From a methodology standpoint, this work advances previously published algorithms to assess POC and SPM from ocean color remote sensing to derive POC/SPM via a neural network based algorithm with thresholds used to define different oceanic particle types.

I am not sure I understand the motivation of this work. The authors develop a new algorithm to derive POC/SPM, which would seem to be a significant scientific advance for the community, but the final product is a general classification of water type, i.e., mineral dominated, mixed types, and organic dominated. What is the point exactly of the general classification? Is this to help inform application of other algorithms or global model configurations? As I mention below, the thresholds for this classification appear arbitrary. The intermediate product of POC/SPM would ostensibly seem more valuable from a science application perspective.

The NN POC/SPM algorithm presented in this work appears new, but there are almost no details given. Minimum details needed are: wavebands used, training data set used, and metrics used to determine valid results.

In 2.2 it is stated that 2 approaches are tested here, the POC/SPM derived from separate POC and SPM derivations via previously published NN algorithms and the new NN algorithm based on POC/SPM. But results do not specify which algorithm was used for what analyses.

The analysis associated with Fig 1 confused me in the context of study objectives. Thresholds are being set for POC/SPM to specify water types dominated by different particle types, but why is bbp/cp being regressed with POC/SPM? What does bbp/cp add? The objective in exploring these kinds of relationships, as the authors point out, is typically to used bbp/cp as a proxy for POC/SPM. I do not understand how bbp/cp is somehow influencing the decision on where to set thresholds.

Where thresholds are set, in fact, appears arbitrary. The authors state these thresholds are set “Based on our knowledge of the in-situ data points” but what does that mean? Setting of thresholds should be done quantitatively and independently of preconceived expectations in results. On the contrary, it appears the authors varied thresholds until results, i.e., maps of particle types for different regions, fit their preconceived bias. This is not a sound approach in my opinion.

Fig 2b looks like a shot gun target whereas the relationship in Fig 2a looks quite encouraging. In my opinion, 35% misclassified data for particle type in Fig 2b is quite poor. The authors state this is primarily driven by poor atmospheric corrections. Because these are match up data, why can’t the in-situ measured Rrs be used to assess algorithm performance independently of the atmospheric correction? In my opinion this should be done. And an important conclusion of the paper should be, apparently, that we need better atmospheric corrections.

As mentioned above, I would be more interested in performance assessments for the POC/SPM with associated error metrics such as mean absolute % error, mean absolute error, etc, tabulated. From the relationship in Fig 2a, it looks like the authors may be on to something! If the authors feel compelled to include a general (arbitrary) classification of particle type as well with global maps, OK, but the performance and applicability of the POC/SPM algorithm should be rigorously documented. 

The Authors present a neural network approach for characterizing coastal organic vs. inorganic fractions of suspended particulate matter using satellite ocean color radiometry.  Broadly speaking, the intent of the work is important and timely and the focus on application to a mission-long time-series (in this case MERIS) is appreciated and novel. I have one major concern, however, before recommending this work for publication, namely that the Methods are inadequately presented.  There was little-to-no information provided on the neural network approach (or its training dataset), the thresholds applied were not well defended, and important details were omitted here and there.  I recommend a major revision of the Methods, after which the manuscript could be suitable for publication.  All the best to the Authors for evaluating a satellite time-series of this size.  Specific comments/recommendations follow.  

Line 20:  Please provide units the first time a variable is introduced

Line 41:  Change to “related”

Lines 14-49:  This is a monster paragraph.  For improved readability, suggest breaking it up into several shorter paragraphs.

Line 51:  Change to “such as”

Section 2.1.1:  Are these data available to the community and, if so, where?  Also, a map and a table that more deeply describe the temporal and geophysical ranges for each sub-dataset would both be useful.

Line 80:  Was DS0 also used to train the neural network?  If not, what dataset was used for this?

Line 88:  Is this how MERIS data were processed for the full suite of spatial and temporal analyses, or just for the match-up exercise?  If the former, please make it clear this section describes ALL MERIS data processing methods.

Line 91:  Do you mean 400m?  If not, does your analysis require any additional explanation re: differences that could be realized in a shallow system where resuspension is meaningful vs. a system off a steep shelf break?

Section 2.1.2: This is a 3^rd in situ dataset, correct?  If so, as above, additional description would be useful.

Line 95:  Maybe change to “Two different approaches were initially tested to assess PPC.”?

Section 2.2: Please consider adding 1-2 sentences each for HA16 and TR19 to briefly describe how these two algorithms operate.  Also, please elaborate substantially on the neural network approach and its formulation. Is there a reference for the NN approach?

Line 108:  “correspond”

Line 114:  Please reword around “that is” or add a comma or semi-colon.

Lines 116-122:  How new thresholds were assigned reads as subjective. Please elaborate on how the in situ information (“based on our knowledge of the in-situ points”) allowed you to define the new thresholds.  For example, how did you define “mineral-dominated” in the field data?  As mentioned previously, a more complete description of the in situ datasets might help with this.

Line 138:  Are median absolute % difference, mean ratio, and slope calculated in log-transformed space or normal space?  Recommend presenting the equations for the first two.

Line 145:  Some points are only misclassified by a short distance, whereas others are much farther off the mark.  Is there a metric that can be reported to encompass this?  Thinking that since the thresholds are somewhat arbitrary (e.g., is there really a difference between 0.08 and 0.09?), you might want to take credit for “near-misses”.

Line 200:  Just reiterating that the PPC (neural network) approach needs to be better described within this manuscript.