In this study, we focus on the South-East Pacific (SEP) region (100–70° W, 30–10° S), the main oceanic basin for the occurrence of marine stratocumulus (StCu). Statistics are computed for the months of August and September, when StCu coverage reaches its seasonal maximum occurence over the SEP . The analysis periods depend on satellite data availability: 2003–2025 for MODIS (Terra and Aqua satellites) and 2020–2025 for ABI observations (GOES16 for the 2020–2024 period and GOES19 for the year 2025). The combination of these long-term datasets and the focus on the major StCu season provide a statistically robust framework for characterizing mesoscale cloud organization.

To investigate the mesoscale organization of StCu, we rely on three complementary satellite datasets covering the same SEP domain. Readers may refer to the “Data Availability” part for references of the data sources:

MODIS Liquid Water Path (LWP).

We use the Level-2 (L2) Liquid Water Path (LWP) product from MODIS instrument onboard the Terra and Aqua satellites. The native spatial resolution of LWP is 1 km. For consistency with GOES observations, MODIS LWP fields are interpolated onto a 2 km grid. MODIS provides two snapshots per day over the SEP, at approximately 10:30 and 13:30 Local Solar Time (LST) (respectively Terra and Aqua overpass times). The MODIS dataset (2003–2025) serves as the reference for characterizing mesoscale convective cell regimes.

While MODIS provides a long-term climatological perspective with two daily overpasses, the use of geostationary observations from GOES-East allows us to go beyond this snapshot-based framework. By exploiting the high-frequency sampling and infrared measurements, we investigate the full diurnal variability of stratocumulus mesoscale organization, including nighttime conditions that are not accessible from passive daytime LWP products. Table  summarizes general information related to the four satellite datasets used in our study.

Figure  shows a SEP situation through LWP and ΔTb data. A weaker spatial contrast is observed in the ΔTb field compared to LWP. In particular, the cellular convective systems are more clearly perceived in the LWP product, where variations in liquid water content enhance the visibility of mesoscale structures. In the south-western corner of the scene, cirrus clouds can be identified. These high-altitude clouds are only weakly perceived in ΔTb, likely because of their really small optical thickness contrasting with the underlying stratocumulus. In the LWP product, however, regions of cirrus clouds, as well as their surrounding environment and optically thin clouds appear as missing values in the dataset. Their values will need to be carefully handled in the subsequent calculations, as detailed in the methodology Sect. .

This full region (both for MODIS and ABI data) is divided into sub-domains measuring about 300km×300km, with a 150 km overlap between two adjacent sub-domains, in order to maximise the use of the data. Hereinafter, these sub-domains will be referred to as “cloud scenes”. As a result of our choice to divide the scenes in pixels rather than in lon-lat grid, the 300km×300km cloud scenes may not appear perfectly square in a geographic projection (examples in Fig. ). However, we have verified this slight geometrical bias has a negligible impact on our computations and analysis we describe in the following.

The full IR GOES-East (Table  last row) analysis represents a total of 15 389 772 cloud scenes from 366 d analysed (August and September between 2020 and 2025). However, to ensure that the analysis focuses exclusively on marine stratocumulus cloud scenes, a set of exclusion criteria was applied to remove scenes featuring continental lands or dominated by high-level clouds. All selection criteria were applied at the pixel level within each 300km×300km subdomain.

As we mention before, our work is based on previous MODIS-based studies, that used LWP data and a spectral approach to determine the characteristic spatial scale of the convective systems . Using these former results as a reference, we use the same methodology and apply the spectral computation to the products referenced in Table .

We use a spectral approach to perform a scale decomposition of the variance of a scalar field X (where X=LWP or ΔTb), in order to characterize the typical horizontal scales of stratocumulus mesoscale cellular organization. This method allows us to retrieve the dominant wavelength λ associated with the maximum contribution to the variance, which is interpreted as the characteristic size of the mesoscale cloud cells .

The spatial variance of the scalar field X of a scene is expressed as: 1σX2=∫kmin⁡kmax⁡PX(k)dk, where PX(k) represents the distribution of variance across spatial wavenumbers k, between kmin=2π/L with L=300km and kmax=2π/Δx, Δx=2km.

We therefore analyze the power spectral density (PSD) from this spectral decomposition, referred as PX(k) in the following.

To compute these spectra, we use the cloudmetrics toolbox , which gather various metrics to quantify cloud spatial organization. This spectrum computation performs a two-dimensional Fourier transform of the scalar field and reduces it to a one-dimensional radial spectrum by averaging over the azimuthal component. The justification of this reduction to 1D is that most of the variability in cloud field can be retrieved from this radial component.

To reduce small-scale spectral noise and to facilitate the identification of dominant modes, the radial spectrum is averaged into discrete logarithmically spaced wavenumber bins, uniformly distributed in log⁡10(k) between kmin and kmax from Eq. (). Characteristics of horizontal scales of the convective system are associated to the dominant peaks or local maxima of this binned spectrum. For each detected local maximum, a wavenumber interval bounded by the two neighboring bins is defined. A second-order polynomial is fitted to the native PSD values within these three binned spectral values, and the wavenumber corresponding to the maximum of the fitted polynomial is retained as the refined spectral peak. The associated wavelength is finally computed as λpeak=2π/kpeak and is associated to the cell scale λcell, called λ hereafter. If several local maxima are detected, only the one corresponding to the highest wavenumber k (smaller wavelength) is kept for this study.

This two-step approach combines the robustness of bin-based peak detection with the precision of a continuous fit applied to the original spectrum, and reduces the dependence of the detected λ on the arbitrary choice of binning. Given the finite domain size, the minimum resolved wavenumber is kmin⁡=2π/L, and a limited number of structures can be represented at large wavelengths. For wavelengths approaching the size of L, the number of detectable cells decreases rapidly (fewer than four cells for λ≥L/4 for instance), leading to increasingly discrete spectral modes and reduced statistical robustness. To avoid domain-size artefacts and sampling limitations at large scales, we therefore restrict the analysis to cell wavelengths smaller than 80 km.

Figure  depicts four contrasted cloud regimes taken from the situation of Fig. , together with their associated power spectral densities computed from the LWP and ΔTb fields. For clarity, the spectra shown correspond to normalized PSDs: each spectrum is divided by the variance of the corresponding scalar field (P^X=PX/σX2), which allows a direct comparison of the dominant spatial scales λLWP and λΔTb.

For all four regimes, a well-defined spectral peak emerges in both LWP and ΔTb, from which the characteristic cell scale λLWP and λΔTb are retrieved. The consistency between the two estimates across such contrasted scenes highlights the robustness of our spectral approach in capturing the spatial scale of the cloud cells, regardless of their organization mode.

Nevertheless, we notice λΔTb tends to be slightly larger than λLWP. We attribute this systematic overestimation to the weaker cloud–scene contrast in the ΔTb field compared to LWP. We believe reduced contrast likely misses small-scale variability and limit the detection of the smallest organized structures, thereby shifting spectral energy toward slightly larger wavelengths.

To be noted, the first two scenes are assigned a blue L label in Fig. , as they correspond to low clouds with a detected λ below 80 km. The third scene is labeled with a red L because, although a λ is detected, it exceeds 80 km. The fourth scene is also assigned a red L, as no λ is detected.

Beyond the illustrative examples shown in Fig. , the robustness of our method and our choice to use ΔTb=Tb12.3µm-Tb10.3µm, is moreover supported by the comparison performed over the full dataset (Fig. ). When considering all scenes for which both LWP and ΔTb are available and both λ under 80 km (representing a total of 1 762 287 scenes ≈11.45% of the dataset), the two variables exhibit a satisfying and statistically robust linear relationship.

To account for variability of both variables, we perform an orthogonal distance regression fit, that yields a slope of 1.24 and an intercept of 6.95 km. The slope larger than unity and the positive intercept indicate an overestimation of the characteristic scales when derived from the ΔTb field. This behavior is consistent with the reduced small-scale contrast in ΔTb (observed in Fig. ), which likely smooths fine-scale structures and redistributes spectral energy toward larger spatial scales.

Although the best-fit line departs from the 1:1 relationship, the overall dispersion remains structured along a well-defined linear trend, indicating that both computations consistently capture similar variability in cloud organization across regimes. Thus we consider λ by ΔTb as reliable proxy for characterizing the mesoscale organization of marine cloud cells when LWP retrievals are unavailable.

When it comes to stratocumulus organization, another key metric to investigate is the aspect ratio (AR) associated to the cell scale. AR is a dimensionless variable, defined as the ratio of λ to the mean depth of the marine boundary layer HMBL: 2AR=λHMBL. Over the Southeast Pacific stratocumulus region, low cloud tops are assumed to be representative of the MBL depth, as the cloud layer is mainly contained in the boundary layer and overshooting above the inversion is negligible . Therefore, we define HMBL from the CTH field with a low-level mask and averaged over each scene (0<CTH≤3000m).

This definition implies that our estimate of HMBL – and consequently of AR – directly depends on the characteristics of the cloud top retrieval. MODIS and ABI instruments rely on different retrieval schemes and have different spatial resolutions, that can lead to differences in the CTH field and thus in the estimated MBL depth. Since AR scales linearly with 1/HMBL, any bias or variability in the cloud top product propagates directly into the AR values.

In ABI, the estimation of the MBL depth is primarily based on the CTH of low-level clouds. For MODIS observations, such retrieval based on CTH can be biased in partially cloudy scenes, as thin clouds and cloud edges tend to artificially lower the average CTH. To account for this, we follow the approach of , which adjusts the cloud top temperature (CTT) sampling based on the scene's cloud fraction to retrieve a more representative CTH, using the parameterization described in and . In the following, this method is used for MODIS as it improves how MBL is estimated in broken cloud regimes. More discussion on this in the Appendix .

Table  gives the number of retained scenes and confirms the robustness of the statistical framework. These wavelengths and their associated aspect ratios (ARLWP and ARΔTb) constitute the key variables used in the subsequent analyses

Figure  presents the mean daytime evolution (2020–2025) of the characteristic cell scale λLWP and the aspect ratio, ARLWP.

The mean daytime evolution of λLWP exhibits a clear three-phase structure. During the morning, cell size increases steadily with rising solar radiation, with an upscale growth rate of approximately 1.17 kmh-1 with ABI and 1.07 kmh-1 with MODIS. The characteristic scale grows from about 30 km at 08:00 LST to nearly 37 km around 13:00–14:00 LST. This coarsening phase is followed by a quasi-stationary plateau in early afternoon that seems to be followed by a slight decrease later in the day. The transition between the upscale and downscale phases occurs around 14:00 LST, slightly after the maximum in incoming shortwave radiation, indicating some inertia in the cloud-field response.

ARLWP is underestimated when derived from ABI than from MODIS. This offset is more pronounced for Aqua overpass time, suggesting that the cloud top height field, used for MBL depth estimation, is more sensitive to retrieval differences and sensor characteristics than the cell scale spectral computation. Despite this difference in magnitude, the temporal evolution remains highly coherent between the two datasets.

The aspect ratio displays a remarkably similar temporal structure. ARLWP increases from values near 20 in the early morning to about 27 around midday. The observed growth rates are 0.98 h-1 with ABI and 2.09 h-1 with MODIS. Both λLWP and ARLWP consistently capture the same sequence: morning increase, early-afternoon stabilization, and weak late-day reduction. This indicates that the upscale dynamics coherently modulate both the horizontal scale and the geometric organization of the cloud field. This is further supported by their similar relative changes (+23% for λ, +29% for AR), while the relative change in MBL depth remains negligible (-2%). Together, these results point to a diurnal evolution of the AR more linked to the cloud dynamism itself rather than the MBL variability.

Before further analyses, particular caution is required near sunrise and sunset, when it comes to LWP-based results. λLWP and ARLWP are derived from cloud properties retrieved in the UV–visible spectral range. At low solar zenith angles, rapid changes in illumination conditions, enhanced three-dimensional radiative effects, and increased sensitivity to shadowing introduce substantial retrieval biases. These effects can lead to artificial variability or systematic underestimation/overestimation of cloud properties during dawn and dusk transitions. Consequently, we consider these large uncertainties at the edges of the daytime window as radiative artifacts rather than real dynamical evolution of the cloud field.

We next examine the interannual variability of these daytime evolutions, shown in Fig. . For λLWP, all years display the same daytime phasing: a steady morning growth, a mid-afternoon maximum with a short plateau phase, followed by a small decrease. The timing of these phases is nearly invariant across years. Interannual differences mainly affect the mean value of the cycle. The afternoon maximum varies by about 3–5 km between years, with 2024 exhibiting slightly smaller values and 2021–2023 lying at the upper end of the range. This spread remains weaker than the total diurnal amplitude (about 10 km), demonstrating that daily solar forcing dominates over year-to-year variation.

This daytime pattern robustness is also reflected in the inferred upscale growth rates (Table ). ABI-based yearly rates range from 1.07–1.27 kmh-1, whereas MODIS-based estimates are less clustered, between 0.88 and 1.65 kmh-1. Although the magnitude of the growth slightly varies from year to year, the growth phase consistently occurs within the same local solar time window, confirming the stability of the diurnal phasing.

The 2020–2025 period encompasses contrasted climatic variability modes, including a prolonged La Niña, from 2020–2022, followed by a strong El Niño event in 2023–2024, together with a transition of the Pacific Decadal Oscillation (PDO) toward its positive phase around 2023. In the South-East Pacific stratocumulus region, these modes strongly modulate sea-surface temperature, lower-tropospheric stability, and the MBL depth. We might expect these thermodynamical adjustments to modulate the characteristic horizontal scale of the cellular pattern (λLWP) and its organization, as the stratocumulus cover mirrors the circulation and physics within the boundary-layer depth. La Niña conditions typically enhance coastal upwelling and inversion strength, favoring a shallower and more stable MBL, thus supporting smaller and more compact structures. Conversely, traditional El Niño and PDO-positive conditions promote warmer SSTs, reduced stability, and a deeper/decoupled boundary layer and are expected to favor larger and potentially less organized cells .

However, such a clear monotonic relationship is not evident in our dataset. The ordering of λLWP amplitudes across years does not follow standard El Niño–Southern Oscillation (ENSO) phases, and years characterized by different large-scale states display comparable cell sizes (Fig. ). Importantly, as shown in Fig. , the 2020–2025 period reveals a very small interannual variability, except for the two EL Niño years 2023 and 2024, where the MBL depth reaches a notably low altitude in 2023 and decreases even further during the peak of the event in 2024. Moreover, despite ABI and MODIS estimation offsets–driven by differences in CTH field retrieval (see Appendix for details) – the diamond markers show consistent interannual variability and minimal daytime evolution between Terra and Aqua overpasses.

These observations go along with the smallest λLWP values of the series, contrasting with the larger, fractionated structures typically expected under El Niño conditions. This apparent paradox can be explained by specificity of El Niño 2023–2024 event, that is characterized by an the atypical large-scale thermodynamic change. Recent studies have demonstrated that this anomaly was majorly driven by internal ocean dynamics rather than traditional local trade wind-SST coupling . These dynamics led to enhanced convective activity in the equatorial Pacific, intensifying the large-scale subsidence over the SEP subtropical gyre. This mechanical and thermal compression from above significantly strengthened the Estimated Inversion Strength (EIS), acting as a powerful atmospheric “cap” that prevailed over the buoyant forcing from the warm local SST anomalies observed . Consequently, this persistent upper-level warming and enhanced subsidence forced the MBL to remain exceptionally shallow and compressed for the 2023–2024 period. This atypical thermodynamic regime provides a physical explanation for the 2023 and 2024 outliers, with decreased values in MBL depth observed in Fig.  and smaller cellular scales captured in Fig. . This seems to indicate the particular nature of an ENSO event can prevail over traditional large-scale expectations.

Interestingly, this interannual gap almost vanishes when considering the aspect ratio. The ARLWP curves are nearly superimposed throughout the day, with differences generally within about one unit. Hence, while λLWP retains a measurable interannual amplitude change, ARLWP exhibits a strong year-to-year stability. This dispersion of λLWP remains small compared to the total diurnal increase (approximately 6–8 units) (Fig. a). Both the morning growth and the early-afternoon plateau occur at nearly identical local solar times across all years. While MODIS exhibits larger fluctuations in diagnosed growth rates than ABI, the overall temporal structure remains unchanged (Table ).

As ARLWP remains remarkably stable from year to year, this indicates that while the horizontal scale of organization λLWP may undergo disparities linked to interannual thermodynamical variability, the proportional relationship between horizontal scale and vertical MBL depth remains largely invariant, highlighting the robustness of the AR regarding internannual variability.

The daytime evolution of λΔTb and ARΔTb closely mirrors that inferred from λLWP and ARLWP. Both mean cycles (2020–2025) exhibit a robust three-phase structure: a steady morning upscale growth, a short early-afternoon quasi-plateau, and a rapid late-afternoon downscale transition (Fig. ). As expected, the artefacts we could observe near sunrise and sunset are not seen anymore, when using this infrared-based index. Both upscale rate of λΔTb and ARΔTb are followed by a substantially faster downscale rate (twice faster for both metrics): λΔTb upscale and downscale of respectively 1.28 kmh-1, of 2.31 kmh-1 and ARΔTb speeds reaches 1.08 and 1.79 h-1 respectively (Table ). This confirms a systematic asymmetry between growth and decay phases, with a faster evening dynamics. As for LWP, λ and AR evolve coherently in time, indicating that both horizontal scale and geometric structure are sensitive to the same dynamical forcing.

After sunset, λΔTb rapidly stabilizes around 36 km and remains nearly constant until sunrise, while ARΔTb settles around 25 with minimal variability. The strong daytime dynamics followed by a quasi-steady nocturnal regime further emphasizes the dominant role of solar forcing in driving mesoscale cloud organization. This remarkably stable nighttime observation thus add a fourth phase to the common shape diurnal evolution of the metrics. Overall, ABI/GOES-East observations enable continuous day–night monitoring of mesoscale cloud organization and confirm that ΔTb provides a robust proxy for capturing both the dynamical transitions. The likeness of LWP- and ΔTb-based daytime evolution make the nighttime observation trustworthy.

This day–night contrast of ARΔTb dynamics is also perceived from a spatial point of view. The climatological maps in Fig.  show that the large-scale structure of ARΔTb over the South-East Pacific is remarkably similar between nighttime (00:00–06:00 LST) and afternoon (13:00–17:00 LST). We observe a general value in the center of the bassin (25 at night and 30–35 in the afternoon) with an intensification of the gradient near the coast, where persistent StCu overcast decks show higher values of ARΔTb up to 45. These organization patterns and spatial gradients are preserved, indicating that the geometric structure of the field is robust to the diurnal cycle. However, a clear amplitude increase is observed during the afternoon, consistent with the temporal evolution described above: the afternoon-minus-night difference highlights a general amplification of ARΔTb at the bassin-scale rather than a spatial change in organization, with a stronger enhancement near the coast.

These results indicate that the diurnal cycle primarily modulates the intensity of ARΔTb while leaving its large-scale spatial organization mainly unchanged, supporting the idea of ARΔTb as a dynamically constrained field responding to radiative forcing .

All years show the same temporal patterns, with nearly identical timing of the morning growth onset (around 08:00 LST) and the late-afternoon transition (around 17:00 LST) (Fig. ). As previously found for λLWP, interannual variability primarily affects the amplitude rather than the phase of the cycle. Upscale rates for λΔTb range from 1.08–1.40 kmh-1, while downscale rates vary between 2.03 and 2.44 kmh-1 (see Table ). Although slightly larger than in the LWP-derived metric, this dispersion remains small compared to the total diurnal amplitude (of order 10 km), confirming that solar forcing dominates over year-to-year variation.

Here again, a clear contrast emerges between the cell size and the aspect ratio. While λΔTb testifies of a moderate interannual spread in magnitude, this dispersion largely cancels out when considering the normalization by the MBL depth. ARΔTb upscale rates vary only between 0.92 and 1.21 h-1 and downscale rates between 1.61 and 1.93 h-1. The diurnal curves from all years almost overlap (Fig. b). This mirrors our LWP-based results: interannual variability is primarily expressed in the absolute horizontal scale, whereas the aspect ratio remains tightly constrained over the years.

Finally, as observed in Sect. , although ENSO-related changes in sea surface temperature and inversion strength are expected to modulate MBL depth and thus the horizontal scale of organization, the observed year-to-year variability in λΔTb remains difficult to link with a clear, traditional ENSO-phase dependence. However, consistent with our findings for λLWP, the cell sizes λΔTb for 2024 stand out as a clear outlier with distinctly lower values. This anomaly aligns with the atypical nature of the 2023–2024 El Niño event discussed previously; the intensified large-scale subsidence and enhanced inversion strength (EIS) effectively compressed the MBL from above, preventing the spatial expansion of cells traditionally expected during warm ENSO phases. Interestingly, this structural distinction becomes minimal when considering the aspect ratio ARΔTb consistent with our results for ARLWP. This suggests that while the structures of the cellular patterns were vertically and horizontally constrained by the enhanced capping inversion, their proportionality remained largely conserved.