panel C shows the layer finding and cloud–aerosol discrimination results from CNN data processing. It is evident from Figure 3 that the CNN algorithm is performing well when compared qualitatively to the traditional cloud–aerosol discrimination technique.

Figure 4 shows the confusion matrix, comparing the cloud–aerosol classification made by the CNN method with the traditional cloud–aerosol discrimination technique, using the data shown in Figure 3. The layer detection performance is similar to that in Figure 2, showing over 98% agreement between traditional and CNN-based methods for “No Layer” detection. The cloud–aerosol discrimination provides additional insight into the discrepancies between traditional and CNN-based methods. Over 86% of cloud layers detected by traditional techniques are identified similarly by the CNN-based method, and nearly 75% of aerosol layers detected by traditional techniques are identified similarly by the CNN-based method. The largest discrepancy between traditional processing techniques and CNN-based methods remains the misclassification of “Layer” bins found by traditional methods as “No Layer” by the CNN-based method. We find that of these misclassified bins, about 8% of cloud layers determined by traditional cloud–aerosol discrimination are classified as “No Layer” by the CNN-based method, and that slightly over 18% of aerosol layers determined by traditional cloud–aerosol discrimination are classified as “No Layer” by the CNN-based method. The next largest discrepancy happens between cloud–aerosol layers found by traditional and CNN-based methods. We find that 5% of the bins identified by traditional techniques as cloud are classified as aerosol by the CNN-based method. Conversely, nearly 7% of the bins identified by traditional techniques as aerosol are classified as cloud by the CNN-based method.

We further demonstrate the ability of the CNN model to perform cloud–aerosol discrimination by showing an additional processed atmospheric profile from 1 April 2019 in Figure 5 together with an associated confusion matrix in Figure 6. Panel A in Figure 5 shows the ICESat-2 532 nm profiles of calibrated attenuated backscatter on which layer detection and cloud–aerosol discrimination is performed. Panel B shows the layer-finding results from standard ICESat-2 data processing. Panel C shows the layer finding and cloud–aerosol discrimination results from CNN data processing. We again see good qualitative agreement between cloud–aerosol discrimination performed by the traditional and CNN-based methods, although some discrepancies are now visibly identifiable. For example, there are cloud layers identified by the CNN-based method near 00:35:00 UTC time and 00:40:00 that are classified by traditional techniques as indeterminate. There is some additional disagreement between cloud–aerosol discrimination that occurs after 01:10:00 UTC time. The layer detection performance shown in the confusion matrix demonstrates again over 98% agreement between traditional and CNN-based methods for “No Layer” detection. Over 93% of cloud layers and nearly 74% of aerosol layers detected by traditional techniques are identified similarly by the CNN-based method. We find that about 5% of cloud layers determined by traditional cloud–aerosol discrimination are classified as “No Layer” by the CNN-based method, and that 17% of aerosol layers determined by traditional cloud–aerosol discrimination are classified as “No Layer” by the CNN-based method. We find that 1% of the bins identified by traditional techniques as cloud are classified as aerosol by the CNN-based method and, conversely, nearly 10% of the bins identified by traditional techniques as aerosol are classified as cloud by the CNN-based method.

3.3. Role of CNN Receptive Field

Figure 7 shows the results of validation set evaluation using trained CNN models, where the CNN depth is uniformly increased from one to six. The CNN depth controls the size of the CNN receptive field, defined as the window size of the data input used to make a single bin prediction. The CNN receptive field plays a similar role in the CNN model as the sliding window size used for layer detection in traditional data products. The four panels in Figure 7 show (from left to right) the cloud classification precision, the cloud classification recall, the aerosol classification precision, and the aerosol classification recall, as the CNN receptive field size is varied from a 5 × 5 window (CNN depth 1) to a 224 × 224 window (the training data sample size, CNN depth 6). The precision and recall metrics are calculated individually for each ATL09 data granule, and the data points in Figure 7 show the average metric value across the validation set together with 95% confidence intervals. Each validation metric data point is fit to a simple saturation growth model as a function of CNN receptive field, shown by the solid black lines together with model uncertainty computed by Monte Carlo error propagation shown in gray.

The results in Figure 7 show that the CNN-based data processing approaches the performance of traditional processing algorithms near CNN depth 2 or 3 (receptive field size 17 × 17 or 41 × 41, respectively). These results suggest there is little benefit to extending the model deeper. We can rationalize this finding by considering that the “ground truth” data were generated utilizing a 7 × 13 receptive field window during the layer detection step. We observe the cloud precision and cloud recall in the high CNN depth limit to be 80% and 90%, respectively, indicating that cloud layer classifications are correct 80% of the time and that we identify over 90% of cloud layers found by the traditional data-processing algorithms. We observe the aerosol precision and aerosol recall in the high CNN depth limit to be 70% and nearly 75%, respectively. These findings are consistent with the CNN model predicted results evaluated on individual validation set profiles and discussed in Section 3.1 and Section 3.2.

Figure 8 shows a daytime portion of the scene from 1 April 2019 processed for cloud–aerosol discrimination using CNNs of varying depth. Panel A in Figure 8 shows the ICESat-2 532 nm profiles of calibrated attenuated backscatter on which layer detection and cloud–aerosol discrimination are performed. Panel B shows the layer finding results from standard ICESat-2 data processing. Panel C shows the layer finding and cloud–aerosol discrimination results from CNN (depth one) data processing. Panel D shows the layer finding and cloud–aerosol discrimination results from CNN (depth four) data processing. The visual inspection of the calibrated backscatter shows a segment of the April 2019 scene where clouds are embedded within an aerosol layer. This scene highlights a weakness present in the current operational algorithm. The ICESat-2 atmosphere operational algorithm successfully identifies the layer as primarily aerosol but produces “blocky”, columnar cloud–aerosol predictions. The operational algorithm is assigning the layer to be either only cloud or only aerosol, and fails to detect smaller cloud layers embedded within the aerosol layer. The CNN (depth one) data processing successfully detects several cloud layers embedded within the aerosol layer, although there are several small, false positive layers detected in the noisy daytime scene. The CNN (depth four) data-processing method shows cleaner layer boundaries and better overall agreement with the traditional data product, but the CNN cloud–aerosol discrimination predictions begin to take on the columnar structure present in the traditional data product. Indeed, as shown in Figure 7, the CNN predictions tend toward faithfully reproducing the characteristics of the traditional data product in the limit of high depth.

3.4. Ensemble CNN Single Profile Comparisons

Panel C of Figure 8 demonstrates the ability of the low-depth CNN-based data-processing technique to perform cloud–aerosol discrimination at a fine resolution. However, the presence of many small, false-positive layers detected near layer boundaries lowers the fidelity and limits the utility of these predictions. Conversely, Panel D of Figure 8 demonstrates the ability of the moderate depth CNN-based data-processing technique to perform accurate layer detection while the cloud–aerosol discrimination predictions are similar to that of the traditional data product. We utilize the strengths of these two models by ensembling their predictions. Specifically, we apply the atmospheric layer mask produced by the depth four CNN to the cloud–aerosol discrimination predictions of the depth one CNN to obtain the final multi-class feature detection map. This ensembling results in a superior data product, where atmospheric layers are accurately detected and the cloud–aerosol discrimination is performed at a fine resolution. We investigate this approach through a comparison of the ensemble CNN-based data-processing method with the traditional data product at the resolution of a single backscatter profile in the following selected scenes from the validation dataset.

Figure 9 shows the 1 April 2019 scene examined previously. We select 500 records from the data granule and zoom in to display cloud–aerosol prediction at a higher vertical and horizontal resolution. This figure depicts a daytime scene where there are several cloud layers embedded inside an aerosol layer contained within the planetary boundary layer. Figure 9 is complemented by Figure 10, which shows the calibrated attenuated backscatter profile, DDA-derived backscatter density field, and a comparison between the cloud–aerosol layer predictions based on CNN-based data processing and the traditional data product obtained at the location of the red line in Figure 9A. The calibrated attenuated backscatter profile (left) is horizontally averaged over five profiles, which we choose based on the receptive field size of the depth one CNN. Over the backscatter profile, we plot the top and bottom heights and characterization of each layer identified by the CNN-based data-processing technique. The backscatter density profile (right) is obtained from the ATL09 data product. Over the backscatter density profile, we plot the top and bottom heights and characterization of each layer identified by the traditional data-processing technique.

From Figure 9, we observe the improvements offered by the CNN-based data-processing technique over the traditional data product. We observe from the backscatter profile the presence of a geometrically thin cloud layer sitting at the bottom of an optically thin aerosol layer. We find good agreement between traditional and CNN-based methods for the predicted layer top and bottom heights. Examination of the backscatter density profile in Figure 10 indicates that the thin cloud is identified in the first pass of the DDA algorithm, then is removed from the signal for the subsequent second pass where the aerosol layer is identified. The combination of layer masks from the two separated passes of the DDA algorithm, however, produces only a single atmospheric layer. The signal returns due to the cloud are much higher than those due to the aerosol, and the mean calibrated attenuated backscatter of the layer computed for use in the traditional cloud–aerosol discrimination algorithm results in high confidence that this layer is entirely cloud. The ensemble CNN-based data-processing technique successfully resolves the thin cloud layer, and several other neighboring cloud layers, embedded within the aerosol layer.

Figure 11 and its complement Figure 12 shows another scene selected from the 01 April 2019 data. Here, we can observe from the calibrated attenuated backscatter profile the presence of a cloud layer embedded within the center of an aerosol layer. Inspection of the backscatter density profile indicates that the cloud layer is identified in the first pass of the DDA algorithm, then removed from the signal for the subsequent second pass where the aerosol layer is identified, again resulting in only a single layer being identified. In this instance, the layer is instead typed as aerosol, and the traditional approach fails to identify the cloud layer. The ensemble CNN-based data-processing technique is able to successfully resolve the cloud layer embedded within the aerosol layer.

The final scene selected from 1 April 2019 data is shown in Figure 13 and its complement Figure 14. Here, we can observe from the calibrated attenuated backscatter profile a complex scene with intermixing between several cloud and aerosol layers stacked vertically. The ensemble CNN predicted layer top and bottom heights are in good agreement with the traditional data product. In the selected atmospheric profile, the traditional data-processing technique identifies two unique layers, and classifies them both as cloud layers. The ensemble CNN similarly identifies these two layers but classifies the layer beginning near 3 km above surface level as cloud, and the second layer beginning near 2 km above surface level as aerosol. Within the aerosol layer, a thin cloud layer is also identified. In the overall extended scene, we can observe that the traditional cloud–aerosol discrimination struggles to distinguish the intermixing of cloud and aerosol layers, resulting in “blocky” columnar layer classifications. This example demonstrates the ability of the ensemble CNN-based technique to analyze complex scenes and offer some improvements to the traditional data product.

3.5. Test Evaluation of Ensemble CNN Approach

We assess the global applicability of the CNN model to perform cloud–aerosol discrimination by processing all ICESat-2 ATL09 data granules collected during April 2023. Figure 15 shows the confusion matrix, comparing the cloud–aerosol classification made by the CNN method with the traditional cloud–aerosol discrimination technique across 449 data granules contained in our test set. The layer detection performance across the test set is similar to that in previous validation assessments, showing over 98% agreement between traditional and CNN-based methods for “No Layer” detection. Over 90% of cloud layers detected by traditional techniques are identified similarly by the CNN-based method (95% CI [90.3%, 90.7%]), and 56% of aerosol layers detected by traditional techniques are identified similarly by the CNN-based method (95% CI [54.3%, 55.8%]). We note that the cloud and aerosol recall results on the test set are similar to the cloud and aerosol recall results evaluated on the validation set at a low CNN depth reported in Figure 7. We find that 25% of atmospheric bins classified as aerosol by the traditional cloud–aerosol discrimination technique is classified as cloud by the ensemble CNN-based approach, which would include observations where the ensemble CNN-based method is able to identify multiple cloud layers embedded with aerosol layers in the representative case studies presented in Section 3.4.

We select one representative granule from the test set to visualize. Figure 16 shows a scene from 30 April 2023. Both the ATL09 operational algorithm and CAD CNN identify the layers present in the atmospheric profile. When we view the complement Figure 17, we observe from the atmospheric profile the presence of numerous cloud layers embedded within an aerosol layer. While the CAD CNN is able to accurately resolve the cloud layers present within the aerosol layer, the traditional algorithm fails to identify multiple layers present at a lower altitude and instead assigns the single layer identified within each single profile as either cloud or aerosol.

4. Discussion

This work represents our first attempts at constructing improved ICESat-2 atmospheric data products by a deep learning-based approach. We have performed qualitative and quantitative comparisons between our method and the traditional data product for atmospheric feature detection and cloud–aerosol discrimination. The model validation results are used to compare the CNN-based prediction against the layers detected and classified by traditional processing algorithms. Because the ICESat-2 data are real and not derived from a model simulation, the actual ground truth classification of the atmospheric data is unknown, and the ground truth for comparison is considered to be that produced by the current, traditional processing algorithms. We compare the results of CNN-based prediction on validation data against traditional data products but understand the traditional data product itself is not 100% accurate and does contain errors.

We find that the CNN method classifies 1% of the bins predicted by the traditional method as “No Layer” as “Layer”, and nearly 12% of bins identified by the traditional method as “Layer” are instead classified as “No Layer” by the CNN. This level of agreement on CNN-based layer detection is similar to that found in McGill et al. [22], where the authors report 10% of data bins with a similar misclassification when performing CNN-based layer detection on real-time CPL lidar profiles. Both their work and ours perform layer detection by semantic segmentation with a fully convolutional neural network, predicting the presence or absence of a layer in a pixel-by-pixel approach. The depth of the CNN, and by turn the CNN receptive field size, controls the extent of signal averaging. Since the bin classifications are jointly predicted by feedforward computation, our layer detection algorithm also takes into account the classifications of neighboring bins. This is in contrast to the traditional approach based on thresholding, where the method identifies the top and bottom bin heights of an atmospheric feature only. In addition, although the DDA algorithm is used to aggregate signals from neighboring data profiles, once the aggregation has occurred, the thresholding function used for layer detection operates on a single profile basis. We refer to this approach as profile-by-profile prediction.

The results presented in Section 3 highlight an important finding of this work: the CNN-based data-processing method is able to detect and classify (embedded) layers at a higher resolution compared with traditional data products. This finding is notable considering the only CNN training data available are those of the traditional data product. One explanation for this phenomenon is the ability of CNNs to generalize in weakly and inaccurately supervised learning settings [42]. Inaccurate supervision occurs when the label information used for training contains errors or label noise [43,44]. This situation commonly arises in computer vision applications based on deep learning where the labeled data are collected by crowdsourcing or other unreliable data-labeling techniques. In this work, the traditional cloud–aerosol discrimination algorithm is reliable but not perfect, and the classifications given in the ATL09 data product can be viewed to contain some level of noise. We have highlighted mixed columnar predictions of cloud and aerosol layers provided by the traditional data product in scenarios where embedded layers are present as one example of label noise. A recent study has demonstrated that CNNs are able to generalize on unseen data after training on massively noisy data, instead of merely memorizing noise [45]. Rolnick et al. [45] have shown that the minimum dataset size required for effective training increases with noise level, that a large enough training set can accommodate a wide range of noise levels, and that dataset noise can be partly compensated for by adjusting model hyperparameters such as batch size and learning rate. We consider that the training set size used in this work is sufficient to accommodate the level of noise present in the ATL09 traditional data product and use model depth to control the extent to which the model can memorize the noise. We have observed in Section 3.3 that in the limit of large depth, the CNN model can memorize the noise present in the traditional data product. We consider that the CNN receptive field size plays a role in regularizing the neural network, which improves its ability to accurately generalize on unseen data.

Yorks et al. [21] also report the ability of a similar CNN-based approach for detecting and classifying layers at a higher resolution compared to that needed for the traditional processing algorithm when analyzing CATS atmospheric backscatter profiles. Their model operates on processed backscatter data profiles averaged to a horizontal resolution of 5 km. They are able to detect atmospheric layers at 5 km horizontal resolution, which would require profile averaging to a horizontal resolution of 60 km for detection by traditional methods. Our approach operates on data at native horizontal resolution (280 m on track) without the need for profile averaging. Our studies also differ in how the machine learning models are trained. While previous studies approached the learning problem by multi-class classification, we split the multi-class learning problem into a series of binary classification problems trained jointly (multi-task learning). This allows us to combine layer detection and cloud–aerosol predictions from models of varying receptive field size, which enables our reported improvements over traditional data products.

5. Conclusions

We have demonstrated the application of CNN-based data-processing techniques to atmospheric feature detection and cloud–aerosol discrimination using ICESat-2 profiles of atmospheric backscatter. The atmospheric feature detection performance of the CNN-based method shows good agreement with the traditional data product. We have shown how cloud–aerosol discrimination by CNN can faithfully reproduce layer classification assigned by the traditional processing algorithms in the limit of large CNN depth. More importantly, we have shown through qualitative analysis of backscatter lidar scenes and individual profile comparisons between CNN-based predictions and traditional data product that an ensemble of low- and moderate-depth CNNs can produce high-quality layer boundaries and successfully overcome some limitations of the traditional data product, showing the ability to classify and resolve complex cloud–aerosol scenes. In several instances of cloud layers embedded within aerosol layers contained inside the planetary boundary layer, the CNN-based approach outperforms the traditional data product, identifying additional layers and providing cloud–aerosol discrimination at a finer resolution. Our study demonstrates the potential of machine learning-based methods for building improved ICESat-2 atmospheric data products.

Importantly, the results shown are obtained using only ICESat-2’s single wavelength intensity data. Whereas conventional wisdom typically asserts that multiple wavelengths and/or a linear depolarization measurement are required to enable cloud–aerosol discrimination, we find that is not true and single-wavelength intensity data are sufficient. The ML/CNN method allows the detection of cloud embedded in aerosol layers based only on pattern recognition. Beyond improving the ICESat-2 atmospheric data product, the implication is that incorporating our CNN-based data-processing techniques into airborne or space-based lidar will enable the real-time determination of both layer heights and cloud–aerosol discrimination. The ability to generate these data products that are essential for aerosol, cloud, and air quality models with low latency and high accuracy is essential to enhance the utility of future space-based lidar remote sensors.