In the context of strengthening the construction of ecological civilization and accelerating the “carbon peak” in China, the regional ecological pattern and its connection with carbon sink capacity have become an urgent topic. Given that Inner Mongolia is a large carbon emission province and the conflict between economic development and ecological protection is particularly prominent, we took Inner Mongolia as an example to extract its ecospatial network, then calculated the integrity index, topological indices, and recovery robustness of the network and evaluated integrity and other properties of the ecospatial network structure by combining them with the ecological background. In addition, we analyzed the relationship between the topological indices and net primary productivity (NPP). The results showed that the network was scale-free and heterogeneous, with low integrity, connectivity and stability, which were the focus of future optimization. The nodes with important functions were mainly distributed in the farm-forest ecotone, grasslands, and the agro-pastoral ecotone; under the simulation attack, the node recovery robustness was stronger than the corridor recovery robustness, and NPP was negatively and significantly correlated with the woodland nodes and grassland nodes. In terms of ecological restoration, the unused land in the west is a key area, and it is necessary to add new ecological nodes and corridors. In terms of enhancing carbon sequestration capacity, under the premise of ensuring network connectivity, the appropriate and rational merging of ecological nodes and corridors within woodlands and grasslands is a particularly effective means. This study provides a reference for evaluating and optimizing the ecological pattern of areas with prominent ecological problems and improving the carbon sink of ecosystems in terms of their ecospatial network structure.

The soil phase-transition temperature (PTT) is critical for accurately determining the soil freeze–thaw (FT) state, which is essential for developing surface FT state identification algorithms based on remote sensing technology and reliably validating FT products. In this study, we develop an innovative approach for determining soil PTT in the region of the dense soil moisture and soil temperature monitoring network in the Qinghai Lake Basin (QLB-NET). We employ four advanced models—polynomial regression (PR), random forest (RF), feedforward neural network (FNN), and transformer (XFMR) models—to establish relationships between soil PTT and various environmental factors for daily and site-specific soil PTT estimation. Soil PTT values derived from a sigmoid curve fitting (SCF) method that combines physical theory with empirical modeling are used in the models as the target data considering soil properties and topographic features as predictors. Additionally, we systematically evaluate the models’ performance and compare the results against those of the fixed 0°C threshold methods. The results indicate that most soil freezing temperatures range between -1°C and 0°C, and thawing temperatures range between -0.5°C and 0.5°C, with occasional freezing temperatures above 0°C. Model-derived soil PTT significantly outperforms the traditional 0°C threshold in terms of soil FT state classification accuracy, particularly during the frozen and FT transitional seasons. The XFMR model achieves superior performance in an internal comparative analysis, highlighting its particular effectiveness for soil PTT determination. These results provide valuable insights for improving remote sensing-based FT product validation and permafrost monitoring.