Thank you very much for your helpful comments. They were very focused and allowed us to elaborate on details in the manuscript and of the tobac package itself in greater depth, and we appreciated the opportunity to improve the manuscript.

• How does the “watersheding” decide on the time when the boundary of two features meets?

Watershedding is performed on a data field independently on each timestep, and as such does not incorporate the time dimension directly. The watershedding algorithm acts by “flowing” outwards from a marker set at the feature location, with the expansion rate governed by the magnitude by which a particular grid point exceeds the segmentation threshold prescribed. Thus, where the boundaries of two or more features will meet depends both on the size of the region being segmented and the magnitudes of the data being segmented.

• Will segmentation be twisted due to the vertical structure?

 Yes, tobac can accurately segment vertically tilted or irregular structures in 3D spatial data like a growing convective cloud in a highly sheared environment.

• In the detection of weather phenomena like atmospheric rivers, there are thresholds in geometric shape, size, and direction, how do these thresholds be applied in tobac? In your example, magnitude thresholds are used. But for atmospheric rivers, the magnitude threshold is the very first step. Could you explain more about the process of detecting atmospheric rivers?

As the reviewer points out, detection of atmospheric rivers (hereafter, ARs) typically requires more than a simple magnitude-based threshold detection. tobac is intended to be a flexible tracker, and as such isn’t specifically designed around any particular use case. However, tobac does provide a variety of options to allow users to tailor the feature detection to their particular phenomena of interest.

Size thresholds can be imposed for feature detection by prescribing the minimum number of contiguous grid cells that must exceed the magnitude threshold. A different value of this parameter can be set for each feature detection threshold, allowing for more flexible detection of different-sized features at different magnitudes.

Additional parameter settings and processes needed to detect ARs would likely depend on the particular case of use. One possibility is to follow a methodology similar to that used by Guan and Waliser (2019). ERA5 (they used ERA-interim) specific humidity and wind component fields can be used to calculate zonal and meridional components of Integrated Vapor Transport (IVT, see Equations 1a and 1b in Guan and Waliser, 2019). From these components, an IVT magnitude field and IVT direction field can be determined.

Guan and Waliser use a progressive application of percentile thresholds and geometry thresholds on IVT to identify coherent ARs. Percentile thresholds can straightforwardly be calculated from the overall ERA5 IVT field and used within the feature detection step of tobac. Imposing the geometry thresholding after the magnitude thresholding is more complex due to the cyclical nature of wind direction, but one approach would be to perform watershed segmentation on the IVT magnitude field based on the detected features. This segmentation field can then be used to analyze the IVT zonal and meridional components, as well as the IVT direction field within the segmentation mask. Guan and Waliser prescribed a poleward (meridional) IVT component >50 kg/m/s; more than half of IVT directions in the prospective area being within 45 degrees of the mean IVT; length > 2,000 km; and a length/width ratio > 2. These former two criteria can be assessed by applying the segmentation mask to the poleward IVT field and IVT direction field; and the latter two can be assessed by just examining the dimensions of the segmentation mask itself.

If all the above criteria are met, we have detected the presence of an AR feature that can then be spatiotemporally tracked using tobac. Otherwise, we would either discard this potential AR feature or follow the methology of Guan and Waliser to impose higher thresholds.

• The manuscript is a pleasant read and I congratulate the authors on the substantial improvements made to a promising, versatile tool for meteorological feature tracking. I just have a few minor comments as a community member.

Thank you for your kind words and thoughtful comments. They have improved the manuscript, particularly regarding our consideration of Atmospheric Rivers and discussion in the introduction.

• The manuscript did not mention whether the detection/identification of features (that is, the step before segmentation and tracking) can be based on multiple input fields (not just multiple thresholds for the same field) or can only take one input field in that step. For reference, in the case of atmospheric rivers (ARs), there are algorithms that rely on two or more of the following fields to detect the features: integrated water vapor (IWV), IWV transport, wind, precipitation, etc.

Thank you for your thoughtful comment. We agree with the reviewer that this would be extremely useful. Unfortunately, at this point in time, this version of tobac cannot be used on multiple input fields without multiple feature detection steps. However, implementing multivariate tracking while preserving our ability to remain variable-agnostic is already a key feature of our upcoming development plans for tobac, and initial results in the implementation of these capabilities are promising.

• Some of the improvements introduced in the manuscript have also been introduced in AR tracking algorithms, such as handling of separations and mergers, periodic boundary conditions, and grid agnosticity (Guan & Waliser, 2015, 2019).

Thank you for mentioning these publications. We were remiss in not citing these papers, and now include a description of their findings as well as the appropriate citations as shown below:

“Guan and Waliser (2015, 2019) have developed a tool called Tracking Atmospheric Rivers Globally as Elongated Targets (TARGET), which is designed for the detection and tracking of atmospheric rivers (ARs). TARGET includes techniques such as split and merger processing; periodic boundary condition treatments; and grid agnosticity; but can only be applied as presently designed to ARs.” (105-108)

• Given the number of studies dedicated to AR detection in recent years, including those contributing to the Atmospheric River Tracking Method Intercomparison Project (ARTMIP; Shields et al., 2018), some general discussion of AR detection might be helpful to provide more context and motivation for the development of tobac.

While we agree that AR detection and tracking are very important, we were reluctant to spend too much time expanding upon this particular use case due to the multitude of other atmospheric phenomena that can be detected with tobac, and the variety of improvements that needed to be discussed in the paper. Nonetheless, we have added a mention of ARTMIP and short discussion of the innovative techniques in TARGET:

“In recent years, there has also been a greater research focus on Atmospheric Rivers (ARs), including many existing within the Atmospheric River Tracking Method Intercomparison Project (ARTMIP; Shields et al. 2018). Guan and Waliser (2015, 2019) have developed a tool called Tracking Atmospheric Rivers Globally as Elongated Targets (TARGET), which is designed for the detection and tracking of atmospheric rivers (ARs). TARGET includes techniques such as split and merger processing; periodic boundary condition treatments; and grid agnosticity; but can only be applied as presently designed to ARs.” (103-108)

We have included our response to Reviewer 2 as a supplemental PDF in order to best preserve the figures and formatting in our response.

We have included our response to Reviewer 3 as a supplemental PDF in order to best preserve the figures and formatting in our response.