Use of multispectral satellite images for the vegetation monitoring in the ecuadorian amazon using a u-net network in areas of industrial oil activity

The behavior of the vegetation dynamics in areas of oil industry activity in the Ecuadorian Amazon region was analyzed in order to quantify the ecological impact from 2013 to 2018. The implementation of deep learning in the measurement of plant coverage was set taking advantage of the spatial, temporal and spectral resolution of the Landsat 8 multispectral images. The bands of the visible and the near-infrared spectrum were analyzed by a semantic segmentation algorithm developed by labeling each pixel of the images with eight labels (wa-ter, forest, shrub, grass, wetlands, barren land, infrastructure, and clouds) and 2 classes (vegetation and no-vegetation). The region of study is characterized by high cloudiness which made it difficult to obtain satellite images for the training process of the convolutional neural network. Due to the lack of training data available, using a U-Net architecture represented an advantage because have an ability to learn in environments of low to medium quantities of training data like in the present study case. The obtained training overall accuracy was 0.996, and the validation overall accuracy of 0.989. The deep-learning semantic segmentation algorithm developed allowed analyzing the historical behavior of the vegetation cover in the study area between 2013 and 2018. The changes noticed in the zone of the Ecuadorian Amazon in the Blocks 31 and ITT located at the National Park Yasuní showed that from 2013 to 2018 the vegetation cover was reduced by 4.32% (99.17% in 2013, 94.85% in 2018). © 2021, Univelt Inc. All rights reserved.

Сборник материалов конференции
Univelt Inc.
  • 1 RUDN University, Academy of Engineering, Department of Mechanics and Mechatronics, address: 6 Miklukho-Maklaya Str., Moscow, 117198, Russian Federation
  • 2 RUDN University, Faculty of Ecology, Department of Geoenvironment, address: 6 Miklukho-Maklaya Str., Moscow, 117198, Russian Federation
Ключевые слова
Convolutional neural networks; Deep learning; Infrared devices; Near infrared spectroscopy; Petroleum industry; Semantics; Space applications; Space flight; Space platforms; Vegetation; Learning semantics; Multispectral images; Multispectral satellite image; Near infrared spectra; Semantic segmentation; Temporal and spectral resolutions; Vegetation dynamics; Vegetation monitoring; Image segmentation
Дата создания
Дата изменения
Постоянная ссылка

Другие записи