Latlong Geocoding API: now 4× more accurate than Google Maps

The history of Indian astronomy (ज्यौतिष) is both fascinating and illuminating. It is also the story of focus on measurement, accuracy and continuous improvement, over several millennia. This approach is summarised beautifully by Parameswara, a celebrated astronomer and mathematician:

"कालान्तरे तु संस्कारः चिन्त्यतां गणितोत्तमैः।" — "With time, Refinement (in measurement) needs to be planned by good metrologists."

I am excited to announce the results of our latest R&D efforts: Latlong's Geocoding API is now 4× more accurate than the corresponding Google Maps API for Indian addresses.

The New Benchmark for Accuracy

In 2023, in a benchmarking study verified by IIT Kanpur's National Centre for Geodesy, we showed that Latlong's API was 2× more accurate than Google Maps for Indian addresses. Our current benchmark, performed on an enhanced test bed 1.6× larger and with a far more diverse set of addresses (especially rural ones):

• Latlong's Average Error: 297 m
• Google Maps' Average Error: 1,280 m

For a like-for-like comparison on the original 2023 test bed, Latlong's average error of 231 m compares favourably to Google's 1,067 m.

The Journey from 2× to 4×

Three things stand out this time:

1. Revamping parsing technique: We refined our "string parsing" method for tokenization, focusing on "max area matches through polygon chain". This made the algorithm faster and more tolerant of 'address fuzziness'.

1. Dealing with 'fuzzy area delineation': One of the challenges with Indian addresses is that area boundaries (like pin code or localities) are not precise. Our algorithm addresses these challenges directly.

1. Far more data: We have more than doubled the number of Points of Interest (POI), helping us precisely geocode a higher percentage of addresses, with 58% of test results geocoded within 100 m. We added granular data for almost every single one of the 6.6 lakh villages in India.

Our Hypothesis on Geocoding

Most platforms start by solving for the "autocomplete / search" function and then try to adapt that logic to geocoding. Core to our methodology is to treat "geocoding as a spatial match problem." We believe we got this fundamental piece of the spatial algorithm right from the beginning.

Shreyas Bharadwaj, SVP of Varahe Analytics, said: "Latlong's Geocoding and location APIs have been instrumental in transforming our campaign communication strategy. By enabling hyper-local precision, we've been able to identify and engage specific voter clusters with issue-based messaging that truly resonates."

The 4× gap is a testament to the state-of-the-art quality a 30-member company in south Bengaluru is capable of delivering. We invite you to experience our APIs firsthand at apihub.latlong.ai.