Deep learning-based land use characterization study

Land use and Land cover (LULC) is an important part of natural resource research by remote sensing technology. Efficient land use classification is essential for the conservation and management of natural resources. Traditional manual classification methods are time-consuming laborious and difficult to meet the dynamic classification requirement. For that, this study aims to develop an automated land use classification model. This model is developed based on dilated convolutions and classical U-Net architecture by training GF-2 data, which is named Dilated Convolution U-Net (DU-Net). The model is evaluated by the detailed classification and analysis of land use on Changxing Island and its northern surroundings using GF-2 data. Experiments show that the classification accuracy of the classical U-Net is 79% and the training loss rate converges to 60%, which makes it difficult to effectively support the task of feature classification. For the DU-Net, the classification accuracy is increased to 84%, the training loss rate is reduced to about 30%, and the Kappa coefficient reaches 0.71. Moreover, the classification accuracy for built-up land and farmland reaches more than 70%, while the accuracy for water reaches up to 97%. However, the recall is only 49% for farmland, for which model optimization is needed. Overall, the DU-Net significantly enhances the efficiency of the dynamic classification of tabular natural resources. It means the DU-Net can provide dynamic and efficient support for land use and natural resource management.