Research on Landscape Feature Recognition Based on Three-branch Fusion with Convolutional Neural Network

Normally, Land Use and Land Cover (LULC) refer to different physical areas, such as open water or forest, while land use reflects human activities on the land. However, the existing classification models based on Machine Learning (ML) and Deep Learning (DL) methods have drawbacks in classifying land cover areas due to high-scale images. To overcome these limitations, a Three-branch fusion– Convolutional Neural Network (TRI-CNN) method is proposed for improving landscape feature recognition of land use mapping and land covers. The EuroSAT dataset is utilized and by preprocessing the data is increased in certain classes. Then, features from different landscape images are extracted by proposed TRI-CNN model that extracts relevant features for recognizing land use. At last, classification was executed by Softmax layer of the proposed model that effectively classifies according to classes. The proposed approach attained a high accuracy of 99.4%, recall of 99.1%, precision of 99.3, % and f1-score of 99.3 % which is greater than existing deep learning models such as CNN, 1 Dimensional – CNN (1D-CNN), Lightweight Wide Residual Networks (LWRN), Deep CNN-ResNet-50, and SHapley Additive exPlanations (SHAP).