A History-Aware Residual-Learning Framework for Enhancing SMAP Soil Moisture Retrieval Over Organic-Rich Tundra

Soil Moisture Active Passive (SMAP) L-band retrievals often show less seasonal soil-moisture variation than field measurements in organic-rich tundra. Freeze–thaw state and organic-layer wetness change through the year, whereas the operational algorithm uses mostly static soil parameters. We introduce the Composite Shared Latent Dynamic Feature (CSLDF), a residual-learning method that summarizes the previous 365 days of surface temperature, SMAP SCA soil moisture, vegetation optical depth (VOD), and H- and V-polarized brightness temperatures into five predictors learned separately within each validation fold. We compare it with Operational SMAP Single-Channel Algorithm (SCA), a static Random Forest (RF), and calendar-seasonal, equal-weight 365-day, and fixed-window history RF baselines. All features and models use training data only. In rolling next-month validation, Operational SMAP SCA had R 2 = −0.055 and RMSE = 0.131 m3 m −3 , whereas Learned History C-SLDF achieved R 2 = 0.548 and RMSE = 0.086 m3m −3. An analytical seasonal branch identified importance peaks near independently derived thaw and freeze windows. In leaveone-site-out testing, a spatial extension achieved R 2 = 0.165 and RMSE = 0.124 m3 m −3, but its site-block 95% interval for R 2 included zero. A secondary rolling test across four available years (16 thawed monthly blocks; 6873 observations) showed a smaller gain over Equal365 (∆R 2 = +0.017; ∆RMSE = -0.0030 m 3 m −3). Preceding microwave and thermal observations therefore add information beyond annual calendar phase. Temporal enhancement was consistent; spatial and cross-regional transfer require broader validation.

Hali tarjima qilinmagan