SWIR-cGAN: Synthesizing SWIR Bands Using VNIR Imagery from CBERS and AMAZONIA-1

Alisson Cleiton de Oliveira, Rennan de Freitas Bezerra Marujo, Thales Sehn KörtingarXiv 2026 · DOI · arXiv

TL;DR

Researchers have developed SWIR-cGAN, a deep learning framework that synthesizes missing Short-Wave Infrared (SWIR) bands for Brazilian satellites using only their existing Visible and Near-Infrared (VNIR) data. By "hallucinating" these spectrally distant bands, the model enables high-frequency environmental monitoring—such as fire and deforestation detection—on platforms that lack the physical hardware to see them.

What Problem Does This Paper Address?

In remote sensing, there is often a trade-off between how often a satellite visits a spot (temporal resolution) and what parts of the light spectrum it can see (spectral resolution). The Brazilian satellites CBERS-4/4A and AMAZONIA-1 are workhorses for environmental monitoring, offering a high revisit rate of 1–3 days. However, their Wide-Field Imager (WFI) sensors lack Short-Wave Infrared (SWIR) bands.

SWIR is critical for distinguishing between different types of vegetation, soil moisture, and—most importantly for Brazil—detecting burned areas and active fires. While satellites like Sentinel-2 have SWIR sensors, they don't visit as frequently. The authors aim to bridge this gap by using conditional Generative Adversarial Networks (cGANs) to synthesize Sentinel-2-like SWIR bands from the readily available VNIR data of the Brazilian satellites.

Methods

The authors propose SWIR-cGAN, a framework built on a sophisticated image-to-image translation architecture:

  • Generator: An attention-enhanced U-Net. It uses residual blocks to stabilize training and attention gates to help the model focus on spatially relevant features (like the edges of a forest or a burn scar) while suppressing irrelevant background noise.
  • Discriminator: A PatchGAN discriminator that looks at local image patches to ensure the synthesized SWIR bands have realistic textures and high-frequency details, rather than just being "blurry" averages.
  • Multi-Loss Objective: Instead of relying on a single error metric, the model optimizes a composite loss function including:
  • $L_1$ loss for pixel-wise accuracy.
  • SSIM loss to maintain structural consistency.
  • Gradient loss to preserve sharp transitions.
  • Perceptual loss (via a pre-trained VGG-16) to ensure textural realism.
  • Adversarial loss to drive the "realism" of the generated data.
Fig. 2. Overview of the proposed attention U-net–based cGAN. 16px: added pixels; c: number of channels; H (height) and W (width): spatial dimensions; Conv: convolution; s2: stride = 2; LReLU: Leaky ReLU; ResBlock: residual block; IN: instance normalization; f: number of filters; Att: attention; NN: nearest neighbor; s1: stride = 1; adv loss: adversarial loss.

Datasets

The study utilized a custom-built dataset of spatiotemporal patch pairs:

  • Input: VNIR imagery from CBERS-4, CBERS-4A, and AMAZONIA-1.
  • Target (Ground Truth): SWIR bands 11 and 12 from Sentinel-2/MSI, resampled and coregistered to match the WFI resolution (55m or 64m).
  • Scope: 5,151 pairs of $128 \times 128$ pixel patches. The training set focused heavily on the Brazilian Cerrado biome, a region frequently impacted by fire.
Fig. 1. Geographic distribution of the spatiotemporal patch pairs. The open dataset has 5,151 paired WFI–MSI patches acquired on the same day: 1,742 from CBERS-4A (1,361 train/val, 381 test), 1,796 from CBERS-4 (1,446 train/val, 350 test), and 1,613 from AMAZONIA-1 (1,318 train/val, 295 test).

Key Insights and Ideas

  • The Power of NIR: The authors highlight that while RGB imagery provides basic context, the Near-Infrared (NIR) band is the "secret sauce" for predicting SWIR behavior, as it offers much richer information about vegetation health and structure.
  • Multi-Loss is Mandatory: The ablation study revealed that the multi-loss formulation was the single most important factor in performance. Without it, the model struggled to maintain the structural integrity of the landscape.
  • Band-Specific Personalities: Interestingly, the optimal hyperparameters for synthesizing SWIR-1 (1,610 nm) and SWIR-2 (2,190 nm) were quite different. SWIR-1 synthesis relied more on structural and gradient losses, whereas SWIR-2 required a more balanced multi-objective setup to handle its higher spectral complexity.

Key Results

SWIR-cGAN consistently outperformed baseline models like standard U-Nets, ESRGAN, and Vision Transformer-based U-Nets (ViT-U-net).

  • Quantitative Performance: On the test set, the model achieved an SSIM of 0.87 and a PSNR of ~29-30 dB for both SWIR bands.
  • Qualitative Success: The model proved particularly effective at identifying burned areas. In Figure 3, the synthesized bands clearly delineate fire-affected regions, even capturing the high-error signals associated with active fire fronts.
Fig. 3. Examples of image patches containing vegetation and burned areas (represented by dark tones at the center of the false-color composites). (a) Predicted SWIR-1 band; (b) Predicted SWIR-2 band. There is an active fire (black circle) in the southern portion with higher errors in SWIR-2 band.

Limitations and Future Work

  • Dataset Diversity: The current model is heavily weighted toward the Cerrado biome. The authors acknowledge that performance might vary in the dense Amazon rainforest or during different seasons, requiring a more diverse training set for true "cross-biome" generalization.
  • Operational Integration: While the results are promising, the next step is integrating this into real-time systems like Deter (Brazil's deforestation detection system) to see if "hallucinated" SWIR bands actually improve the speed and accuracy of human analysts.
  • Spectral Distance: Synthesizing bands that are spectrally far apart remains a challenge; future work may explore incorporating other data sources (like radar) to further constrain the synthesis.

Check out the code: The implementation and pre-trained weights are available on GitHub at marujore/swir-cgan.

Summary updated

Abstract

Spectral band synthesis is an important research topic in remote sensing. The Brazilian satellites CBERS-4, CBERS-4A, and AMAZONIA-1 provide high temporal resolution (1-3 days) visible and near-infrared (VNIR) imagery through the Wide-Field Imager (WFI), but lack short-wave infrared (SWIR) bands. Recent deep learning methods have shown promising results for spectral synthesis, but challenges remain for cross-satellite generalization and spectrally distant bands. This paper presents a conditional Generative Adversarial Network (cGAN) with an attention-enhanced U-net generator and a multi-loss objective to synthesize Sentinel-2/MSI SWIR bands 11 (SWIR-1) and 12 (SWIR-2) from WFI VNIR inputs. Experiments used 4,125 patches for training and cross-validation and 1,026 unseen patches for testing, with hyperparameters optimized using a subset of the training data. On SWIR-1, the model achieved SSIM of 0.900, PSNR of 31.44 dB, and MAE of 0.019 in TOA reflectance. On SWIR-2, it achieved SSIM of 0.917, PSNR of 33.35 dB, and MAE of 0.015. Benchmarking was performed on the test dataset, with each model optimized separately prior to evaluation. SWIR-cGAN outperformed the baselines, particularly for SWIR-2. The ablation study identified the multi-loss as the main performance contributor, while the attention and residual modules provided additional improvements. Our results demonstrate the feasibility of synthesizing SWIR bands from WFI imagery. The SWIR-cGAN is available at https://github.com/marujore/swir-cgan.
BibTeX
@misc{oliveira2026swir,
  author = {Alisson Cleiton de Oliveira and Rennan de Freitas Bezerra Marujo and Thales Sehn Körting},
  title = {{SWIR-cGAN: Synthesizing SWIR Bands Using VNIR Imagery from CBERS and AMAZONIA-1}},
  year = {2026},
  eprint = {2611.22922},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22922},
}
Figure