Forecasting landslide deformation is challenging due to influence of various internal and external factors on the occurrence of systemic and localized heterogeneities. Despite the potential to improve landslide predictability, deep learning has yet to be sufficiently explored for complex deformation patterns associated with landslides and is inherently opaque. Herein, we developed a holistic landslide deformation forecasting method that considers spatiotemporal correlations of landslide deformation by integrating domain knowledge into interpretable deep learning. By spatially capturing the interconnections between multiple deformations from different observation points, our method contributes to the understanding and forecasting of landslide systematic behavior. By integrating specific domain knowledge relevant to each observation point and merging internal properties with external variables, the local heterogeneity is considered in our method, identifying deformation temporal patterns in different landslide zones. Case studies involving reservoir-induced landslides and creeping landslides demonstrated that our approach (1) enhances the accuracy of landslide deformation forecasting, (2) identifies significant contributing factors and their influence on spatiotemporal deformation characteristics, and (3) demonstrates how identifying these factors and patterns facilitates landslide forecasting. Our research offers a promising and pragmatic pathway toward a deeper understanding and forecasting of complex landslide behaviors. (c) 2025 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/ 4.0/).

Skill in predicting where damaging convective storms will occur is limited, particularly in the tropics. In principle, near-surface soil moisture (SM) patterns from previous storms provide an important source of skill at the mesoscale, yet these structures are often short-lived (hours to days), due to both soil drying processes and the impact of new storms. Here, we use satellite observations over the Sahel to examine how the strong, locally negative, SM-precipitation feedback there impacts rainfall patterns over subsequent days. The memory of an initial storm pattern decays rapidly over the first 3-4 days, but a weak signature is still detected in surface observations 10-20 days later. The wet soil suppresses rainfall over the storm track for the first 2-8 days, depending on aridity regime. Whilst the negative SM feedback initially enhances mesoscale rainfall predictability, the transient nature of SM likely limits forecast skill on sub-seasonal time scales. Early warning of severe weather is particularly important in Africa, where resilience to storm hazards such as flash flooding is weak. Given large-scale atmospheric conditions favorable for convective activity, understanding where storms will occur is challenging for conventional weather prediction models. In semi-arid regions such as the Sahel, the spatial distribution of SM provides additional predictability of convective rain, via its impact on heating and moistening of the atmosphere. Given that convection is favored over drier soils and that storms create new SM patterns every few days during the wet season, the extent to which knowledge of today's SM aids rainfall prediction in future days is unclear. Here we use 17 years of satellite observations to document how surface properties evolve over 20 days after a storm, and how the surface influences subsequent rainfall patterns. We find that even in regions of West Africa where storms are frequent, the suppression of rain over recently-wetted soils is evident out to 2 days. In climatologically drier regions, this predictability extends out to 8 days. Overall, the feedback between SM and rainfall enhances rainfall predictability in the short-term (days), but effectively degrades the skill of longer-term (weeks) forecasts. Satellite observations over the Sahel reveal how the land surface evolves in the 20 days after a Mesoscale Convective System (MCS) After an MCS, rainfall is suppressed over wet soils for 2 days in humid regions and up to 8 days in drier areas Initially soil moisture enhances rainfall predictability, but the strong land feedback degrades skill at longer lead times

As a vital source of the climate change predictability, the snow depth predictability originates from its own persistence and the external forcing factors. In order to investigate the root of snow depth predictability at the North Hemisphere, this study conducted an ensemble of 20 simulations spanning 50 years with the Community Earth System Model (CESM). With a regression model constructed via the canonical correlation analysis method, we analyzed the temporal and spatial distribution characteristics of snow depth predictability on the global scale, as well as the effects of snow depth persistence and sea surface temperature (SST) on snow depth predictability. The results show that the predictability due to snow depth persistence depends on both season and location. The persistence of snow depth can reach more than 3 months in high latitude region. After considering the SST forcing, the predictability is increased in many parts of the Northern Hemisphere, such as northern North America, Europe, and central Siberia. The areas where SST significantly influences snow depth predictability mainly overlap the snow cover transition zones. We further investigated the possible pathways of the impact of SST on snow depth predictability, and found that in North America and Europe, SST improves the predictability mainly through affecting the surface temperature, while in central Siberia and eastern Europe, the pathway also includes snowfall and shortwave radiation, respectively. Additionally, we conducted a similar analysis with three other climate models from the Atmospheric Model Intercomparison Project phase 6 (AMIP6), and the results can also verify the conclusions of CESM ensemble simulations.