Add Geopandas support

[mazzma12]

Hi,

Thanks for your amazing work, many custom function can now been deprecated and lots of keystroke are saved. If I may have a feature request it would be to support geopandas API for the geo plots.

If you are not familiar with this library, it inherits pd.DataFrame and embedd a custom geometry column that stores the geo object (Points, Polygone, Line ...).
It would be great if the plots could be done based on the geometry automatically, without casting points, or specifing the that you want polygons...

Tell me if you want more details about this.

[nicolaskruchten]

Sounds like an interesting idea! Can you tell me more about the API you would envision and the resulting output?

[mazzma12]

Glad you like it!

Basically, I think px could set rational defaults for different geom_types of the GeoDataFrame by using the API provided by geopandas.

Here is an example in pseudo-code just to showcase the API if you are not familiar with it

if isinstance(df, geopandas.GeoDataFrame):
    gdf = df # I know I am geo
    geom_type = gdf.geom_type.unique()
    
    if all(geom_type) == 'Polygon':
        # Treat as a polygon
    elif all(geom_type) == 'Points':
        # Set rational defaults
        lon = gdf.geometry.x
        lat = gdf.geometry.y # Mind the x, y
        bbox = gdf.bbox # Might be useful for zoom 
    else:
        NotImplementedError("Only Point and Polygon supported atm")

If I remember plotly uses geojson format for the API. In this case calling gdf.__geo_interface__ might be more advised instead of accessing the geometry property (at least for polygons)

More about the "geometry" column ehre

[nicolaskruchten]

OK cool, so what would you envision as a px API here? px.scatter_geo(line="geometry_column") or something like that? I'm not sure I see how geometry columns map onto the px or Plotly primitives at the moment...

[mazzma12]

I will try to give more details about the API using an example from the gallery show :

px.scatter_mapbox(carshare, lat="centroid_lat", lon="centroid_lon", color="peak_hour", size="car_hours", 
                  color_continuous_scale=px.colors.cyclical.IceFire, size_max=15, zoom=10)

I would expect that if the carshare is an instance of geopandas.GeoDataFrame with Point geometry types that the lat and lon column would be discovered by the method automatically so you just have to call :

px.set_mapbox_access_token(open(".mapbox_token").read())
px.scatter_mapbox(carshare, color="peak_hour", size="car_hours", 
                  color_continuous_scale=px.colors.cyclical.IceFire, size_max=15, zoom=10)

The same thing could be done if your geometry is of type Polygon.

[nicolaskruchten]

I see. I think that just grabbing a Point column is a bit too automatic for my tastes... what about when there are two Point columns? I could see a case for px.scatter_mapbox(point="point_col") for sure though.

[mazzma12]

Actually, the point of geopandas is to bring some structure to a DataFrame, with only one active geometry at a time. Hence it's rational to display this geometry by default, according to its geom_type.

There is no reason to have a geodataframe with one geometry active and wanted to display another one. If that occurs one can just use geopandas.set_geoemtry() methods that acts just like pandas.set_index(), or override the lon, lat kwargs in the plot method

[nicolaskruchten]

Ah I see, thanks for that extra bit of context :)

Seems like a reasonable and reasonably small thing to add... Any obvious downsides?

As a sidenote: right now px doesn't look at the index of a data frame at all, and I couldn't think of a good default behaviour for it in most cases. Any opinions?

[mazzma12]

Seems like a reasonable and reasonably small thing to add... Any obvious downsides?

I can't think of any

As a sidenote: right now px doesn't look at the index of a data frame at all, and I couldn't think of good default behaviour for it in most cases. Any opinions?

I encountered that problem several times in other circumstances. I haven't found a nice solution either, I usually ignore the index, and assume that one shall reset if one that uses it.

but I reckon it's boring to pass a column name and then realize it is an index. You could try to reset the index at the beginning (it's a copy anyway) but you'll have to deal with other problem such that potential duplicate in column name...

[mazzma12]

As a sidenote: right now px doesn't look at the index of a data frame at all, and I couldn't think of a good default behaviour for it in most cases. Any opinions?

I just met a use case where it might be useful: when you pass an instance of series to the scatter plot, you would like the default to assuming x is in index and y is the values. At the moment the only way I found to do it is a bit tedious :

1. Reset the index (will cast the Series into a DataFrame)
2. Pass the index name to x
3. Pass the column name to y

[nicolaskruchten]

OK, thanks for the input! Basically in certain cases (2d-cartesian plots) you would like the default value of x to be the data frame index? This basically precludes the notion of having multiple data points at the same x value, as index values must be unique, right? Also I don't think we can easily support multi-level indices just yet. (plotly.js supports 2-level axes for 2d cartesian plots but this isn't exposed in px at the moment).

At this point I don't think we're going to support passing in Series rather than DataFrames directly, as we need the column names all over the place for labelling.

[mazzma12]

Hi, I try not to derive too much on this as it is not related to this issue.

OK, thanks for the input! Basically in certain cases (2d-cartesian plots) you would like the default value of x to be the data frame index?
yes
This basically precludes the notion of having multiple data points at the same x value, as index values must be unique, right?
Not intuitive at first first glance, but indices are not necessarily unique in pandas
Also I don't think we can easily support multi-level indices just yet. (plotly.js supports 2-level axes for 2d cartesian plots but this isn't exposed in px at the moment).
Ofc I only intend a simple case with 1D index atm, just raise not implemented error instead
At this point I don't think we're going to support passing in Series rather than DataFrames directly, as we need the column names all over the place for labelling.

It's actually quite easy to grab the x and y column names from the series ts' by doing x=ts.index.nameandy=ts.name. Then when you call ts.reset_index()it will return a new dataframe object with columns [x, y]`

Happy to detail a bit longer in another post if needed :)

[nicolaskruchten]

OK so re indexes there's another issue here now #37 where I outline a different approach :)

[DmitriyG228]

That would be great to have geopandas support to be able to plot shapely Polygons. imho, I even better solution would be to enable this function in Plotly first

[andychang-1]

I would love to see this, I am drawing a map a zip code overlay as well as individual colored data points in Folium currently.

Folium can't even support >1000 points without clusters and I would heavily prefer to use plotly express for my task due to its way better speed.

[nicolaskruchten]

This is something I'm looking into in September! :)