Towards a Deep Learning Based Approach for Semantic Segmentation of Coca-Leaf Growing Regions in Satellite Images of Perú

Rodriguez R.M.; Llenque J.C.E.; Asato F.T.; Verde J.L.; Espiritu T.W.C.; Gonzales J.J.P. · 2024 7th IEEE Biennial Congress of Argentina, ARGENCON 2024 (2024) · DOI: 10.1109/argencon62399.2024.10735830

Resumen

The use of remote sensing to detect illicit crops is a trend that has increased over time. Regular in-person surveys are labor-intensive, time-consuming, expensive, and potentially life-threatening. Therefore, remote sensing techniques allow a small number of analysts to locate scattered sites of illicit crop production in large areas. The National Commission for Development and Life Without Drugs (DEVIDA), deliver a report on the monitoring of coca cultivation in the country on an annual basis. The report is based on the visual analysis of high spatial resolution satellite images recorded between July and November of the year prior to the submission of the report. The present study shows the first study based on deep learning for the semantic segmentation of coca leaf growing regions in Perú. For this purpose we use a U-Net architecture and SPOT-6 satellite images for the Pichari district - Cusco, Perú. We can conclude that the results are promising achieving an accuracy of 94.10% on the test image from 2019.

Instituciones: Pontificia Universidad Católica del Perú