Deforestation detection in Amazon rainforest through deep learning module TerraNet using Sentinel-1 image.
- Paper number
GLOC-2023,T,IP,x75306
- Author
Mr. Saleh Nabiyev, Azerbaijan
- Year
2023
- Abstract
Worldwide forest destruction has been steadily increasing over the last 50 years, causing a number of significant problems for humans, animals, and the ecosystem. Satellite images are a reliable source to identify and verify the extent and the level of forest loss. Observation of forest and other natural landscapes and analysis of changes occurring through active sensors are widely used. Recently, various remote sensing solutions have been proposed for the study of natural and anthropogenic impacts in the Amazon forests. When using optical satellite images we can face cloud cover over rainforest causing some difficulties for observation. In this regard, the application of Radar images can fill the gaps created during optical observation. Mütəmadi SAR təsvirlərinin əldə olunması və analizi operativ olaraq hadisələrin müəyyən olunmasına imkan verir. Change detection will be applied to detect deforested areas which appear with a lower radar backscatter signature compared to forested area. The presented project dedicated to determining the areas with destroyed forest cover through the TerraNet module developed by us and working on the basis of artificial intelligence algorithms. We are using a modified version of U-Net, a fully convolutional neural network architecture introduced by Ronneberger et al. for medical image segmentation. In general, a U-Net-like architecture consists of an encoder path to capture the context and a symmetrical decoder path that enables precise localization. The encoder consists of multiple blocks, each containing convolutions, and pooling layers, while decoder consists of upsampling layers and convolutions. Before determining the deforested areas with different classification methods, the SAR images pass (several) preprocessing stages as filtering and etc. Changes in the forest area from 2018 to 2023 have been observed and change detection results were described with the map. Deforestation determined based on the proposed methodology was compared with the results of Sentinel 2 optical satellite imagery. The difference between the results is not more than 10%.
- Abstract document
- Manuscript document
(absent)