Enhancing Forest Image Segmentation in Aerial Images using Deep Learning

Forest cover segmentation is required for monitoring and managing forest ecosystem, yet due to the heterogeneous and complex nature of forest landscape it poses several challenges. This study explores and evaluates the performance of latest deep-learning architectures, focusing on semantic image segmentation using DeepResNet-U architecture. We utilized an extensive dataset of HD aerial images having detailed and diverse forest cover information. After training, testing and validating our model we determined their accuracy and efficiency. A comprehensive analysis demonstrated an impressive performance of DeepResNet-U model with an IoU of 0.7813, dice loss of 0.1229 and overall accuracy of 0.8771. The proposed method turned out to outperform other standard models due to the incorporation of residual connections and hence proved to be a better approach to improve feature extraction and gradient flow.