Land use Land Cover Classification using Convolutional Neural Networks

Building accurate Land Use Land Cover Maps to identify various land resources, such as forest areas, water bodies, agricultural land, urban areas, etc. have become essential. The information acquired from these Land Cover Maps can be used for various environmental applications and infrastructure planning. Creating automated techniques to generate maps using remotely sensed data using different computer vision and engineering techniques is of great importance. Many attempts have been made to develop Land Use Land Cover classification maps for different study areas. In this research work, we have developed a method to generate a seven-class Land Use Land Cover map using Quantum GIS-tool and the recent Convolutional Neural Network Unet-DenseNet121. Unet deep learning architecture backed with DenseNet121 backbone model is trained to perform classification task on 111,180 patches generated from the study area. The model gives an average Overall-Accuracy of 94.09%, whereas the Matthews-Correlation-Coefficient is 0.66 across all LULC classes, which indicates excellent model performance.