Sub-Annual Precipitation Reconstruction

• Sub-annual precipitation reconstruction is a method to estimate monthly or daily rainfall using AI models and diverse proxy data.
• It integrates conditional diffusion frameworks, deep learning architectures, and multi-source conditioning to fuse sparse and heterogeneous inputs.
• The approach supports robust evaluation of hydrological extremes, climate variability, and teleconnection patterns over historical periods.

Sub-annual precipitation reconstruction refers to the quantitative estimation of precipitation fields at temporal resolutions finer than a year (typically monthly or daily), from incomplete, noisy, or indirect observations and multi-modal climate data. This field integrates generative modeling, data assimilation, atmospheric dynamics, and proxy analysis to provide high-resolution temporal and spatial precipitation datasets. These reconstructions support the evaluation of hydrological extremes, climate variability, and teleconnections over intervals not covered by direct measurements, offering essential records for both climate science and the broader geosciences.

1. Methodological Foundations

The dominant methodologies for sub-annual precipitation reconstruction employ generative AI frameworks, particularly conditional diffusion models and deep learning architectures, to synthesize spatiotemporal precipitation fields from sparse or heterogeneous input data.

Diffusion Modeling Framework

A diffusion probabilistic model defines a forward stochastic process that incrementally perturbs fields $x_0$ (e.g., monthly or daily precipitation maps) with Gaussian noise, and a learned neural network to approximate the reverse process, reconstructing a clean signal from noise using available conditioning information: $$q(x_t | x_{t-1}) = \mathcal{N}\left(x_t; \sqrt{1-\beta_t}\, x_{t-1}, \beta_t I\right)$$

$$p_\theta(x_{t-1} | x_t, c) = \mathcal{N}\left(x_{t-1}; \mu_\theta(x_t, t, c), \Sigma_\theta(t) I\right)$$

where $c$ encodes the observational or proxy information (e.g., event indicators, reanalysis fields, regime indices).

For example, Sida He et al. trained a 3D spatiotemporal diffusion model to invert historical event indicators extracted from Chinese archives, reconstructing monthly precipitation over eastern China from 1368–1911 AD with high consistency and physical validity (He et al., 30 Jan 2026). Other models, such as PRIMER, apply coordinate-based infinite-dimensional diffusion processes for fusing satellite, reanalysis, and gauge observations at sub-annual scales, supporting downscaling, gap-filling, and bias correction (Sun et al., 13 Jun 2025). Conditional architectures leverage U-Net backbones, FiLM conditioning, and attention modules for effective spatiotemporal feature extraction.

2. Data Sources and Conditioning

Sub-annual precipitation reconstructions utilize diverse input sources, often with contrasting characteristics:

• Historical archives and qualitative proxies: Event-based records (e.g., flood/drought dating, local chronicles), digitized into event codes and mapped onto spatiotemporal grids. This approach yields event channels (+1 flood, –1 drought, 0 none) and, where possible, sub-annual timing (He et al., 30 Jan 2026).
• Instrumental and remote sensing data: Gridded reanalyses (e.g., ERA5 at 0.25°), satellite-derived estimates (IMERG), and irregularly-distributed gauge measurements (Sun et al., 13 Jun 2025).
• Atmospheric state variables and regimes: Large-scale dynamical predictors, such as weather regime (WR) indices derived by clustering geopotential height EOFs, or selected atmospheric prognostics for physically-constrained reconstructions (Camilletti et al., 16 Jun 2025, Aich et al., 1 Apr 2025).

Encoding approaches align diverse input sources via gridding, interpolation, or embedding. Conditioning is often realized through cross-attention, FiLM layers, or explicit source embeddings injected at each neural network block (Sun et al., 13 Jun 2025). For models relying on regime indices, static features (latitude, land–sea mask, orography) are combined with learned latent vectors and atmospheric predictors (Camilletti et al., 16 Jun 2025).

3. Model Training, Calibration, and Validation

Calibration involves synthesizing pseudo-observations when ground truth is unavailable. For example, the event synthesis approach maps model-simulated precipitation to event classifications matching observed archive counts, supporting supervised loss formulation (He et al., 30 Jan 2026). Multi-source frameworks implement two-stage training: learning large-scale, seasonal, and monthly climatology from reanalysis/satellite, with subsequent refinement using gauge data for local statistical correction (Sun et al., 13 Jun 2025).

Model evaluation leverages a suite of deterministic and probabilistic metrics depending on the reconstruction task:

• $R^2$ (explained variance), coefficient of efficiency (CE), and anomaly correlation coefficient (ACC) for deterministic skill at annual, seasonal, and monthly resolutions (He et al., 30 Jan 2026, Camilletti et al., 16 Jun 2025).
• Continuous Ranked Probability Skill Score (CRPSS), spread–skill ratio (SSR), mean absolute error (MAE), and power spectral density (PSD) for probabilistic and structural fidelity (Sun et al., 13 Jun 2025, Aich et al., 1 Apr 2025).
• Ensemble approaches quantify uncertainty, using CRPS and spread vs. RMSE to ensure calibration (Aich et al., 1 Apr 2025).

Cross-validation includes tests on withheld ensemble members, comparisons with parallel proxy reconstructions, and physical consistency checks via empirical orthogonal function (EOF) analysis and teleconnection indices (He et al., 30 Jan 2026).

4. Temporal and Spatial Resolution

Reconstructions typically target native monthly or daily resolution grids, yielding fields $x \in \mathbb{R}^{T\times H \times W}$, where $T$ is the number of sub-annual intervals (e.g., $T=12$ for months), and $H\times W$ specifies spatial dimensions. The spatial grid is determined by archival source or model framework: 1° zonal × 0.57° meridional for eastern China (He et al., 30 Jan 2026); $0.1^\circ$ or $0.25^\circ$ grids for coordinate-based and diffusion-based models (Sun et al., 13 Jun 2025, Aich et al., 1 Apr 2025); and 1°×1° for Europe (Camilletti et al., 16 Jun 2025). Models directly generate the full grid without interpolation, except where multi-source fusion requires feature projection between irregular gauge locations and dense grids.

Reconstructions allow aggregation to seasonal or annual scales, preserving intra-annual variability and supporting decomposition into physical modes (e.g., monsoon dipoles, regime influences), as revealed through EOF analysis or cross-correlation with climate indices (He et al., 30 Jan 2026).

5. Large-scale Dynamics, Extremes, and Teleconnections

Sub-annual reconstructions enable the quantification of hydrological extremes and the diagnosis of climate teleconnection patterns over centuries.

• The reconstructed series for China quantifies the full spatial/seasonal evolution of major droughts and floods (e.g., Ming Great Drought, 1593 Great Flood), capturing annual z-scores >+3 or <–2 and revealing spatial extent, temporal persistence, and intra-annual migration of anomalies (He et al., 30 Jan 2026).
• Principal large-scale precipitation modes, identified via EOFs, align with climatological patterns: uniform wet/dry, meridional and zonal dipoles, and meridional tripoles, capturing up to 20% of variance per mode.
• ENSO teleconnections are assessed via correlation of reconstructed precipitation with prior-winter Niño 3.4 indices. The canonical southeast–north dipole, spring-to-winter phase reversal, and persistent signals over five centuries are demonstrated, highlighting the ability to probe dynamics inaccessible in modern records (He et al., 30 Jan 2026).
• For Europe, AI models leveraging WR indices reproduce monthly anomaly fields, capturing both mean and extreme precipitation with competitive or superior skill to operational forecast systems (e.g., ECMWF SEAS5), even under regime index uncertainty (Camilletti et al., 16 Jun 2025).

6. Computational and Algorithmic Considerations

State-of-the-art reconstructions utilize deep diffusion architectures: spatiotemporal U-Nets (4 down/up blocks, 20M parameters), temporal self-attention, and large-batch or ensemble-based inference (e.g., 80–100 samples per year) (He et al., 30 Jan 2026). Optimizers include Adam with $lr=5 \times 10^{-4}$, and training uses mixed-precision on high-end accelerators (e.g., Ascend 910B, A100 GPUs) (He et al., 30 Jan 2026, Sun et al., 13 Jun 2025).

For multi-source fusion, coordinate-based and sparse convolutional modules (e.g., SparseConvResBlocks) process irregular input, while plug-and-play strategies (inpainting, SDEdit) impose observational constraints during sampling (Sun et al., 13 Jun 2025). Hierarchical training (two-stage with source-conditioned loss weights) allows encoding of both climatology and fine-scale event variability.

Post-processing includes baseline renormalization, seasonal aggregation, and ensemble spread analysis for uncertainty quantification. Power spectral diagnostics confirm recovery of high-frequency precipitation features beyond the reach of traditional upsampling or regression methods (Aich et al., 1 Apr 2025).

7. Scientific and Practical Implications

Sub-annual precipitation reconstructions furnish long-term datasets critical for hydrological, climatological, and interdisciplinary studies:

• They enable event timing, magnitude, and spatial extent analysis for historic extremes, with quantified uncertainty at grid and basin scales (e.g., ensemble standard deviation $\sigma \approx 0.07$–$0.12$ for major droughts/floods).
• Seasonal to sub-seasonal reconstructions using regime indices democratize surface anomaly estimation, facilitating modular, fast-response forecasting systems for water resource management, agriculture, and risk assessment (Camilletti et al., 16 Jun 2025).
• Multi-source frameworks generalize to new models and forecasts without retraining, acting as plug-and-play Bayesian priors for bias correction and downscaling (Sun et al., 13 Jun 2025).
• A plausible implication is that the integration of qualitative archives with generative AI reconstructions will advance paleoclimatology and socio-environmental history, extending high-resolution climate records into the pre-instrumental era (He et al., 30 Jan 2026).

The field continues to evolve, with future directions including improved out-of-distribution robustness, computational scaling, direct ESM coupling, and incorporation of additional atmospheric or proxy controls.

---

Key References:

• Generative AI inversion of Chinese archives: (He et al., 30 Jan 2026)
• Multi-source coordinate-based generative modeling: (Sun et al., 13 Jun 2025)
• AI anomaly reconstruction from synoptic regimes: (Camilletti et al., 16 Jun 2025)
• Diffusion models for sub-annual precipitation generation: (Aich et al., 1 Apr 2025)