Detection of Areas of Deforestation in the Amazon Through Convolutional Neural Networks in the Period of 2020-2022

Gamarra-Moreno A.; Aroni-Ordonez L.; Ambrosio-Carrasco E.; Arias-Rafael S.; Huaman-Guizgueta J.; Baldeon-Huaccho Y. · Proceedings - 2023 49th Latin American Computing Conference, CLEI 2023 (2023) · DOI: 10.1109/clei60451.2023.10346180

Resumen

This study aimed to develop a deep learning application for detecting deforestation in the Amazon rainforest using satellite images. Two convolutional neural network (CNN) architectures, VGG16 and MobileNetV2, were evaluated, considering adjustments such as fine-tuning, k-means, and variations in hyperparameters using the Keras tuner. Utilizing a dataset of 6760 labeled images, a total of eighteen experiments were conducted. The results highlighted that with fine-tuning, 12 epochs, and the Adam optimizer, VGG16 achieved an accuracy of 93.474% on the test set, underscoring the effectiveness of deep CNNs in identifying deforestation in satellite images.

Instituciones: Universidad Nacional del Centro del Perú