Advanced Deep Learning Architectures for Remote Sensing-Based Segmentation

DenseNet, InceptionNet, MobileNet, and ResNet are four state-of-the-art designs whose segmentation abilities are examined in this work using remote-sensed benchmark images. The primary objective is to identify algorithms proficient in fine object segmentation, anticipating their future applications in diverse domains. DenseNet121 scores better on Indian Pines with an overall accuracy of 99.7%, which is a result of its novel architecture that addresses gradient issues and encourages the reuse of features. Inceptionv3 excels on Salinas (99.913%) because of its unique modules catching characteristics at various sizes and reducing vanishing gradient concerns, showcasing its adaptability in hyperspectral data processing. Heterogeneous architectures lead to different performances of deep learning models on hyperspectral datasets. Long-range dependencies and diminishing gradients provide problems for ResNet50. Complex pattern extraction is where Inceptionv3 shines, and DenseNet121’s connectivity improves information flow potentially surpassing ResNet50. Especially in complex datasets like Salinas, Inceptionv3 outperforms because of its unique design, which incorporates factorized convolutions, auxiliary classifiers, various filter sizes, pooling layers, and other characteristics that enhance spatial and spectral feature acquisition. It is expected that the results of this study will contribute to the advancement of more efficient segmentation algorithms, meeting the needs of upcoming applications that depend on precise and detailed object segmentation in remote sensing scenarios.