Google DeepMind Expands Farm AI Models From India to Africa and UN's FAO

Google DeepMind is expanding two artificial intelligence models originally built for Indian agriculture into additional countries across Africa and Asia, while the United Nations Food and Agriculture Organization (FAO) has begun using their data to update global crop statistics. The news was published on Google's India blog on August 26, 2026.

Two models, one goal

The models were built by the AnthroKrishi team within Google DeepMind. Agricultural Landscape Understanding (ALU) identifies field boundaries, water bodies and vegetation from satellite imagery using machine learning. Agricultural Monitoring & Event Detection (AMED) provides field-level information, including which crops are being grown, when they were planted and when they were harvested, refreshing the data roughly every 15 days.

Both systems were trained on three years of historical data and satellite imagery, allowing them to automatically detect crop type, field size and planting timelines without manual ground inspection. The blog post was written by Alok Talekar, who leads agriculture and sustainability research at Google DeepMind.

Farm credit and advisory services

Several partners in India are already using the models. Sugee.io is embedding data from the ALU API directly into its agricultural lending system to speed up loan applications and improve credit quality and compliance for banks. The company also plans to use AMED to monitor credit risk and track events that could affect a farmer's ability to repay a loan over the life of the agreement.

Krishi DSS, a government system run by Amnex, uses the models to monitor crop health, estimate planted acreage and advise policymakers on irrigation. The Council on Energy, Environment and Water (CEEW) is building crop diversification analysis on top of this data and integrating it with its own climate platform.

FAO and global farm statistics

The most significant part of the expansion is the collaboration with FAO under the geoAI4stats initiative. The UN agency plans to fold ALU and AMED data into its CROPGRIDS platform to improve the accuracy and update frequency of agricultural statistics shared with member states. The new capabilities are meant to automate crop-type detection, generate agricultural maps and speed up the release of statistical data.

The geoAI4stats project received $2.5 million in support from Google.org through its AI Collaborative: Food Security program. The funding is intended to help countries use AI foundation models and local datasets for climate resilience planning, crop insurance, fertilizer subsidies and monitoring agricultural sustainability.

Scale of deployment in India

The scale of use within India itself is already substantial. The Agriculture Data Exchange in the state of Telangana supports more than 5 million farmers, while the Karnataka Water Resources Information System manages irrigation across 2.6 million hectares. Terrastack has analyzed more than 140 million hectares of farmland nationwide, showing that the models have moved well beyond pilot status and now operate as national-level infrastructure.

The expansion into Africa and Southeast Asia means the same models must now serve farmers under climate conditions and cropping systems very different from those they were originally trained on. Google has not yet disclosed details on model accuracy outside India or the number of farmers covered by the program in Africa.

What it means for global agriculture

For developing countries where reliable crop data is often scarce, tools based on satellite imagery and machine learning could replace costly field surveys. The integration with FAO means data from Google's models may feed into official statistics that governments use to plan agricultural policy, insurance and subsidies.

At the same time, the growing role of private AI models in farm lending systems, as with Sugee.io, raises questions about who is accountable when flawed risk assessments affect a farmer's access to financing. Google has not yet disclosed details on auditing or accuracy verification for the models in the new regions.