Autores: Pavel Coronado, Pedro Achanccaray

Resumen

The Earth’s population growth has continuously increased the demand for agricultural production. Consequently, acreage and crop yield information have become increasingly important. Techniques based on satellite images are one of the most attractive options for agricultural monitoring over large areas. In this context, this work proposes a methodology for estimating the production of agricultural crops. Our study area is the La Libertad department, located in the northwest of Peru. The proposed methodology is based on convolutional neural networks (CNN) trained on high-resolution remote sensing imagery. The CNN is trained in an active learning cycle to assist in the labelling of new areas, to increase the dataset, and to improve the model’s performance. The dataset used to train the CNN models has been developed in cooperation with the Ministry of Agrarian Development and Irrigation (MIDAGRI), which conducted field visits and surveys on a statistically representative sample of plots. Our experiments demonstrated that the proposed methodology is able to recognize different crops, even with only a few labelled parcels. Finally, the trained model is used to increase the number of labelled parcels for further improvements.

Expositor

Pavel Coronado

Doctor of Philosophy in Government and Politics and Master of Arts in Economics from the University of Maryland at College Park. Master's in Economics from the Pontifical Catholic University of Rio de Janeiro, and Bachelor's in Economics from PUCP. His main areas of interest are economic development and political economy. Methodologically, he has focused on applied microeconometrics and artificial intelligence applied to the Social Sciences.  He is a professor in the Department of Economics at PUCP and the first Head of the Laboratory of Artificial Intelligence and Computational Methods in Social Sciences at PUCP-QLAB.