I noted a couple of things, the data lack substantial periods of low flow, it has perennial flow of typically periods where the stream is rising due to rainfall or snow melt, and falling due to dry conditions.  And compared to streams I typically work in, your peaks flow rates are not several orders of magnitude larger than low flows.  Your yellow line appears to be in frequency close to bankfull flow or channel maintenance flow that occurs every year or two  I would try to separate by periods when stream is rising, periods when stream is falling.  If you wanted to separate annual flood peaks into two categories, what you have is fine, except there are actually three categories, below the yellow line, between yellow and red lines and above red line.  If this is snowmelt system, you might want to divide storms by rainfall derived storm peaks and snowmelt derived storm peaks, and probably your worse peaks would be warm or intense rains on snow.  I would have to say though, there are plenty of other ways to analyze 20 years of streamflow data, perhaps wet years, dry years and average years (that would be three categories).  And if there was some activity in the watershed in the last 10 years, might that be connected to the more recent high peaks, or is it storm related or temperature related (what some try to suggest a climate change connection -- which is premature with this record).  In one of my analyses, it was to help identify boatable flows, so you could divide into fllow too high to boat, flows suitable and flows too low to boat (raft, kayak, etc.).  If you were a fish biologist, you might want to divide flows according to some fishery concerns.  If a geomorpholgist, dividing flows below bankfull channel and above bankfull on floodplain may be another way.

 Hi Philipp,

You pose an interesting problem of identifying similarity in flow behavior between two periods. But, I can't understand why you want only two periods.

My suggestion would be to divide your daily flow records into 2 or 3 year periods and construct the Flow Duration Curve for each of these periods. You may then compare the shapes of the flow duration curves or the flow quantiles (Q10, Q20, etc) between each period and try to identify similarity between them.

You could test for changepoints in the time series pretty easily in R (package changepoint), or you could calculate a suite of flow metrics (R package hydrostats) and then do some sort of ordination on these metrics to explore groupings in the data (i.e. which years cluster together). 

You might check out the software program "Indicators of Hydrologic Alteration" - developed by the Nature Conservancy. 

https://www.conservationgateway.org/ConservationPractices/Freshwater/EnvironmentalFlows/MethodsandTools/IndicatorsofHydrologicAlteration/Pages/indicators-hydrologic-alt.aspx

I am positive there are tools in that package for comparing flows between two different time periods, but I am not sure if there is a tool for identification of separate/similar regimes or periods. If not perhaps you could use one of the techniques described above to identify the periods or hydrologic difference or similarity, and then use the IHA software (or analogue in R) to quantify hydrologic variations between periods in more detail. Worth investigating perhaps.

OK, I think I see what you want, two time periods of different response for some reason, I do not know.  The rising stage of hydrographis respond to storm intensity, but the falling stages call the recession curve is now the watershed is functioning due to lack of rain after a storm, and to some degree, this can be affected seasonally due to growing or non-growing season and associated transpiration.  If you not a shift in the recession curve response that is not associated with growing or non-growing season, that might be a reason to separate into two periods. If you analyze land use activity during this period, and there has been some significant activities within the watershed during this period, that may be another separation possibility (ie, before and after major activity or increase in imperveous surfaces, installing a dam or water control structures).  Likewise, this could be done if you had other reasons to separate to look for differences.  However, along the Broad River in SC, USA, we had one year with about 8 bankfull events, including two extreme floods.  Flood frequency or severity is not something you can say, we had the 100 flood, so now we should be safe for awhile.  You could have another flood of the same magnitude this year or next.  The question you might want to determine is that is my watershed of interest responding to storms differently now that it used to, and if there is a difference, when was that difference noticable and then were there things going on in the watershed that could be connected to the change.  If you are thinking climate change, then you need long periods of record as climate can also be cyclic with wet and dry periods part of the natural system.  

So I would try to detect differences in the slopes of the hydrograph recession curves, which you might need to separate dormant and growing season to compensate for ET.

Some of the time dependent detail may be lost when looking at this graph, so if you have this in Excel, take the graph and extend it completely across the page to accentuate and expand the time detail, then the differences in the recession curves and dormant/growing seasons may be more noticeable. Rainfall-snow records may also have meaning to consider relative to this data.  

Otherwise, unless there is some reason to separate into two time periods, your choice may be arbitrary, and I would assume that is not something you want.  

Belatedly, another way to split the flow data into 2 time periods (at least here on the N. Am. Pacific coast) is by interdecadal cycle, which is relevant to wetness & thermal trends.

-Bob Vadas, Jr.