Classification of Remote Sensing Dataset Imagery Using Deep Learning Techniques

Integrating deep learning with remote sensing is contentious due to the disparity between natural and remote sensing images, raising the question of whether deep learning techniques can transform the remote sensing domain. This study covers prior research employing deep learning methodologies for land use and land cover categorization in the last five years. Our study is categorized into three groups based on the data type utilized in each analysis. The initial group comprised roughly 20 prior studies of multispectral data, the subsequent group contained 8 studies of hyperspectral data, and the final group encompassed 10 studies of aerial photos obtained via drones or radar. Each data type is physically distinct and serves a specific purpose. Furthermore, we chose one of the studies and implemented it with the specified data to categorize land cover with the CNN technique, employing a 4-band CNN patch size of 5x5. An accuracy of 95.0% was achieved, whereas a 4-band CNN with a patch size of 9x9 attained an accuracy of 91.5% on the training data for a region in Pakistan. The accuracy for the test data in Lahore City was 93.6% for the 4-band CNN with a patch size of 5x5 and 91.1% for the 4-band CNN with a patch size of 9x9.