With a large population and rapid urbanization, there are still many challenges to optimize the ecological-agricultural-urban space. Here, taking Yulin City, situated on the Loess Plateau of China as a case in point, we explored the spatial suitability evaluation of ecological-agricultural-urban space. Building upon the Chinese government’s concept of “resource and environmental carrying capacity and territorial development suitability evaluation” (hereinafter referred to as “double evaluation”), this study applies machine learning to the planning of ecological-agricultural-urban space. It explores an intelligent evaluation method for land space patterns using multi-source data. Based on the random forest (RF) algorithm and geographic information system (GIS), resulting in evaluated spatial patterns for ecological-agricultural-urban in the Yulin area. The results showed the constructed random forest models achieved an accuracy of 93% for ecology, 90% for agriculture, and 92% for urban space in Yulin City on the test dataset. By means of suitability analysis, the results indicated that the extremely important ecological space were predominantly located in the southwestern and eastern regions of the study area, while suitable space for agricultural production were primarily scattered throughout the southeast. In contrast, suitable space for urban construction were concentrated mainly in the central part of the study area. The use of machine learning has proven to be effective in addressing multicollinearity among spatial evaluation factors across three different areas. By eliminating human subjectivity in weight assignment during evaluation, it introduces fresh perspectives for land space planning and status assessment. These findings may offer support for the scientific delineation of ecological-agricultural-urban space (three districts and three lines) in China.

With the increasing global population and escalating ecological and farmland degradation, challenges to the environment and livelihoods have become prominent. Coordinating urban development, food security, and ecological conservation is crucial for fostering sustainable development. This study focuses on assessing the "Ecology-Agriculture-Urban" (E-A-U) space in Yulin City, China, as a representative case. Following the framework proposed by Chinese named "environmental capacity and national space development suitability evaluation" (hereinafter referred to as "Double Evaluation"), we developed a Self-Attention Residual Neural Network (SARes-NET) model to assess the E-U-A space. Spatially, the northwest region is dominated by agriculture, while the southeast is characterized by urban and ecological areas, aligning with regional development patterns. Comparative validations with five other models, including Logistic Regression (LR), Naive Bayes (NB), Gradient Boosting Decision Trees (GBDT), Random Forest (RF) and Artificial Neural Network (ANN), reveal that the SARes-NET model exhibits superior simulation performance, highlighting it’s ability to capture intricate non-linear relationships and reduce human errors in data processing. This study establishes deep learning-guided E-A-U spatial evaluation as an innovative approach for national spatial planning, holding broader implications for national-level territorial assessments.

With the ever-growing availability of massive geo-data, deep learning has been widely applied to geoscientific questions such as sedimentary provenance analysis. However, randomly selected initial weights (and also biases) and possible loss of population diversity in traditional neural network learning remain problematic. To address this issue, in this study, we proposed a new deep neural network model by incorporating genetic algorithm (GA) and simulated annealing algorithm into the BP neural network, i.e., the GA-SA-BP model. We then applied this new model to rare earth element (REE) geochemical data of the Liuling Group of the East Qinling Orogen to investigate its provenance. Our results showed that among other deep learning algorithms, the new model presents the best performance with good measuring metrics (e.g., over 85% of accuracy, over 0.82 of F1-macro-average, F1-micro-average, and Kappa coefficient, and smallest (<0.15) Hamming distance). Here, we interpreted in accordance with the classification results that the southern margin of the North China Craton and the South Qinling Orogen are likely two major sources of the Liuling Group, suggesting a bidirectional deposition route of sediments from the north and south. Therefore, we proposed a foreland basin environment as the likely tectonic setting for the Liuling Group, which is consistent with current geological understanding. Our observations suggested that the GA-SA-BP model (or improved deep learning models) coupled with REE geochemistry is capable of provenance analysis.

Accurate reservoir identification from well-logging data is crucial for hydrocarbon exploration, yet challenges persist due to a series of factors, including limitations such as low efficiency and subjectivity of manual processing for massive datasets, as well as class imbalance and its impact on machine learning model training. This study develops an intelligent identification model using a Convolutional Neural Network (CNN) enhanced with Focal Loss, applied to real well-logging data from the Chang 8 Member of the Yanchang Formation in the Jiyuan Oilfield, Ordos Basin. A well-based data partitioning strategy is adopted to ensure the model’s generalization ability to new wells, avoiding the overoptimistic performance associated with random sample splitting. Experimental results demonstrate that the proposed model achieves an Accuracy of 84% and a Recall of 83% for oil-bearing layers. In comparison, the Random Forest model achieves a lower Recall of 56% for oil-bearing layers, and the CNN-LSTM model achieves 77%. The key influential well-logging parameters identified are bulk density (DEN), spontaneous potential (SP), true resistivity (RT), and natural gamma ray (GR). The findings confirm that the Focal Loss-enhanced CNN effectively mitigates class imbalance issues and provides a reliable, automated method for reservoir identification, offering significant practical value for the secondary interpretation of well logs in similar tight sandstone reservoirs.