dc-l1-bangkok-attribute-coercion

Analyst notes

Description

Tests whether the agent recognises an attribute-type problem in a malformed vendor GeoJSON: every numeric column arrives as a JSON string, so client-side parseFloat in the dashboard silently returns NaN. The prompt deliberately names the symptom (NaN averages) instead of the operation (type coercion), the affected columns, or the int-versus-float split, so the agent has to inspect the file and work out what the defect is on its own.

Approach

1. Open the input GeoJSON and inspect a few features to see how each property is encoded.
2. Notice that every numeric column is serialised as a JSON string and that the dashboard's parseFloat fails on those.
3. Decide which fields are integer identifiers and which are floating-point measurements based on what the values look like.
4. Rewrite every numeric property as a proper JSON number, keeping the station_id as an integer and the measurement columns as floats.
5. Leave the Thai-script name field, any other string columns, and the point geometry untouched, and write the result back as GeoJSON.

Pitfalls

• Reading the file with GeoPandas and writing it straight back does not fix the defect, because the GeoJSON driver re-emits the object-typed columns as strings and the dashboard keeps producing NaN.
• Coercing the three measurement columns to float but leaving station_id alone is the most common partial fix, and it leaves the identifier column as a string which the grader treats as a clear miss.
• Casting station_id through float on the way to int turns the identifiers into 1.0, 2.0, ... and the JSON type subcheck for integers fails.
• Round-tripping the file through a non-UTF-8 stage like Shapefile or KML corrupts the Thai name_th values into question marks or mojibake.
• Reprojecting the points to a metric CRS to clean them up shifts every coordinate and changes the output CRS away from WGS84, even though the task is purely about attribute types.
• Dropping one of the required columns (station_id, name_th, sensor_value, pm25_ug_m3, elevation_m) on the way out fails Gate 1 outright.

Map

Recent runs task v3