SEVIRI IR_108 Infrared Channel

• SEVIRI IR_108 is an infrared channel at 10.8 μm that measures upwelling radiance to infer cloud-top brightness temperatures for deep convection.
• It provides high-frequency (15-minute) imaging, enabling detailed convective cloud analysis and real-time rainfall nowcasting through deep learning frameworks.
• It supports operational meteorology by converting predicted brightness temperatures into rainfall rates using non-linear mapping and 3D event clustering.

The SEVIRI (Spinning Enhanced Visible and InfraRed Imager) infrared channel at 10.8 μm is a principal data source for convective cloud analysis and precipitation nowcasting using satellite observations. Centered within the thermal IR window, this channel (denoted IR_108) is leveraged to infer cloud-top thermal properties, enabling high-frequency, large-area monitoring for operational and research meteorology, particularly in algorithmic frameworks for rainfall prediction and event detection (Bhuskute et al., 14 Nov 2025).

1. Physical and Radiometric Foundations

SEVIRI's IR_108 band is centered at 10.8 μm, a spectral window in which upwelling radiance from Earth's surface and atmosphere, particularly from optically thick cloud tops, is dominant. The instrument measures spectral radiance $L_\lambda$, which—via the inverse Planck function—is reported as a brightness temperature $T_b$. Cold, high cloud tops, typically signatures of deep convection, present as low $T_b$ values, whereas clear-sky or low cloud regions appear significantly warmer.

The radiometric conversion adheres to the Planck law:

$$L_\lambda = \frac{2hc^2}{\lambda^5}\frac{1}{\exp(hc/(\lambda k_B T))-1} \implies T_b = B^{-1}(L_\lambda),$$

where $h$ is Planck’s constant, $c$ is the speed of light, $k_B$ is Boltzmann’s constant, $\lambda$ is wavelength, and $B^{-1}$ is the inverse Planck operator. This formalism ensures that $T_b$ provides a physically-motivated interpretation directly related to cloud-top emission characteristics.

2. Spatial and Temporal Sampling Regime

Native SEVIRI IR_108 imagery is provided at a resolution of 252 × 252 pixels, corresponding to spatial samples of approximately 4 km at the equator and up to approximately 12 km in northern domain limits. For compatibility with convolutional neural networks, this is conventionally padded to 256 × 256 pixels. The refresh rate is one image every 15 minutes, so operational nowcasting typically aggregates four consecutive frames (covering a one-hour period) as the minimal spatiotemporal ensemble for predictive modeling.

3. ConvGRU-Based Brightness-Temperature Forecasting

A two-stage deep learning framework utilizes the SEVIRI IR_108 channel for short-term rainfall prediction. The initial modeling stage predicts future brightness temperatures; the architecture comprises an encoder–ConvGRU–decoder cascade:

• Encoder: Two convolutional layers (kernels 3×3) with channel progressions 1→16→32, using ReLU activations.
• Temporal Core: Two stacked ConvGRU layers maintaining hidden states of shape 256 × 256 × 64.
• Decoder: Expanding convolutional layers with channel reductions 64→32→16→1, reconstructing the predicted $T_b$ fields.

The ConvGRU evolution at each timestep $t$ for input $x_t$ and preceding hidden state $h_{t-1}$ is:

$$\begin{aligned} z_t &= \sigma(W_z * x_t + U_z * h_{t-1} + b_z), \ r_t &= \sigma(W_r * x_t + U_r * h_{t-1} + b_r), \ \tilde h_t &= \tanh(W_h * x_t + U_h * (r_t \odot h_{t-1}) + b_h), \ h_t &= (1 - z_t)\odot h_{t-1} + z_t\odot \tilde h_t, \end{aligned}$$

where $*$ denotes convolution and $\odot$ denotes element-wise multiplication.

Input preparation: Four sequential $T_b$ frames $\{T_b^{t-3},T_b^{t-2},T_b^{t-1},T_b^t\}$ are normalized by 300 K, mapping into [0,1]; non-cloud pixels are assigned a value of 1 using Otsu’s thresholding, and the resulting maps are stacked along the channel dimension.

Training: Separate ConvGRU models are trained for each forecasting horizon (1–4 h), each predicting a 4-frame block without iterated autoregression. Optimization uses Adam (learning rate $10^{-3}$, batch size 25, 10 epochs) and minimizes mean squared error between predicted and measured normalized $T_b$ fields: $$\mathcal{L}_{\rm BT} = \frac{1}{N}\sum_{i=1}^N (\hat T_{b,i} - T_{b,i})^2.$$

4. Transformation to OPERA-Compatible Rainfall Rates

Predicted $T_b$ fields are spatially upscaled (from 252×252 to the OPERA precipitation grid, 1512×1512), then passed through a nonlinear infrared-rainfall mapping. The empirically-derived power-law:

$$R = f(T_b; \alpha, \beta, \gamma) = \alpha[\max(0, \gamma - T_b)]^\beta$$

with parameters $\gamma = 300$ K (cloud-base threshold), $\alpha = 0.02$ (mm h$^{-1}$/K$^\beta$), and $\beta = 1.5$ as fitted via least squares to collocated SEVIRI–OPERA data, provides conversion to rainfall rates $R$. This mapping captures the rapid nonlinearity of rain-formation as inferred from IR coldness, consistent with established empirical literature (Goodman et al. 1994; Vicente et al. 1998).

5. Event Prediction: 3D Connected Component Analysis

For event-level prediction, sequences of 16 hourly rainfall-rate frames (post-IR mapping) are treated as a three-dimensional volume (x, y, t). Precipitating gridpoints exceeding the 2 mm h$^{-1}$ threshold are grouped using 18-connectivity 3D labeling to identify contiguous rain voxels—interpreted as discrete precipitation events in space–time.

For each detected event cluster, characteristic descriptors are extracted:

• Maximum intensity,
• Duration (number of timesteps occupied),
• Spatial footprint (voxel count at peak intensity),
• Centroid location and bounding box in the central time slice for event tracking.

The submission retains the five highest-intensity events per sequence.

6. Performance Evaluation and Benchmarking

The ConvGRU pipeline demonstrates robust skill at multiple horizons:

Forecast	$T_b$ RMSE [K]	$T_b$ SSIM	Rainfall RMSE [mm]	Rainfall SSIM	F1 score (0.5 mm h$^{-1}$)
1h	17.13	0.761	2.48	0.747	0.6822
2h	18.53	0.756
3h	20.51	0.754
4h	20.94	0.756

Key additional scores:

• Probability of detection (POD) at 0.5 mm h$^{-1}$: 0.7713;
• False alarm ratio (FAR): 0.3883.
• Final cumulative-rainfall CRPS on Weather4Cast leaderboard: 3.37 (2nd place).

Brightness temperature nowcasts surpass a persistence baseline and are at parity or marginally superior to ConvLSTM, especially for longer lead times (Bhuskute et al., 14 Nov 2025).

7. Significance and Operational Implications

The exploitation of SEVIRI’s 10.8 μm IR window for short-term precipitation forecasting demonstrates that a single-channel, computationally-efficient pipeline can yield competitive rainfall and event nowcasts. The approach combines physical radiometric principles with modern deep learning architectures and objective clustering for event identification. This methodology is particularly notable for its capacity to operate at rapid cadence and broad spatial coverage without reliance on multi-spectral, ground-based, or radar data. The streamlined reliance on IR_108 sets a baseline for operational scalability and transfer to nowcasting tasks in bandwidth-limited or sensor-restricted contexts (Bhuskute et al., 14 Nov 2025).