The data for our Agricultural Landscape Understanding (ALU) model is refreshed roughly every 6 months. The Agricultural Multi-crop Identification (AMED) model data, which builds upon ALU, is refreshed more frequently, approximately every 15 days, to provide more timely insights.
ALU: High-Resolution imagery used for Google Maps procured from private vendors.
AMED: Builds on top of the ALU foundation, and leverages public satellite data from sources like Sentinel 1 & 2 and LandSat 8 & 9.
The outputs for both ALU and AMED are provided in a vectorised format, which means they are represented by shapes and polygons rather than a raster grid.
ALU: The underlying imagery for ALU has a resolution of 30-50 cm, from which features are inferred at a 1-meter resolution.
AMED: AMED uses satellite pixels with a resolution of 10-30 meters but infers and provides crop predictions at the field level, as defined by the ALU boundaries.
At launch, AMED covers 12 crops:
Bajra
Chilli
Corn
Cotton
Gram (Bengal Gram and Chickpeas)
Groundnut
Mustard
Rice
SorghumSoybean
Sugarcane
Wheat
We will continue to expand the crops that we cover. Please follow the API page for the latest updates.
Our definition of fields (and other landscape features like water bodies and trees) is described in our ALU ArXiv paper. To simplify, satellite based segmentation is performed based on visual cues, so the differences are based on separation of adjacent fields like gaps, bunds, trees, cropping pattern changes etc. Satellite imagery doesn’t directly provide any ownership-related signals.
Our external collaborators at IIT Bombay (Professor Milind Sohoni) have used ALU to digitise land records for the Department of Land Records (DoLR) of Maharashtra. The ArXiv paper also contains these details in the appendix authored by IIT Bombay.
Both the Agricultural Landscape Understanding (ALU) system and the Agricultural Multi-crop Identification (AMED) model undergo comprehensive validation processes, combining field-level assessments with comparisons against large-scale official data. See the respective ALU and AMED papers on arxiv for details.
ALU validation includes rigorous on-ground validation in collaboration with external partners. Aggregate analysis of ALU's output against the Agricultural Census Data showed comparable results in plain topography but some divergence in hilly and coastal regions.
AMED validation primarily relies on two dimensions: field-level predictions using labeled data and aggregate predictions against national crop census data. Model's performance is significantly better in the winter (Rabi) season compared to the monsoon (Kharif) season. For aggregate evaluation, AMED's predictions were compared against India's national crop census. The model demonstrated an overall agreement of 94% in the winter season and 75% in the monsoon season.
We err on the side of providing comprehensive data by default. To fine-tune the output for your specific use case, you can use the confidence scores provided with both ALU and AMED predictions.
For ALU: Partners have used a combination of size, confidence, and shape parameters to determine the quality of a landscape feature for their specific needs.
For AMED: In addition to the ALU parameters mentioned above, you can use the AMED crop prediction confidence, season length, and predicted crop history for a given field to filter the output.
The API request can accept a location in the form of lat, lng or s2 cell id. The response is a GeoJSON of the corresponding s2 cell. See the API documentation for more information.
S2 cells are a clever way to create a grid for the entire surface of the Earth, solving the problem that you can't wrap perfect squares around a sphere. It works by first projecting the globe onto the six faces of a cube and then continuously dividing those square faces into four smaller ones. This process can be repeated until the squares are roughly 1km x 1km (level 13), which is what we use for the API. Each of these tiny squares gets a unique ID, which acts like a highly specific zip code. For computers, searching for these simple number IDs is much faster than using latitude and longitude, making it perfect for apps that need to quickly find nearby landscape features.
Prerequisites
Make sure you have Google Cloud project and billing set up. Use this guided flow to enable and use all the necessary services.
Provide the following information to the Google team:
Google Workspace Customer ID (GWCID). See Find your customer ID for instructions on how to find it.
The email address of the person who will enable the API for your Google Cloud project.
Wait for the Google team to add your email and GWCID to the API allowlist.
Set up
Enable the API in your project by visiting Agricultural Understanding API.
Create an API key. See Create an API key for more details.
Currently the APIs are available for use in India, however, geographical expansion to other global regions is in the roadmap. Kindly stay tuned for updates regarding the same.
Both ALU and AMED APIs are free to use the API.
There is no rate limit for maximum number of API calls.
