| 01 Oct 2026
Dust detection algorithm for the EarthCARE Multi-Spectral Imager over the ocean
Abstract. Following the launch of the EarthCARE (Earth Clouds, Aerosols and Radiation Explorer) satellite in May 2024, the Multi-spectral Imager (MSI) supplements the vertical information retrieved from the Atmospheric Lidar (ATLID) and the Cloud Profiling Radar (CPR) instruments by providing information on cloud and aerosol properties in the across-track direction. MSI features a 150 km swath and a spatial resolution of 500 m across its four solar and three thermal channels. The operational cloud mask algorithm does not distinguish between clouds and aerosols, potentially leading to misclassifications of5 thick aerosol layers over ocean originating from dust storms, wildfires, or volcanic eruptions as clouds. Existing dust detection algorithms rely on spectral channels that are not available on MSI. This study therefore introduces a dust detection algorithm over ocean for MSI based on a random forest (RF) classification model that operates with the reduced spectral information available from the instrument. RF is a machine-learning approach that constructs multiple decision trees, whose outcomes are aggregated to generate accurate and reliable predictions. Pre-launch of the EarthCARE mission, a precursor dust detection10 algorithm was developed with Moderate-resolution Imaging Spectroradiometer (MODIS) data, using channels similar to those of MSI and adjusted to match MSI’s swath dimensions. Two case studies of dust storm outbreaks captured by MSI demonstrate the model’s capability and highlight its potential for operational use.
In the manuscript “Dust detection algorithm for the EarthCARE Multi-Spectral Imager over the ocean” by Walter et al., the authors test an application of the RF machine-learning approach to improve detection of dust plumes over the ocean.
The authors introduce and prove the concept, showing that improved aerosol/cloud classification enables dust aerosol retrieval rather than screening dust events as clouds.
The manuscript is clearly structured and well-written. However, two successful case studies may not be sufficient to prove the concept; additional examples of the RF application, such as an intercomparison of monthly AOD aggregates over the Central Atlantic using current cloud screening and the RF approach, could provide stronger evidence for the concept.
I suggest a minor revision to the manuscript.
COMMENTS
MSI-based cloud mask can be described in more detail.
A paragraph can be added on the discussion of cloud masking for cloud and aerosol retrievals.
Why not use TC from ATLID to discriminate dust episodes? Has RF dust classification been compared with TC?
P1, L5: The statement “operational cloud mask algorithm does not distinguish between clouds and aerosols” sounds a bit too critical.
P2, L34: in “Mxd35”, the meaning for “x” should be defined
P3, L69: I suggest adding “M-RGR product” and “M-CM product”, and specifying which variables from those products are used in the study; do the same for ATLID.
P4: What are the aerosol types west of the Sahara in Figure 2(b)?
P5, L96 : Have manual and RF thresholds been compared?
P9, Fig.5: Indeed, the BT11-12 difference is a typical dust test when measurements at those wavelengths are available.