Boosting Geospatial Land Use Mapping Accuracy via Self-Attention Mechanism-Based Fusion Transfer Learning Approach on Remote Sensing Imagery

For land use classification, the use of remote sensing imagery is essential in different fields, including land resource management, agriculture, and environmental protection. Advanced accessibility to higher-resolution RSI and the capability to collect multitemporal and source imaging from other geographical areas give new opportunities for numerous time-scale LC classifications. Recently, remote sensing using unmanned aerial vehicles has become a novel technology for studying Earth’s features and surface characteristics. The advantage of UAV-RS enables land cover mapping at various spatial resolutions and ranges. Nevertheless, computer vision-based RSI classification methods are required for monitoring the changes despite temporal, radiometric, spatial, and spectral resolutions. Recently, deep learning models have been presented in the domain of land-use RS mapping. In this study, a self-attention mechanism-based fusion transfer learning model for land use classification using remote sensing imaging (SAFTL-LUCRSI) technique is proposed. The main goal of the SAFTL-LUCRSI technique is to enhance UAV land use classification using RS images. Initially, bilateral filtering is applied in the image preprocessing stage to improve the image quality by removing unwanted noise. For the feature extraction process, the fusion models such as MobileNetV2, EfficientNetB0, and ResNet101 are employed. Furthermore, the SAFTL-LUCRSI model utilizes the self-attention mechanism-based long short-term memory (SA-LSTM) model for classification. Finally, the improved seagull optimization algorithm finetunes the hyperparameter values of the SA-LSTM model, resulting in better performance. The experimental evaluation of the SAFTL-LUCRSI approach is performed under the UC Merced Land Use dataset. The comparison analysis of the SAFTL-LUCRSI approach revealed a superior accuracy value of 99.00% compared to existing models.