u-net convolutional neural network

Deforestation remains a significant concern regarding climate change and biodiversity conservation. At the same time, the development of new image processing techniques and wide access to high spatial and temporal resolution satellite imagery have created unique conditions for monitoring deforestation. This is particularly important in areas such as the Arasbaran Biosphere Reserve. Existing methods for monitoring deforestation rely on a combination of visual inspection, spectral profiles, statistics, and machine learning techniques. Given recent advances in image processing using Convolutional Neural Networks (CNNs), this study aims to evaluate the performance of a refined U-Net architecture for identifying forest cover to monitor deforestation in multi-temporal satellite images. In this regard, a deep learning model for monitoring deforestation in the Arasbaran Biosphere Reserve based on the classification of Landsat satellite images from 2000 to 2022 was developed. In this study, the Normalized Difference Vegetation Index (NDVI) was used to create masks, which were then visually corrected. Additionally, the refined U-Net model presented was compared with Random Forest and Artificial Neural Network (ANN) models. The results showed that the refined U-Net outperformed traditional methods in classifying images into forest or non-forest categories, resulting in an overall accuracy of 96.53%, a kappa coefficient of 91.55%, an F1 score of 94.68%, and an IoU of 90.04%. The proposed model can accurately show forest changes and estimate the amounts of forest area increase or decrease. Overall, it was observed that the forest area in Arasbaran increased during the 2000-2022 period. This research indicates that using a refined U-Net can be an effective tool for sustainable forest resource monitoring and management