Google DeepMind launches domain-specific agricultural AI models for farmland mapping
Developed by Google's AnthroKrishi team, the models map field boundaries and track agricultural events in near real-time.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- Google DeepMind has launched two domain-specific agricultural AI models named ALU and AMED.
- The models utilize satellite imagery and machine learning to map field boundaries and track events on over 140 million hectares.
- The technology is expanding to 11 countries, including Japan, Vietnam, and Malaysia, to support sustainable farming.
Google has launched two 'India-first' domain-specific agricultural artificial intelligence models developed by Google DeepMind’s AnthroKrishi team. Designed to provide granular, field-level intelligence in data-scarce environments, the models—Agricultural Landscape Understanding (ALU) and Agricultural Monitoring and Event Detection (AMED)—leverage satellite imagery and advanced machine learning to transition from broad regional averages to individual farm insights.
The ALU model acts as a digital base layer by automatically identifying physical land patterns, mapping precise field boundaries, and detecting local water bodies and vegetation. The AMED model operates as a monitoring system, tracking time-based changes such as crop sowing, growth cycles, and harvesting events. By refreshing its observations approximately every 15 days, AMED provides local agricultural agencies with near real-time insights into agricultural productivity and land-use events.
The technology has already seen extensive real-world deployment, mapping more than 140 million hectares of Indian farmland in partnership with companies like TerraStack. The Karnataka Water Resources Department is currently utilizing the data across 2.6 million hectares to optimize public water allocation and irrigation management. Following its success in India, Google announced that it is expanding the deployment of these agricultural models to 11 additional countries, including Malaysia, Vietnam, Indonesia, and Japan.
This release demonstrates the value of training compact, highly specialized domain models to solve massive, real-world resource problems. By mapping individual smallholder farms that are often invisible to district-level census data, Google’s models allow local agencies and agritech companies to deliver highly targeted interventions—such as localized fertilizer recipes or water allocations—proving that custom AI systems can directly optimize regional food security and carbon-efficient farming.
Will other major technology companies release competing field-level agricultural monitoring models for smallholder farmlands by the end of 2026?
Still open. When the paper finds out, it will say so here and on the open questions page — including if it got this wrong.