I'm using plotly's Scattermapbox to overlay a map with a shaded image of polygons created by datashader's shade function (based on https://plotly.com/python/datashader/), but the projections do not seem to align, see picture below. Any suggestions how I can overcome this problem using plotly's Scattermapbox and datashader?

Reproducible example:

import geopandas as gpd
import plotly.graph_objects as go
import spatialpandas as spd
import datashader as ds
from colorcet import fire
import datashader.transfer_functions as tf

# load data
world = gpd.read_file(
    gpd.datasets.get_path('naturalearth_lowres')
)
# world = world.to_crs(epsg=3857)
# create spatialpandas DataFrame
df_world = spd.GeoDataFrame(world)
# create datashader canvas and aggregate
cvs = ds.Canvas(plot_width=1000, plot_height=1000)
agg = cvs.polygons(df_world, geometry='geometry', agg=ds.mean('pop_est'))
# create shaded image
tf.shade(agg, cmap=fire)

shaded image

# create shaded image and convert to Python image
img = tf.shade(agg, cmap=fire)[::-1].to_pil()

coords_lat, coords_lon = agg.coords["y"].values, agg.coords["x"].values
# Corners of the image, which need to be passed to mapbox
coordinates = [
    [coords_lon[0], coords_lat[0]],
    [coords_lon[-1], coords_lat[0]],
    [coords_lon[-1], coords_lat[-1]],
    [coords_lon[0], coords_lat[-1]],
]

fig = go.Figure(go.Scattermapbox())
fig.update_layout(
    mapbox_style="open-street-map",
    mapbox_layers=[
        {
            "sourcetype": "image",
            "source": img,
            "coordinates": coordinates,
        }
    ]
)
fig.show()

overlayed map

I read that Scattermapbox only supports Mercator projection which I found confusing as the examples in plotly's documentation seem to be in long/lat format, but I tried converting the coordinates of the GeoDataFrame to epsg 3857, see

# world = world.to_crs(epsg=3857)

The results is that the shaded image becomes invisible. Any help would be highly appreciated.