Improving Feature Learning in Remote Sensing Images Using an Integrated Deep Multi-Scale 3D/2D Convolutional Network

Developing complex hyperspectral image (HSI) sensors that capture high-resolution spatial information and voluminous (hundreds) spectral bands of the earth’s surface has made HSI pixel-wise classification a reality. The 3D-CNN become preferred approach because its ability to extract discriminative while maintaining data integrity. However, datasets are characterized by high nonlinearity, features, limited training sample data. Therefore, developing deep methods purely utilize 3D-CNNs in their network structure often results computationally expensive models prone overfitting when model depth increases. In this regard, paper proposes an integrated multi-scale 3D/2D convolutional block (MiCB) for simultaneous low-level high-level feature extraction, which can optimally train on strength proposed MiCB solely lies innovative arrangement convolution layers, giving (i) simultaneously convolve with features; (ii) use multiscale kernels abundant contextual information; (iii) apply residual connections solve degradation problem increases beyond threshold; (iv) depthwise separable convolutions address computational cost model. We evaluate efficacy our using three publicly accessible benchmarking datasets: Salinas Scene (SA), Indian Pines (IP), University Pavia (UP). When trained small amounts data, is better at classifying than state-of-the-art used comparison. For instance, achieves overall accuracy 97.35%, 98.29%, 99.20% 5% IP, 1% UP, SA respectively.