Inversion of soil organic carbon content based on the two-point machine learning method.

Soil organic carbon (SOC) is vital for the global carbon cycle and environmentally sustainable development. Meanwhile, the fast, convenient remote sensing technology has become one of the notable means to monitor SOC content. Nowadays, limitations are found in the inversion of SOC content with high-precision and complex spatial relationships based on scarce ground sample points. It is restrained by the spatial difference in the relationship between SOC content and remote sensing spectra due to the problem of different spectra for the same substance and the influence of topographic and environment (e.g. vegetation and climate). In this regard, the two-point machine learning (TPML) method, which can overcome above problems and deal with complex spatial heterogeneity of relationships between SOC and remote sensing spectra, is used to invert the SOC content in Hailun County, Heilongjiang Province, combined with derived variables from Sentinel-1, Sentinel-2, topography and environment. Based on 10-fold cross-validation and t-test, results indicate that the TPML method boasts the highest inversion accuracy, followed by random forest, gradient boosting regression tree, partial least squares regression and support vector machine. The average r, MAE, RMSE, and RPD of TPML are 0.854, 0.384 %, 0.558 %, and 1.918. Further, the TPML method has been proven to be equal to evaluating the uncertainty of inversion results, by comparing the actual and theoretical error of the inversion result in one subset. The spatial inversion result of SOC content with 10 m resolution by TPML is smoother and has more real details than other models, which are consistent with the distribution of SOC content in different land use types. This study provides both theoretical and technical guidance for using TPML method combined with spectral information of remote sensing to predict soil attributes and offer accurate uncertainty estimation, thereby opening up the opportunity for low-cost, high-precision, and large-scale SOC inversion.

A 10-m national-scale map of ground-mounted photovoltaic power stations in China of 2020.

We provide a remote sensing derived dataset for large-scale ground-mounted photovoltaic (PV) power stations in China of 2020, which has high spatial resolution of 10 meters. The dataset is based on the Google Earth Engine (GEE) cloud computing platform via random forest classifier and active learning strategy. Specifically, ground samples are carefully collected across China via both field survey and visual interpretation. Afterwards, spectral and texture features are calculated from publicly available Sentinel-2 imagery. Meanwhile, topographic features consisting of slope and aspect that are sensitive to PV locations are also included, aiming to construct a multi-dimensional and discriminative feature space. Finally, the trained random forest model is adopted to predict PV power stations of China parallelly on GEE. Technical validation has been carefully performed across China which achieved a satisfactory accuracy over 89%. Above all, as the first publicly released 10-m national-scale distribution dataset of China's ground-mounted PV power stations, it can provide data references for relevant researchers in fields such as energy, land, remote sensing and environmental sciences.

Glenoid segmentation from computed tomography scans based on a 2-stage deep learning model for glenoid bone loss evaluation.

BACKGROUND: The best-fitting circle drawn by computed tomography (CT) reconstruction of the en face view of the glenoid bone to measure the bone defect is widely used in clinical application. However, there are still some limitations in practical application, which can prevent the achievement of accurate measurements. This study aimed to accurately and automatically segment the glenoid from CT scans based on a 2-stage deep learning model and to quantitatively measure the glenoid bone defect. MATERIALS AND METHODS: Patients who were referred to our institution between June 2018 and February 2022 were retrospectively reviewed. The dislocation group consisted of 237 patients with a history of ≥2 unilateral shoulder dislocations within 2 years. The control group consisted of 248 individuals with no history of shoulder dislocation, shoulder developmental deformity, or other disease that may lead to abnormal morphology of the glenoid. All patients underwent CT examination with a 1-mm slice thickness and a 1-mm increment, including complete imaging of the bilateral glenoid. A residual neural network (ResNet) location model and a U-Net bone segmentation model were constructed to develop an automated segmentation model for the glenoid from CT scans. The data set was randomly divided into training (201 of 248) and test (47 of 248) data sets of control-group data and training (190 of 237) and test (47 of 237) data sets of dislocation-group data. The accuracy of the stage 1 (glenoid location) model, the mean intersection-over-union value of the stage 2 (glenoid segmentation) model, and the glenoid volume error were used to assess the performance of the model. The R2 value and Lin concordance correlation coefficient were used to assess the correlation between the prediction and the gold standard. RESULTS: A total of 73,805 images were obtained after the labeling process, and each image was composed of CT images of the glenoid and its corresponding mask. The average overall accuracy of stage 1 was 99.28%; the average mean intersection-over-union value of stage 2 was 0.96. The average glenoid volume error between the predicted and true values was 9.33%. The R2 values of the predicted and true values of glenoid volume and glenoid bone loss (GBL) were 0.87 and 0.91, respectively. The Lin concordance correlation coefficient value of the predicted and true values of glenoid volume and GBL were 0.93 and 0.95, respectively. CONCLUSION: The 2-stage model in this study showed a good performance in glenoid bone segmentation from CT scans and could quantitatively measure GBL, providing a data reference for subsequent clinical treatment.

Diagnose Like a Radiologist: Hybrid Neuro-Probabilistic Reasoning for Attribute-Based Medical Image Diagnosis.

During clinical practice, radiologists often use attributes, e.g., morphological and appearance characteristics of a lesion, to aid disease diagnosis. Effectively modeling attributes as well as all relationships involving attributes could boost the generalization ability and verifiability of medical image diagnosis algorithms. In this paper, we introduce a hybrid neuro-probabilistic reasoning algorithm for verifiable attribute-based medical image diagnosis. There are two parallel branches in our hybrid algorithm, a Bayesian network branch performing probabilistic causal relationship reasoning and a graph convolutional network branch performing more generic relational modeling and reasoning using a feature representation. Tight coupling between these two branches is achieved via a cross-network attention mechanism and the fusion of their classification results. We have successfully applied our hybrid reasoning algorithm to two challenging medical image diagnosis tasks. On the LIDC-IDRI benchmark dataset for benign-malignant classification of pulmonary nodules in CT images, our method achieves a new state-of-the-art accuracy of 95.36% and an AUC of 96.54%. Our method also achieves a 3.24% accuracy improvement on an in-house chest X-ray image dataset for tuberculosis diagnosis. Our ablation study indicates that our hybrid algorithm achieves a much better generalization performance than a pure neural network architecture under very limited training data.