Optimise plot images

[jfrost-mo]

They are each almost half a MB. They could easily be half the resolution, and then we could look into crunching the images harder, or maybe using a more efficient format like optimised PNG (no alpha), lossless webp, or even JPEG XL.

[jfrost-mo]

I've now run some experiments on 194 spatial pcolormesh map plots without resizing.

Method	Total Size (MiB)	Mean Size (KiB)	Relative Size	Lossless
Matplotlib PNGs	139.68	737.28	100.00%	Yes
Imagemagick no-alpha PNG	127.79	674.55	91.49%	Yes
oxipng	102.52	541.12	73.39%	Yes
Imagemagick JPEG	99.06	522.89	70.92%	No
Imagemagick webp lossless	88.77	468.54	63.55%	Yes
JPEG XL d0.5	84.02	443.48	60.15%	No
JPEG XL lossless	83.15	438.89	59.53%	Yes
pngquant	69.49	366.77	49.75%	No
pngquant + oxipng	56.07	295.93	40.14%	No
pngquant + JPEG XL lossless	40.70	214.82	29.14%	No

These experiments could do with rerunning with line plots included, and downscaled plots too, as the plots are pretty high resolution currently.

[jfrost-mo]

I'm hesitant to use a lossy algorithm, as I don't want to put the effort in to evaluate what the acceptable loss level is, and that we may sometime be interested in anomalous single grid points.

This gives us JPEG XL and webp as our contenders. webp has browser support, however JPEG XL can use a WASM decoder.

[jfrost-mo]

JPEG XL WASM decoder taken from libjxl, under BSD-3-Clause licence.

[jfrost-mo]

The WASM decoder seems to have issues with the SharedArrayBuffer. I think it needs a bunch of CSP headers set, so that is a no go. Maybe a single threaded one will work, but it might be too slow anyway. Alas, why can't JPEG XL be a default browser image format.

[jfrost-mo]

webp output from matplotlib is heavily lossy by default, however the pil_kwargs={"lossless": True} option can be added to savefig, which causes lossless mode to be used instead.

[jfrost-mo]

Something to consider is how many people want to do something with the images other than look at them, as webp will be less generally compatible than PNG.

[jfrost-mo]

Testing with some resizing to 800 by 800 (100 DPI):

Method	Total Size (MiB)	Mean Size (KiB)	Relative Size
Matplotlib 150 DPI PNG	107.49	759.11	100.00%
Imagemagick 100 DPI PNG	67.54	476.97	62.83%
Oxipng 100 DPI	56.84	401.39	52.88%
Webp 100 DPI	43.88	309.86	40.82%
JPEG XL 100 DPI	40.97	289.34	38.12%

Images were resized with Imagemagick, so it is possible that slightly different numbers would come from directly saving at 100 DPI. All compression was lossless.

[jwarner8]

Given user dpi arg to set outplot plot res, could probaly close this? I can't see users wanting other formats than png?

[jfrost-mo]

Honestly the most viable one would be recompressing the PNG files (the oxipng rows in the table). However the main issue with CSET output is the sheer number of files, which #2160 should resolve.