melt season about over, sniff. fixating on extent, volume, area? Better: physical state of the ice, its 2D distribution and behavior with weather.
As noted above by Neven and Fish, the season looks about over in the Beaufort/Chukchi/Wrangel-arm sector, except for ice displacement by wind which continues through Sept 5th. However the action between the pole and the Barents front will continue for well into the fall because the surface water warmth is anomalously warm.
Along the lines of Shared's proposal, it may be better to look this winter at methods that retain the distribution map of the property being studied, rather than represent the state of the Arctic Ocean by single numbers (eg extent) which have discarded positional information.
Since daily maps of sea ice concentration are available for five years (2012-2016), I looked at methods for comparing whole maps.
First, the Sept 5th images were brought forward in time to a minimum binary state where each grid cell is classified as either open water or ice, done here by setting the whiter pixels to ice and the bluer to water. This was supplemented by a mildly 'greedy algorithm' that captures and converts isolated pixels of one type surrounded by a neighborhood of the other.
The Venn diagram below shows all the water/ice combinations for the five years. This a 5-bit number (32 cases). For example the very center represents that part of the Arctic that always has ice at the end of season (in the context of the last five years).
It is quite difficult to devise a satisfactory color palette for the operation used (multiplication) to fill out all the venn compartments so the images below just compare 2016 to the other four years. This operation works as (upper layer)*((lower layer)/255) on each channel which has the effect of darkening the resultant pixel.
The second image lays out the 32 venn colors in a speadsheet (miraculously drag-n-drop moves gimp palette colors to spreadsheet cell fill) and the third images constructs its (complex) multiplicative palette as 32x32 = 1024 color tabs. Gimp provides 20 logics by which layers can interact and ImageJ another 12; each of these leads to a different cross-palette.
Similarly 2x2 palettes are embedded in the upper right corner of each of binary comparison with 2016, ie a pixel in the 2012-2016 has one of four colors depending on whether both years had ice there, neither, or just 2012 or just 2016.
Quite a bit of distributional variation can be seen at the end of season in the final animation. Outlying ice, if it freezes in place amidst new first year ice, will affect the progression of the 2017 melt season. Consequently, dropping map information in favor of single numbers (integrals of the various scalar fields) will seriously impair forecasts just based on these daily numbers.