Deep Learning Semantic Segmentation for Land Use and Land Cover Types Using Landsat 8 Imagery

Using deep learning semantic segmentation for land use extraction is the most challenging problem in medium spatial resolution imagery. This because of convolution layer and multiple levels steps baseline network, which can cause a degradation small features. In this paper, algorithm comprises an adjustment network architecture (LoopNet) dataset proposed automatic classification using Landsat 8 The experimental results illustrate that (SegNet, U-Net) outperforms pixel-based machine algorithms (MLE, SVM, RF) classification. Furthermore, LoopNet convolutional loop block, superior to other networks U-Net, PSPnet) improvement (ResU-Net, DeeplabV3+, U-Net++), with 89.84% overall accuracy good results. evaluation multispectral bands demonstrates Band 5 has performance terms accuracy, 83.91% accuracy. combination different spectral (Band 1–Band 7) achieved highest result (89.84%) compared individual bands. These indicate effectiveness