The paper presented a source reconstructing procedure by first locating the source location before estimating the emission rates. A machine learning method has been used in the first step to locate the source location. The overall results are quite interesting and encouraging.

However, there are several shortcomings in this manuscript. Some of the statements are not accurate and some terminology uses are also questionable. The presentation of the machine learning method is not easy to follow for those who are not quite familiar with the same method and software. In addition, the dispersion model errors affect the results but are not sufficiently considered or discussed.  It is also a concern that the method is only tested with a single set of experimental data.  More test cases are probably needed.

Major points:

Title: Both “Generalized” and “Spatiotemporally–decoupled” are not accurately reflecting the current two-step method.  The word “non-constant” in the title does not sound appropriate either.  In reality, there are rarely constant releases.  The author should reconsider the title.

Abstract, lines 12-14: This statement is not accurate.  The temporal variation of the release rates may be reflected on the plume shape, not only on the temporal variations of the observations.  In theory, some problems cannot be decoupled.  So the proposed method cannot be a real general framework.  The limitation of the method has to be pointed out in the paper.

Abstract, line 15: Locating a source location is not “localization”.  This needs to be corrected throughout the paper.

Abstract, line 18: A relative error of about 50% for the Oct. 4 total release is probably not deemed “accurate”.  It is better to present the results more objectively with the actual number listed in Table 3.

Line 94: The authors seem to suggest that the correlation-based method only applies when constant-release assumption is made.  This is not accurate.  Constant release is only one assumption that reduces the complexity of the problem.   If the release starting time or duration is not known.  Such assumption may not be enough to guarantee a unique solution of the source location.   On the other hand, if a source is not constant, but the release time period and temporal profile are known, it is probably easy to get the source location even without the constant-release assumption.

Line 108: It is wrong to assume a square matrix.   The dimensions of the observation and source vectors are independent and rarely the same.

Lines 122-124: The statement is not correct.  The emissions combined with the meteorological conditions together determine the concentrations at any given measurement site, including the peak values and its timing.

Line 228: It is very confusing to use “sample” for the different candidate source locations.

Lines 287-288: The hyper-parameters used in the 50 runs should be given in the supplementary document.

Figure S3: What is the sliding window applied here?  It does seem to be a sided window rather than centered one.  Please explain this in the paper.

Minor points:

Line 46, T3-10: Please explain what T3-10 distributions are.

Line 60: What does “deterministic assumption” mean? It is quite confusing.

Figure 1: What do the different shapes and colors in the diagram mean?

Equation (7): Please explain all the parameters here.

Line 160: Why is the amplitude quantity called “wave rate”?

Lines 160-161: The median value is not a central moment.

Line 232: If it is 40^th percentile, the number of samples for Oct.3 and Oct. 4 should be 1200 and 600.

Line 237: The authors probably mean 80^th, 60^th, 50^th, 40^th, 20^th, and 0^th.

Line 238: “A lower percentile” should be “a higher percentile”.

Figure 8: No shade appears for the Bayesian inversion results in the lower left panel.

Line 416:  What do the various pre-screening ranges refer to?

Figure S1: Should it be 20% instead of 10% for the five-fold cross-validation?

Table S1: Brief descriptions of the hyperparamters should be provided.