Abstract
Abstract
Drought‐induced desiccation cracking can trigger several weakening mechanisms in surface soils, which not only potentially create preferential pathways for gravity‐driven flow and contaminant transport, but also contribute to the instability and failure of slopes and earthen structures. With the continuous innovation of artificial intelligence (AI) and the burgeoning emergence of foundation models, soil desiccation crack recognition (SDCR) is currently undergoing a fundamental shift toward a data‐centric paradigm. This study proposes a knowledge‐driven unsupervised framework for SDCR under complicated and volatile field conditions, referred to as U‐CrackNet framework. Pseudo‐labels for crack recognition are generated through a fusion paradigm that combines strong knowledge from the Segment Anything Model (SAM) developed by Meta AI with weak knowledge derived from threshold‐based methods. An innovative low‐parameter and low‐computational semantic segmentation architecture (Lite‐UNet) is proposed to achieve an effective trade‐off between accuracy and inference speed, which facilitates application on portable devices under challenging field environments. Extensive experimental results unequivocally demonstrate the superior performance of the U‐CrackNet framework in recognizing field soil desiccation cracking, surpassing other state‐of‐the‐art methods in both overall and detailed evaluations. The unsupervised and generalized attributes of U‐CrackNet framework hold substantial potential for developing SDCR foundation model, which not only offers an unprecedented level of precision and adaptability in quantification, evaluation, prediction, and decision‐making of soil desiccation crack phenomena, but also provides a novel perspective for unraveling the causal mechanisms underlying observed surface cracking patterns and climate changes.