Clasificación de cultivos en Ecuador utilizando técnicas de aprendizaje profundo con escasos datos

Abstract

CONDITION FOR PUBLICATION OF PROJECT. In Ecuador, the lack of labeled agricultural data is a major obstacle to long-term planning, land management, and the development of evidence-based public policies. This data scarcity hinders the generation of reliable land use maps, especially when aiming to understand the spatial distribution and temporal dynamics of key crops. This thesis tackles this challenge by evaluating different crop classification strategies using satellite imagery, with a focus on low-data scenarios. Three main approaches are compared: traditional models (e.g., Random Forest), Convolutional Neural Networks (CNNs), and Presto, a recent model in the literature. Various combinations of spectral inputs (RGB, 10 bands, and indices) and classification types (binary and multiclass) are tested and evaluated using metrics such as the macro F1-score. The results show that traditional models remain robust in low-resource settings, while CNNs offer modest improvements. Presto demonstrates a promising trade-off between accuracy and computational efficiency, enabling the generation of informative maps for agricultural monitoring. This work highlights that even in data-constrained environments, it is possible to adapt machine learning methods to improve crop mapping and support decision-making in the agricultural sector.