A Deforestation Detection Network Using Deep Learning-Based Semantic Segmentation

Abstract

Semantic segmentation is an important task in which the class label of each pixel is predicted. Thus, it is quite tough compared with classification and classical segmentation. Recently, deforestation has been a serious environmental issue, as it causes numerous environmental concerns: climate change and ecological loss. Hence, it is essential to recognize deforestation to save the environment. In this work, an efficient convolution neural network (CNN) model is proposed to identify deforestation in the Amazon Rainforest more precisely. At first, two modified SegNet methods are presented to make the semantic segmentation more effective. More importantly, an efficient semantic segmentation framework is proposed by integrating the merits of ResNet18-based modified SegNet, ShuffleNet-based modified SegNet, and UNet to yield more effective segmentation. Moreover, the employment of computationally faster ResNet18 or ShuffleNet in modified SegNet leads to improvement of computational efficiency and semantic segmentation performance. Thus, the proposed framework also retains advantages such as residual learning, skip connection, pointwise group convolution, and channel shuffling, which are responsible for making the optimization easier and the network efficient and faster. In addition, a Laplacian-of-Gaussian-based modified high boosting filter (LoGMHF) is employed for deblurring, edge enhancement, and denoising. The experimental analysis also shows that the proposed framework outperforms others.