HyDSoil: A Hybrid Diffusion Model for Event-Centered Block Gaps in Multivariate Soil Moisture Time Series

Soil moisture sensors deployed for long-term monitoring often suffer from prolonged data gaps caused by battery depletion, communication dropouts, or hardware failures. When such gaps overlap with irrigation events, key transient phases are obscured and become difficult for conventional imputers to recover. This study proposes HyDSoil, a hybrid diffusion-based imputation model tailored for event-centered block missingness in multichannel soil moisture time series. HyDSoil is first pretrained on a physically interpretable synthetic generator that mimics the baseline-rise-decay response to irrigation and then fine-tuned on field observations from the Baltimore Ecosystem Study dataset. During reverse diffusion, a mask-guided correction keeps observed values fixed while iteratively denoising missing regions. The denoising backbone integrates one-dimensional convolutions, gated recurrent units, and Transformer components to capture high-frequency event spikes, mid-range temporal dynamics, and long-range cross-depth dependencies, respectively. Experiments on both synthetic and real datasets show that HyDSoil reconstructs irrigation-driven peaks with higher fidelity and achieves consistent improvements over strong baselines in global metrics (MAE and DTW) as well as event-focused metrics (PTE and PAE). Ablation studies further verify the complementary contributions of the convolutional, recurrent, and attention branches, and confirm the benefit of synthetic pretraining for long-duration gaps. Overall, HyDSoil enables more reliable continuous soil moisture monitoring and supports precision irrigation analytics.