Abstract
Convolutional Neural Network (CNN) is achieving remarkable progress in various computer vision tasks. In the past few years, the remote sensing community has observed Deep Neural Network (DNN) finally taking off in several challenging fields. In this study, we propose a DNN to generate a predefined High Resolution (HR) synthetic spectral band using an ensemble of concurrent Low Resolution (LR) bands and existing HR bands. Of particular interest, the proposed network, namely DeepSWIR, synthesizes Short-Wave InfraRed (SWIR) band at 5m Ground Sampling Distance (GSD) using Green (G), Red (R) and Near InfraRed (NIR) bands at both 24m and 5m GSD, and SWIR band at 24m GSD. To our knowledge, the highest spatial resolution of commercially deliverable SWIR band is at 7.5m GSD. Also, we propose a Gaussian feathering based image stitching approach in light of processing large satellite imagery. To experimentally validate the synthesized HR SWIR band, we critically analyse the qualitative and quantitative results produced by DeepSWIR using state-of-the-art evaluation metrics. Further, we convert the synthesized DN values to Top Of Atmosphere (TOA) reflectance and compare with the corresponding band of Sentinel-2B. Finally, we show one real world application of the synthesized band by using it to map wetland resources over our region of interest.
Introduction
Short-Wave InfraRed (SWIR) imagery is a critical asset in remote sensing due to its ability to penetrate atmospheric haze, smoke, and fog, making it indispensable for geology, wildfire monitoring, and material identification. However, capturing high-resolution (HR) SWIR data is technically challenging and expensive. For instance, the Indian Resourcesat-2A satellite carries the LISS-III sensor, providing 24m resolution SWIR data, alongside the LISS-IV sensor, which provides 5m resolution Green, Red, and Near-Infrared (NIR) data but lacks a SWIR band. This paper addresses the challenge of synthesizing a 5m SWIR band by leveraging the spectral relationships captured at lower resolutions.
Methodology
The proposed DeepSWIR model is a Convolutional Neural Network (CNN) designed to learn spectral mapping rather than simple spatial upscaling. The core methodology relies on the assumption of scale-invariant spectral mapping: the relationship between visible/NIR bands and the SWIR band remains consistent across different spatial resolutions.
Network Architecture
DeepSWIR employs a dual-branch architecture to balance detail preservation with complex mapping:
- Global Residual Learning (GRFE): Utilizes 1x1 convolutions to preserve the sharp spatial details of the 5m input imagery.
- Local Residual Learning (LRFE): Uses deep residual blocks to capture complex spectral mappings. Activation functions were omitted at the end of these blocks to better handle the wide dynamic range of satellite radiometric values.
To handle large-scale satellite imagery, the model processes data in 32x32 patches. A novel Gaussian feathering technique is used during stitching to eliminate blocky artifacts and ensure seamless transitions between patches.
Datasets and Training
The model was trained on a custom dataset derived from concurrent LISS-III and LISS-IV imagery. The training set comprised approximately 800,000 patches covering diverse terrains across India, including urban, aquatic, desert, and mountainous regions. Validation was performed on 200,000 patches from different geographical areas and seasons to ensure robust generalization across varying environmental conditions.
Results and Discussion
DeepSWIR demonstrated high quantitative accuracy on 24m test data, achieving a Mean Absolute Error (MAE) of 4.46 and an RMSE of 7.96. Notably, 97% of predicted Digital Number (DN) values were within a 5-count tolerance of the ground truth, indicating high radiometric fidelity.
Cross-sensor validation using Sentinel-2B data confirmed that the synthetic 5m SWIR reflectance closely matches physical sensor readings in homogeneous regions. In practical applications, the synthesized band significantly improved wetland mapping in Goa by accurately delineating narrow water bodies and reducing false positives compared to standard NIR-based methods.
Conclusion
DeepSWIR provides a powerful computational solution for generating high-resolution SWIR data where hardware is limited. While the model successfully bypasses the ground truth bottleneck through spectral translation, it is intended to augment, not replace, physical sensors. Future work will continue to explore the limits of data-driven spectral synthesis in novel environmental conditions and across different sensor platforms.
Summary updated