3.1

Input Data

Aqua MODerate-resolution Imaging Spectroradiometer (MODIS) Collection 6.1 1-km brightness temperatures (BT) and cloud-top phase and cloud optical depths tM are retrieved using the Clouds and Earth’s Radiant Energy System (CERES) Edition 4 algorithms [15] and are referred to as CERES-MODIS or CM products. Reanalyses from version 5.4 of the Global Modeling Assimilation Office Global Earth Observing System GEOS-5.4, an update of the versions described by [20], provide surface skin temperature and vertical profiles of relative humidity.

3.2

Active Sensor Data

CloudSat, CALIPSO, and Aqua are part of the A-Train satellite constellation and take measurements continuously with time differences of less than three minutes. CALIOP and the CloudSat radar are near-nadir viewing and aligned so that they view roughly the same area along their respective flight tracks. Their flight tracks are typically near the Aqua nadir path, so that MODIS views the same scene at viewing zenith angles less than 18°. MLANN is trained and validated using CALIPSO Version 4 [21] and CloudSat R04 [22] vertical profiles of clouds matched with 1-km Aqua MODIS Collection 6.1 radiances in an update of the CloudSat, CALIPSO, CERES, and MODIS (C3M) product of [23]. The C3M first converts the CloudSat CLDCLASS high-confidence cloud profiles from 240–m to 60-m vertical resolution, then merges them with the CALIPSO cloud profile (CPRO) and vertical feature mask (VFM) to produce a complete vertical profile of cloud-filled layers. All CALIPSO averaging resolutions, 0.33 – 80 km, are used to define the cloud profiles. The CALIPSO-CloudSat (CC) data are merged with the corresponding Aqua MODIS radiances and CERES Edition 4 cloud property retrievals, which include tM. In addition to defining the cloud layers, the CALIPSO-CloudSat optical depth for the ice layer tCC was computed for each pixel identified as ML by the CALIPSO-CloudSat profiles to perform the additional analyses. The value of tCC is equal to the CALIPSO ice cloud optical depth, when the CALIOP signal shows a return from the lower layer cloud, otherwise it is equal to the combined CALIPSO-CloudSat optical depth. Similarly, the top and base heights of the ice cloud layer, Z_TCC and Z_BCC, respectively, are also determined for ML cloud pixels.

3.3

Input Layer

The input parameters from MODIS include latitude, longitude, brightness temperatures BT(λ) and brightness temperature differences BTD(λ₁-λ₂) = BT(λ₁) – BT(λ₂), where λ is the wavelength in μm, and the visible-channel optical depth τ_M. Specifically, BT values at 3.7, 6.7, 8.5, 11, and 12 μm are used here along with BTD(3.7-11), BTD(6.7-11), BTD(8.5-11), and BTD(11-12). The GEOS input data consist of the surface temperature and relative humidity at the surface and at 850, 700, 500, 400, 300, 200, and 100 hPa). During the day, the solar zenith angle (SZA) in degrees is also used in the input. Inclusion of the GEOS humidity profiles represents a marked change in the input from that of [16].

3.4

Output Layer: Multilayer Identification

The MLANN as formulated here has four training sets for each parameter of interest: night (SZA > 82°) and day (SZA ≤ 82°) with CM cloud-top phase of ice or water, so that each set yields the weights, w_ij and w_j, along with the constants b and c. Training is performed using the minimum cloud separation threshold ΔZ_min = 1 and 3 km, where the separation ΔZ is the difference in altitude between the bottom of the lowest ice layer and the top of the highest water layer in the CC profile. An ice-cloud layer is assumed to be present in the profile, if at least one layer occurs at a height above the altitude corresponding to 253 K and no temperature inversion exists in atmospheric layer between the altitudes corresponding to 273 K and 253 K. All C3M data having an ice cloud layer by that definition are used as output in the MLANN training. A given profile is assumed to contain a ML cloud system if a cloudy layer exists at least ΔZ_min below the lowest ice cloud layer and the base of that cloud layer is below the 253-K level. Some ice-over-water layer systems will be considered as SL if they are closer than ΔZ_min. ML clouds defined by ΔZ_min = 1 km are classified as SL clouds when ΔZ_min = 3 km. The output consists of a probability. If the probability < 0.5, the classification is SL, otherwise it is ML. For this study, all qualifying data from April 2009 are used in the training process with 60% for training and 20% each for testing and validation.

3.5

Output Layer: Other Parameter Estimation

To estimate the upper cloud optical depth τ_ul, the MLANN is trained with the same input variables in the same sets, but the output variable is τ_CC. Values for the weights are also computed for upper layer cloud top and cloud base heights, Z_tul and Z_bul, respectively, using the top of the highest and bottom of the lowest ice cloud in the CC ML profile. Qualifying July and September 2009 data are used for training these three parameter networks using the same fractions of the data for training, testing, and validating as employed for the ML detection.

3.6

Neural Network Retrievals

Once all of the weights are computed in the training process, they can then be used in the MLANN to produce output values for other input datasets having the same variables. When using the MLANN, the CERES Edition 4 cloud phase and SZA are used to select which set of weights will be used in the MLANN to determine the layering classification.

4.1

Multilayer Clouds

Figure 2 plots the CC cloud profiles determined over a midlatitude area at night around 07 UTC, 2 July 2009. The CC layering classification, gray for SL and orange for ML, is based on ΔZ_min = 3 km, The corresponding MLANN ML detection is shown in orange along the line at 15.4 km. At first glance, it appears that the MLANN detects the ML clouds quite well with gaps mostly corresponding to the gray areas. For the SL areas on the right side of the plot, however, the gaps are noticeably larger than the gray segments. If the CC ML criterion were ΔZ_min = 1 km, the gray area indicated by the oval would have been included in the CC ML category.

Figure 2.

CALIPSO-CloudSat cloud profile from C3M for 2 July 2008with CC ML clouds having ΔZ_min = 3 km indicated in orange and CC SL denoted in gray. The MLANN ML identification for each profile is indicated as an orange dot at 15.4 km. CC profiles that would be considered ML, if ΔZ_min = 1 km, are highlighted with the oval.

In general, the use of ΔZ_min = 3 km captures a significant portion of the ice-over-water ML clouds as seen in Figure 3, which shows histograms of the CC cloud classifications for April 2009 for each of the four CM categories. During the day for CM clouds categorized as liquid (Figure 3a), ΔZ_min = 3 km accounts for 84% of the ML clouds, assuming that clouds with ΔZ < 1 km are SL. At night (Figure 3c), that fraction rises to 87%. The contribution of clouds having ΔZ > 3 km drops to 62 and 67% during the day (Figure 3b) and night (Figure 3d), respectively, when CM identifies the ML clouds as ice. Ideally, the goal of the MLANN should be to detect the most ML clouds as possible while minimizing false ML cloud identification.

Figure 3.

Layering classifications for a CloudSat-CALIPSO data matched CERES Aqua MODIS data for April 2009. Distributions of the various layering categories are shown for the different CERES-MODIS classifications. (a) Daytime, CM liquid, (b) Daytime, CM ice, (c) Night, CM liquid, and (d) Night, CM ice. M3: ΔZ_min = 3 km, M1: ΔZ_min = 1 km

Statistics within 3 x 3 confusion matrices are employed to help assess the MLANN SL-ML discrimination accuracy and reliability. Following the nomenclature of [9], the first column of each matrix lists the correct SL percentage, ST, with the false positive ML, MF, below. The sum of ST + MF, S the total percentage of true SL clouds, is at bottom of the column. The false positive SL, SF, is shown at the top of the middle column, the total percentage of positive ML clouds, MT, sits at the center of the matrix. At the bottom of the column resides the total CC ML fraction M, the sum, SF + MT. The estimated SL and ML fractions, ES and EM, respectively, occupy the top and middle cells of the third column. Normally, the lower right corner would contain the total fraction, but since it is always 100% in this study, the hit rate HR, or total fraction correct, ST + MT, is placed there. The real risk or chance of a misclassification is RR = SF + MF. The confidence in the estimates of SL and ML clouds are C_S = ST/ES and C_M = MT/EM, respectively. The number of correctly identified ML pixels is N_M = MT•N/100, where N is the total number of pixels comprising the matrix.

The MLANN was trained using the CM data for each of the four categories in Figure 3 to obtain a four sets of weights and constants for the two separation distances. Results of applying those coefficients to July 2009 data are summarized in the confusion matrices in Table 1 for ΔZ_min = 1 and 3 km on the left and right halves of the table, respectively. During the day, the 1-km hit rate HR, or percentage correct, is 75.5% for CM ice clouds, with the fraction of ML correctly identified being 0.42 and N_M ∼ 171,000. The layering discrimination during the day is a bit better for CM liquid clouds:

Table 1.

Confusion matrices (each bounded by dashed lines) for MLANN applied to Aqua MODIS relative to layer identification from CloudSat-CALIPSO, July 2009. The bold numbers indicate the percent correct for each matrix.

HR = 83.3% and N_M ∼ 295,000. For ΔZ_min = 3 km during the day, HR is better overall than its 1-km counterparts and the real risk is 15.2% for ice clouds compared to 24.5% for the 1-km case. However, MT is considerably reduced from its 1-km value and many more ML clouds with separation distances of 1 – 3 km are now identified as SL couds. RR is nearly the same, ∼16%, for both 1 and 3-km liquid clouds during the day. For the daytime 1-km data, CM ∼71% compared to ∼67% for 3-km data. At night, RR increases noticeably for ice clouds to 27.5% and 21.9% for 1 and 3-km data, respectively, while the hit rates drop from their daytime values, even though MT rises dramatically. For liquid clouds, MT drops slightly along with HR. Except for daytime 3-km data, the hit rate for ice clouds is signficantly smaller for ice clouds than for liquid clouds. This may be due to the overall larger optical depths of the upper layer clouds for systems identified as CM ice clouds [5]. The larger ice cloud optical depths may muddle the signals distinguishing between pure ice clouds and ML systems. This is especially true during the day.

Overall, the CM ice and water clouds must be considered together, so the results from the two phases were combined for the 1-km data and are listed in Table 2. The 3-km data are not considered anymore since it is clear from Table 1 that maximizing the ML detection is best conducted using the smaller separation distance. The 1-km MLANN detects 1.15 million ML pixels compared to 0.79 million ML pixels by the 3-km MLANN, while suffering minimally in increased false detection. The overall 1-km HRs are 80.4% and 77.1% for day and night, respectively, while the corresponding RRs are 19.6% and 22.8%. During the day, CS and CM are 82.6% and 70.6% compared to 84.7% and 68.5% at night.

Table 2.

Same as Table 1, but includes combined liquid and icer results for upper and lower layer separation ≥ 1 km only.

These results respresent a significant advance from the initial MLANN formulation [16], which did not consider the CM phase categories separately and did not include atmospheric humidity profiles. The changes increase the gap between the various metrics determined from the MLANN and those from other methods reported by [16]. As it is applicable both day and night, it should lead to a more accurate rendering of cloud vertical structure from satellite imagers.

4.2

Upper-layer cloud optical depth

Figure 5 shows the scatterplots of MLANN τ_ul versus the CALIPSO-CloudSat optical depths for the daytime training. When the CM pixel is classified as a water cloud (Figure 5a), the resulting values show some correlation but the origin of any linear fit would be near (0,0.1) and the slope would be fairly horizontal. A large portion of the points lay above the line of agreement. That holds true for the CM ice clouds (Figure 5b), but there appears to be better correlation. There is a significant difference between the range of optical depths for the two CM phases. For CM water, τ_CC is mainly less than 0.4, while for CM ice the significant range extends to τ_CC = 2. As noted above, this discrepancy is due to the fact that as the upper-layer cloud optical depth increases, the CM algorithm increasingly selects ice as the cloud phase [5].

Figure 5.

Scatterplot of MLANN upper layer optical depth versus CC optical depth for the daytime training, July and September 2009.

The distributions of the daytime optical depths and their differences are plotted in Figure 6. It is clear in Figures 6a and 6c that the MLANN is unable to capture the frequency of pixels having very small optical depths, τ_CC < 0.06, for both ML water and ice clouds. This inability to capture extreme values in cloud optical depth seems to be typical as it also appears in the results of [15] and [17]. Although the average CC optical depth, < τ_CC >, is the same as its τ_ul counterpart, the standard deviation of the differences, SDD, is roughly equivalent to 90% of the overall mean. The median differences for CM water and ice in Figures 6b and 6d are ∼0.04 and 0.01, respectively, while the difference histograms are highly skewed to negative values. The nighttime results are similar, but < τ_CC > is less than its daytime counterpart.

Figure 6.

Ice cloud optical depth and difference probability distributions for upper layer cloud of ML ice-over-water systems from CERES-MODIS (τ_ul) and CALIPSO CloudSat (τ_CC), July and September 2009.

4.3

Upper-layer Cloud Top Height

Daytime cloud-top heights determined by the MLANN for the training data are compared to their CC counterparts in Figure 7. The results are well correlated with the CC values for both the CM liquid (Figure 7a) and ice (Figure 7b) clouds. More extreme outliers are evident in the water cloud plot, which apparently balance a cluster of points that is below the 1:1 line at the high end of the range. The ice results are better behaved. Figure 8 shows the histograms of the heights and their differences from Figure 7. Again, the extrema are not well represented in the probability distributions in Figures 8a and 8c, while the number of points in the middle of the distributions are overestimated. The differences for daytime CM water clouds (Figure 8b) are skewed to positive values although the median difference is roughly -0.3 km. For CM ice clouds (Figure 8d) the distribution is less skewed and the median value is close to -0.1 km. The mean differences are 0.00 km and the SDDs are 1.38 and 1.03 km, respectively, for CM liquid and ice clouds. At night, the corresponding SDDs are 1.21 and 1.18 km.

Figure 7.

Daytime cloud-top height scatter plots for matched CM-MLANN and CC upper-layer clouds in ML cloud systems, July and September 2009, for CM (a) water and (b) ice clouds.

Figure 8.

Upper-layer cloud top height (Z_T) and difference probability distributions for for ML ice-over-water systems from CERES-MODIS and CALIPSO-CloudSat (Z_TCC), July and September 2009.

The results here far exceed the accuracy of the single-layer cloud-top height retrievals of most standard algorithms (e.g., [5]). They are also comparable to those from a different neural network approach [19] and may be somewhat larger than those from a dedicated cirrus cloud top height analysis [17]. The exact differences, however, are difficult to determine because of sampling and data selection differences. Nevertheless, it is clear that the skill in determining Z_Tul is sufficient to confidently begin the process of retrieving the properties of the upper and lower layers in ML conditions.

4.4

Upper-layer Cloud Base Height

The MLANN was also trained to estimate the upper-layer cloud base height Z_Bul using the same sets of input parameters. The preliminary results for nighttime are shown as scatterplots in Figure 9 for July 2009. Concentrations of the points along the 1:1 line are not as dense as those seen in Figure 7, but the correlation appears to be as good with most of points scattered about the agreement line. The larger SDD values, 1.6 km for CM water (Figure 9a) and 1.7 km for CM ice (Figure 9b), are not surprising, given that the cloud base might correspond to a single layer of ice or multiple ice layers with clear air between them. Nevertheless, the SDDs are smaller than those found for SL cirrus clouds using more conventional techniques [5]. Results for the daytime analyses are similar.

Figure 9.

Scatterplots of nocturnal upper-layer cloud base heights from CALIPSO-CloudSat and matched MLANN Aqua MODIS retrievals, July 2009.

With the upper-layer cloud top and base estimates, it should be possible define the upper-layer cloud boundaries for scenes identified as ML cloud systems. An example of applying the MLANN to characterize the cloud vertical extent is shown in Figure 10 for a cloud profile taken by CALIPSO-CloudSat over a midlatitude ocean at night, centered at 11:28 UTC, 8 April 2009. The top panel (Figure 10a) shows the cloud-top height retrieved using the standard SL assumption overlaid in blue over the CC profiles. It is clear that few of the CM-retrieved Z_T values coincide with Z_TCC for any of the ML clouds. Where the upper cloud is very thin, the retrieved top is close to that of the lower cloud, while in many instances, it falls between the layers and occasionally exceeds Z_TCC. The lower panel Figure (10b) shows the MLANN values of Z_Tul (black) and Z_Bul (blue) overlaid on the same CC cloud profile. For the ML clouds, Z_Tul tracks Z_TCC quite well. Some exceptions are seen at minutes 26.7 and 29.9, where the two top heights diverge by 1-2 km. Overall, the discrepancies are nothing like those seen in the top panel. Cloud base heights from the MLANN also coincide remarkably well with their CC counterparts. The greatest deviations occur around minutes 26.7, 27.8, and 29.2. Nevertheless, the base heights are better behaved than Z_T from the CM retrievals.

Figure 10.

Comparison of nocturnal cloud height parameters determined from CALIPSO-CloudSat data and from Aqua MODIS over midlatitude ocean area using the (a) CERES-MODIS SL approximation and (b) the MLANN coefficients determined here for Z_Tul and Z_Bul, 8 April 2009.