Drone-and-AI System Aims to Forecast Potato Yields Without Digging Up the Crop

Potatoes present a persistent measurement problem for growers and researchers: the part of the plant that matters most for yield develops entirely out of sight. According to PotatoPro's coverage of the research, monitoring tuber development during the growing season has traditionally required destructive sampling — physically digging up plants to weigh what's underneath. That approach is inherently limited, since every sample destroys a section of the crop and can only be repeated so many times across a field before the data collection itself starts affecting the trial.

The system developed by the University of Tokyo Graduate School of Agricultural and Life Sciences and Kubota Corporation, carried out under the joint Kubota Todai Lab initiative, tackles this by never touching the tubers directly during the growing season. Drones equipped with RGB and multispectral cameras repeatedly photographed potato fields, and researchers extracted plot-level indicators — plant cover ratio, canopy height, color indices, and vegetation indices — from the imagery. Those above-ground signals were paired with actual underground tuber biomass measurements from field sampling to train a machine-learning model, which could then estimate tuber biomass in plots that hadn't been dug up at all.

The second piece is what turns a snapshot estimate into a forecast: the machine-learning outputs were fed into a Gompertz growth curve, an S-shaped model long used to describe biological growth over time. That let the team project tuber development forward and generate a pre-harvest yield estimate rather than just a mid-season measurement.

The research ran across the 2023 and 2024 growing seasons at the University of Tokyo Field Science Center in Nishi-Tokyo City, testing multiple plots with different planting densities and seed potato conditions. Per the research team's reported figures, tuber biomass estimation reached a correlation coefficient above 0.8, and final yield prediction using the growth curve model exceeded 0.7. Those numbers indicate the drone-plus-model pipeline tracked reasonably closely with what was actually dug up and weighed at harvest — enough, per the researchers, to confirm that underground yield can be estimated from above-ground observation and AI analysis.

The stated ambitions extend past a single research trial: the team points to applications in pre-harvest yield forecasting, cultivation management optimization, field monitoring, harvest timing recommendations, and broader AI-based crop phenotyping. Notably, the researchers also flagged that the same drone-plus-growth-curve approach could extend to other crops with underground harvestable organs — meaning the potato work functions as something of a proof of concept for a wider category of crops that share the same visibility problem.

The project brought together academic and corporate researchers: doctoral student Yuto Imachi, Professor Hiroyoshi Iwata, and Associate Professor Wei Guo from the University of Tokyo worked alongside researchers from Kubota Corporation's Next-Generation Research Department, Masahiro Okada of Sarabetsu Prediction Co., Ltd., and Pieter M. Blok, formerly a project assistant professor at the University of Tokyo and now at Eindhoven University of Technology.