Active Learning Strategies for Remote Sensing Estimation of Agronomic Parameters with Limited Calibration Data

This work introduces a new methodology to estimate vegetation parameters from remote sensing data exploiting machine learning techniques. The workflow uses appropriate active learning and regression strategies, based on ensembles, to significantly reduce the number of calibration data needed to run such algorithms. Experimental results, concerning the estimation of vegetation biomass and nitrogen content, will show that it is possible to reduce the calibration dataset up to 80% with a degradation of the estimation performance, against literature data, in the order of 5% for biomass and 15% for nitrogen content.