PRISMA Satellite Imagery: Overview

• PRISMA Satellite Imagery is a hyperspectral system capturing 400–2500 nm data with a push-broom imager, enabling detailed environmental and geological mapping.
• Its robust calibration and preprocessing pipelines, including spectral realignment and cloud masking, ensure high-accuracy reflectance measurements.
• The platform’s modular design supports advanced deep learning and pansharpening methods, driving innovation in remote sensing and geoarchaeological studies.

PRISMA (PRecursore IperSpettrale della Missione Applicativa) satellite imagery refers to data products acquired by the ASI PRISMA platform, which is equipped with a hyperspectral push-broom imager designed for high-resolution earth observation, trace gas detection, fine-grained environmental analysis, and spectral mapping. PRISMA imagery covers several hundred contiguous bands from the visible (VIS), near-infrared (VNIR), and shortwave infrared (SWIR) domains (approx. 400–2500 nm spectral range) at a nominal ground sampling distance of 30 meters, with a panchromatic (PAN) band at 5 meters. The instrument architecture, data calibration pipeline, and open data policies have catalyzed large-scale scientific exploration in remote sensing, deep learning, geoarchaeology, atmospheric science, and aquatic ecosystem analysis.

1. PRISMA Sensor Design and Data Characteristics

PRISMA employs a push-broom spectrometer with 1,000 individual detectors delivering contiguous bands across VIS–VNIR–SWIR (400–2500 nm), each band with a typical bandwidth of 9–12 nm. The sensor’s radiometric and spectral response functions undergo in-orbit calibration by ASI and additional scene-by-scene recalibration, compensating for thermal drifts and mechanical perturbations. The system achieves radiance signal-to-noise ratios (SNR) exceeding 100:1 in SWIR, especially in methane absorption windows (1.6–1.7 µm), and ~12-bit digital quantization is available for reflectance values.

Standard PRISMA surface products include:

• Level-2A (“L2C STD”): atmospherically corrected, surface reflectance for VNIR and SWIR bands (typically 63 bands in VNIR, 174 in SWIR, final cube up to 203 bands post-cleaning).
• Panchromatic product: spatial resolution of 5 m/pixel used to drive pansharpening for high-resolution applications.
• Co-registration tools: all bands co-aligned to facilitate multi-sensor and deep learning fusion.

This spectral and spatial coverage supports applications from planetary geology and mineral mapping to atmospheric gas quantification and underwater benthic discrimination (Groshenry et al., 2022, Gonzalez-Sabbagh et al., 1 Jan 2026, Zini et al., 2023, Sech et al., 2024).

2. Hyperspectral Data Calibration and Preprocessing

PRISMA’s data pipeline mandates multiple calibration and correction steps:

• Spectral and Radiometric Recalibration: Each detector’s response is aligned to synthetic radiances predicted by line-by-line radiative-transfer models (using LOWTRAN/HITRAN2020), removing biases and wavelength shifts.
• Geolocation and Cloud Masking: Integration with georegistration protocols and reflectance-threshold-based cloud screening permits robust spatial referencing and mitigates contamination.
• Band Cleaning: Automated band exclusion for low-SNR or >5% erroneous pixels (VNIR_PIXEL_L2_ERROR / SWIR_PIXEL_L2_ERROR flags).

Preprocessing generates hyperspectral datacubes, e.g., 63, 256, 256 or up to 203, 1259, 1225, normalized to reflectance [0,1], with optional water/land masking (e.g., NDWI thresholds) and alignment to PAN for fusion tasks (Gonzalez-Sabbagh et al., 1 Jan 2026, Zini et al., 2023).

3. Advanced Retrievals: Methane Plume Detection

For trace-gas mapping, PRISMA enables column-averaged CH₄ (“XCH₄”) retrievals at high spatial resolution. Raw radiance vectors $x \in \mathbb{R}^d$ per pixel are converted to enhancement scores $\alpha(x)$ by matched filtering: $$\alpha(x) = (x - \mu)^\top \Sigma^{-1} t / (t^\top \Sigma^{-1} t)$$ where $\mu$ and $\Sigma$ are empirical mean/covariance, and $t$ the methane absorption Jacobian spectrum.

Plume detection utilizes U-Net segmentation backbones on XCH₄ enhancement maps, leveraging a data augmentation protocol that transposes high-resolution plume structures from Sentinel-2 imagery into PRISMA scenes via gamma-distribution sampling and histogram-specification. Training is conducted with cross-entropy loss across synthetic and transfer-learned datasets, achieving mask-based detection precision up to 0.88 and IoU 0.61, substantially outperforming transfer-learning-only baselines. Scalability to future sensors (EnMAP, EMIT, CarbonMapper) is direct given the modular simulation-augmented pipeline (Groshenry et al., 2022).

4. Deep Learning and Pansharpening Methodologies

PRISMA’s dual VNIR+SWIR cubes and PAN band facilitate advanced pansharpening architectures. The 190-scene PRISMA dataset (≈262,200 km² coverage) allows robust benchmarking:

• Deep Networks: PNN, PanNet, MSDCNN, TFNet, SRPPNN, DIPNet adapted for PRISMA cubes, trained to restore high-resolution hyperspectral imagery from downsampled HS and PAN inputs. Metrics evaluated include ERGAS, SAM, SCC, Q2ⁿ, and QNR (no-reference).
• Machine-Learning-Free Methods: PCA, GSA, HySure, MTF-GLP serve as component-substitution and multiresolution fusion benchmarks.

Best-performing models (TFNet and DIPNet) achieve ERGAS 5.18–6.47, SAM 2.37–2.46, SCC 0.87–0.89 (Reduced Resolution), and QNR up to 0.6405 (Full Resolution). These pansharpened products maintain spectral fidelity and spatial detail, supporting large-scale mapping, classification, and environmental monitoring (Zini et al., 2023).

5. Specialized Applications: Aquatic and Geoarchaeological Analysis

PRISMA data has been systematically adopted for aquatic and archaeological imaging tasks:

• Aquatic Colour Restoration: DichroGAN leverages PRISMA VNIR cubes, atmospheric-corrected surface reflectance, and water masks to disentangle and recover diffuse/specular in-air colours of the seafloor. Exploiting full-spectrum supervision and physical dichromatic modeling allows robust transmission and radiance eqution fitting per pixel:

$$N(u, \lambda_i) = J(u, \lambda_i) T(u, \lambda_i) + V(u, \lambda_i) [1 - T(u, \lambda_i)]$$

where $T(u, \lambda_i) = \exp[-r\, \alpha(\lambda_i)]$, and restoration is achieved by generator stacks in the cGAN (Gonzalez-Sabbagh et al., 1 Jan 2026).

• Geoarchaeological Prospection: Pansharpened PRISMA cubes (167 bands, upsampled to 5 m) support detection of buried artefacts and landscape features below standard HS pixel scale. GSA, MTF-GLP, and HySure were quantitatively benchmarked in Aquileia (Italy), with GSA providing low noise, high contrast, and optimal visual delineation. Wald’s protocol (reduced-resolution reconstruction) metrics showed UIQI up to 0.9809 and ERGAS as low as 3.06. Full-resolution assessments underscored the need to balance spectral fidelity ($D_k$), spatial fidelity ($D_S$), and composite quality ($Q^*$), where MTF-GLP achieved the highest $Q^*$ but GSA was superior for archaeological interpretability (Sech et al., 2024).

Method	UIQI (Best)	ERGAS (Best)	Visual Delineation
GSA	0.9809	3.0607	Optimum (palaeochannels, embankments)
MTF-GLP	0.9806	3.1245	Low noise, high spatial detail
HySure	0.9764	3.1051	Highest contrast, more artifacts

6. Evaluation Metrics, Benchmarks, and Implementation Notes

PRISMA remote sensing pipelines integrate multi-protocol assessment metrics:

• Quantitative (Wald’s protocol): UIQI, SAM (degrees), ERGAS, RMSE, SCC, Q2ⁿ.
• Full-Resolution (no-reference): $D_k$, $D_S$, $Q^*$, QNR, Spectral Distortion, Spatial Distortion.
• Visual Interpretability: Ground truth overlays, PCA compositing, and band artefact inspection are essential to bridge the gap between algorithmic scores and application-specific requirements (archaeology, mineral mapping).

Best practices include bandwise pansharpening, sensor-specific MTF estimation, co-registration via AROSICS, noise-floor preservation in augmentation, and careful validation against expert-labeled targets (Groshenry et al., 2022, Zini et al., 2023, Sech et al., 2024).

7. Scalability and Future Directions

PRISMA’s modular calibration, simulation, and augmentation frameworks facilitate direct transfer to next-generation push-broom or imaging spectrometers with ground sampling in the 10–50 m range and SWIR coverage. Methods based on empirical plume or anomaly morphologies rather than synthetic simulation (LES) offer immediate adaptation for EnMAP, NASA-EMIT, CarbonMapper. The necessity for new pansharpening architectures explicitly designed for VNIR+SWIR fusion, deeper spectral–spatial detail transfer, and improved interpretive metrics—especially in domains where qualitative utility outweighs purely numerical fidelity—remains urgent.

A plausible implication is that further expansion of public PRISMA datasets, alongside tailored pansharpening and restoration methods, will drive advances in environmental monitoring, trace gas quantification, and subsurface archaeological detection at previously unachievable scales and spectral richness (Groshenry et al., 2022, Gonzalez-Sabbagh et al., 1 Jan 2026, Zini et al., 2023, Sech et al., 2024).