A Representative Sampling Approach to Few-Shot Transfer Learning

Abstract

Accurate plant growth estimation is essential in controlled
environment agriculture (CEA), where crop monitoring, yield prediction,
and environmental optimization all rely on reliable assessments. However,
this task remains challenging due to the limited availability of labeled
data. In this study, we propose a few-shot transfer learning framework
for estimating cucumber plant growth, using a multi-stage approach.
Our method combines deep feature extraction, representative image se-
lection, and regression to accurately predict cucumber plant height using
limited samples. By identifying key representative images, our approach
reduces data requirements while maintaining comparable predictive per-
formance. Our best model achieved a mean absolute error (MAE) of
20 mm with an R2 value of 0.68 using six representative samples. Ex-
perimental results demonstrate that systematic representative sampling
combined with regression significantly improves predictive performance
and stability in few-shot transfer learning scenarios, thereby minimizing
data collection efforts and reducing computational costs. Moreover, the
integration of clustering enhances explainability, making the approach
more transparent and suitable for practical deployment.