Recently, deep convolutional neural network (CNN)-based super-resolution methods for digital rock images have made significant progress in reconstructing pore structures and mineral composition details from low-resolution inputs. However, these approaches still face clear limitations. On the one hand, the inherently local receptive field of CNNs struggles to accurately reconstruct complex microstructures such as pore networks and tortuous throats. On the other hand, most existing architectures are built with excessive network depth, resulting in high computational costs and limiting their deployment on resource-limited devices. In this work, we propose an Efficient Adaptive Aggregation Network (EAAN) for two-dimensional digital rock image super-resolution reconstruction. The proposed network integrates a dual branch residual attention block (DBRAB) with an embedded linear multi-scale semantic attention (LMSA) module, together with a global feature adaptive transformer (GFAT), forming a hybrid CNN-Transformer cooperation framework. This architecture achieves joint optimization of local feature enhancement and global dependency modeling while maintaining an efficient lightweight design. Validated on the DeepRock-SR 2D datasets, the introduced EAAN architecture exhibits a clear advantage over contemporary methods by simultaneously improving image fidelity and minimizing computational demands. Crucially, beyond standard visual metrics, our evaluations verify that EAAN accurately preserves critical petrophysical properties, providing reliable data support for downstream reservoir characterization.
Soil moisture is critical in hydrology, agriculture, and water management, yet most global products provide it at coarse resolutions (approximately 11 km) limiting local-scale use. Most downscaling studies address only the surface layer and rarely compare against the parent product. This study develops an Artificial Neural Network (ANN) framework to downscale Famine Early Warning Systems Network Land Data Assimilation System (FLDAS) soil moisture to 500 m for two depths (0–10 cm, 10–40 cm) in New South Wales, Australia. The Multi-Layer Perceptron used MODIS-derived indicators, topography, and soil texture as predictors, with spatial block cross-validation to avoid autocorrelation bias. Validation against the in-situ OzNet network showed reasonable surface agreement (R = 0.720, ubRMSE = 0.055 m³/m³). Subsurface accuracy was lower (R = 0.421, ubRMSE = 0.060 m³/m³), as surface-restricted optical and thermal variables cannot capture decoupled subsurface processes. Neither configuration exceeded the native 11-km product (R = 0.719 and 0.372), so the downscaling preserves rather than improves temporal accuracy. Reaggregation to 11 km showed approximate consistency with the parent field. The evaluation period lies within the modelling period and no OzNet sensor matches the 10–40 cm layer, so subsurface metrics are a lower bound. With no fine-resolution reference available, the sub-grid structure is inferred from covariates rather than independently validated. SHapley Additive exPlanations (SHAP) showed that evapotranspiration and temperature metrics contributed most to the model output. The findings indicate that machine learning can refine spatial resolution, but deeper soil moisture requires subsurface hydro-physical variables and active microwave data.
Accurately predicting the permeability of strongly heterogeneous carbonate rocks using simple pore structure parameters remains a considerable challenge. In this study, based on data from 669 carbonate rock samples spanning diverse fabric and pore types, a series of ensemble machine learning models and mechanism-data fusion models were developed for permeability prediction in heterogeneous carbonates. Seven machine learning sub-models were integrated via the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS), and mechanism-data fusion was implemented through the input variable constraints and parameter optimization. Comparisons among the traditional mechanistic models, purely data-driven models, the TOPSIS ensemble models, and mechanism-data fusion models showed that the purely data-driven models achieved significantly higher accuracy than the classical Kozeny’s equation and Winland model, highlighting the limitations of traditional mechanistic approaches in representing flow in highly heterogeneous carbonate rocks. The input variable constraints method further improved the predictive performance, and the modified Winland model constraint with XGBoost achieved the most notable gains, raising the testing-set R2 value to 0.770 ± 0.047 and validation-set R2 values to 0.554 and 0.637. However, the parameter optimization modeling (e.g., Kozeny’s factor optimization and FZI optimization) did not outperform the purely data-driven modeling. The TOPSIS ensemble model under the purely data-driven mode achieved the testing-set R2 of 0.775 ± 0.041 and validation-set R2 values of 0.486 and 0.601, slightly below but still close to the best mechanism-data fusion model, thus offering a reliable alternative when the mechanistic knowledge is limited. These new established models require only a few pore structure parameters, providing a pathway for replacing the costly and time-consuming permeability measurements with the digital image analysis.
Marine ecosystem restoration knowledge is dispersed across scientific literature, technical standards, reports, and other professional sources, limiting efficient computational reuse. This study presents MER-LLM, a lightweight domain-adaptation workflow integrating heterogeneous-source structuring, instruction-corpus construction, parameter-efficient fine-tuning, and professional knowledge evaluation. Three instruction corpora with different source compositions and processing intensities were used to evaluate ten compact Qwen- and Llama-based configurations. Qwen2.5–1.5 B-Instruct trained on Dataset-30 k for four epochs was selected as the final configuration. Relative to its base model, the aggregate specialised question-bank score increased from 126.92 to 137.53 out of 175 (8.36
Routine monitoring captures conditions present during sampling, so sparse schedules may fail to encounter threshold exceedances recorded at other times in the same year. Using 106 Class 2 freshwater campaigns from five Buranhém River sites during 2015–2024, the analysis examined seven water-quality parameters and a composite any-exceedance outcome within site-years. Exact finite-population probabilities were calculated for each positive annual reference set to determine whether hypothetical schedules of one to four campaigns would include at least one recorded exceedance. Fixed common support defined the primary one- to three-campaign comparison, requiring at least three classifiable results for the target outcome; Monte Carlo simulation and independent enumeration served only as computational verification. Across 20 positive common-support site-years, mean conditional detection probabilities for the composite outcome were 58.3
