Land Use and Occupation Detection in Satellite Images Using Deep Learning: Evaluation of LinkNet and D-LinkNet Architectures

EasyChair Preprint 15670

Abstract

Land cover detection in remote sensing images can contribute to several applications in the scientific and social spheres. The development of techniques to perform this detection was driven by the improvement of deep learning algorithms, such as convolutional neural networks. Thus, the present work used two neural network architectures, LinkNet and D-LinkNet, applying different backbones to perform land cover segmentation in high-resolution satellite images. The LinkNetB7 architecture, with the EfficientNet-B7 backbone, demonstrated greater sensitivity and accuracy, providing a more refined segmentation of land cover compared to the other models, reaching an accuracy of 0.91, sensitivity of 0.93 and IoU of 0.84.

Keyphrases: Aprendizado Profundo, Backbones, Redes Neurais Convolucionais, Segmentação de Imagens de satélite, Sensoriamento Remoto

@booklet{EasyChair:15670,
  author    = {Álisson de Oliveira Alves and Luisa Christina de Souza and Luiz Eduardo Nunes Cho Luck and Raniere Rodrigues Melo de Lima and Carlos Augusto Teixeira de Moura and Wesley José dos Santos Marinho and Rafael de Medeiros Mariz Capuano and Bruno Cesar Pereira da Costa and Marina de Siqueira and Arthur Diniz Flor Torquato Fernandes and Jesaias Carvalho Pereira Silva and Pablo Javier Alsina},
  title     = {Land Use and Occupation Detection in Satellite Images Using Deep Learning: Evaluation of LinkNet and D-LinkNet Architectures},
  howpublished = {EasyChair Preprint 15670},
  year      = {EasyChair, 2025}}