IntroductionAccurate segmentation of plant-column and path regions is essential for visual navigation and baseline extraction of autonomous agricultural platforms in Lycium barbarum gardens. During the non-fruiting period, sparse foliage, exposed branches, fallen leaves, weeds, similar soil backgrounds, and variable illumination weaken the visual contrast between plant-column and path regions, rendering conventional segmentation methods based on color or texture unreliable.MethodsWe propose a deep learning-based semantic segmentation method tailored to non-fruiting L. barbarum garden images. A region-based annotation strategy was used to label continuous plant-column and path regions following the requirements for navigation baseline extraction. The encoder–decoder model incorporates depthwise separable convolutions, skip connections, and multi-scale contextual feature extraction to enhance feature representation of complex field scenes. Image data were collected from multiple gardens across different years, growth stages, lighting conditions, irrigation regimes, and soil types; data augmentation was implemented during model training. Model performance was validated using an independent internal test set containing 230 images, and cross-garden generalization was assessed on two unseen field sites with limited fine-tuning.ResultsOn the internal test set, the model achieved precision of 95.24 ± 2.66%, recall of 94.68 ± 3.15%, mean intersection over union (mIoU) of 90.10 ± 5.43%, Dice coefficient of 94.68 ± 3.15%, and mean pixel accuracy (mPA) of 94.64 ± 3.14%. For the two unseen gardens, the raw mIoU reached 86.84% and 84.60%, and increased to 92.43% and 90.23% after fine-tuning with a small set of newly annotated images.DiscussionThe proposed method yields reliable region-level segmentation for non-fruiting *L. barbarum* garden scenes, alleviating domain shifts while preserving strong pixel-level and region-overlap performance. Minor boundary discrepancies may arise from annotation ambiguity. This approach offers solid technical support for navigation baseline extraction in agricultural robotic systems.
Recently, energy consumption forecasting plays an important role in power systems, smart grids, and power generation networks. Given the complex nature and time dependence of electricity consumption data, applying machine learning techniques improves energy forecasting accuracy and performance. In this research, a hybrid XGBoost–Random Forest (XGBoost–RF) model is proposed to make better the forecasting accuracy. To improve the performance of the proposed model, a residual learning-based forecasting framework is presented to learn the remaining prediction errors. In the forecasting procedure, important prediction metrics such as Mean Absolute Error (MAE), and coefficient of determination R2 are evaluated. The experimental results show that the proposed hybrid model achieves an MAE of 15.8683, Root Mean Square Error (RMSE) of 22.1588, R2 of 0.985807, and Mean Absolute Percentage Error (MAPE) of 1.3118%. Compared with the base XGBoost model, the proposed model is able to decrease MAE by 11.19% and RMSE by 11.30%. In addition, the SHapley Additive Explanations (SHAP) method is used to examine the effect of the input features on the model predictions and improve model interpretability. The SHAP results show that temporal features and lag-based variables have an important role in energy consumption forecasting. In particular, lag_1 and lag_168 are identified as the most important features. The experimental results confirmed that the proposed hybrid model achieved lower prediction errors and a higher R2 than the other evaluated models. These findings show that the proposed XGBoost-RF residual learning approach can improve energy consumption forecasting accuracy while providing useful information about the important forecasting features.
Plant leaf diseases threaten agricultural productivity and motivate accurate automated monitoring. This study presents LARU-Net, a unified architecture for binary cucumber lesion segmentation (lesion versus background) with an auxiliary continuous severity-regression output on a bounded 0–4 scale. The network integrates a Disease Feature Enhancement Autoencoder, Residual Multi-Scale Encoder, transformer bottleneck, Attention-Guided Decoder, and Boundary Refinement Module. On the declared held-out test partition (1,080 images), the archived full-model results were Dice 0.9230, IoU 0.8660, Boundary F1 0.9350, and precision 0.9460; the previously emphasized Dice value of 0.9456 is a validation value and is now identified as such. Severity estimates yielded an overall MAE of 0.29, RMSE of 0.37, and R2 of 0.90. These values are single-run point estimates from one greenhouse image collection. Plant/session identifiers, complete acquisition and annotation provenance, repeated-run logs, and an external dataset were not preserved in the supplied study archive; therefore, the results support in-domain evaluation but not broad cross-greenhouse or cross-season generalization. The split was stratified at image level; independence at plant, leaf, or acquisition-session level could not be verified. Confidence intervals and significance tests were not reconstructed because per-image predictions and repeated-run logs were unavailable.
IntroductionThe SingHealth Duke–NUS Academic Medical Center (AMC) oversees 3,000 clinical faculty members and processes 400 appointments and promotions annually; yet, because its promotion and tenure (P&T) policies are dispersed across 30 documents, the relevant information is time-consuming to locate.MethodsIn response, we designed and evaluated “AskADD,” an AI-assisted chatbot for P&T guidance using a two-phase (alpha-beta) approach. Our aim was to iteratively design and refine such a chatbot to meet faculty-identified requirements; capture faculty’s perceptions of its usability, responsiveness, empathy, and accessibility; assess the changes in these self-perceived evaluations between iterations; and determine faculty’s perceptions of its effectiveness relative to traditional methods of locating P&T information. Thirty-five faculty and staff, recruited by email, tested an initial prototype built on Microsoft low/no-code platforms, and following refinements informed by their feedback, they evaluated a subsequent custom generative pre-trained transformer built on the institutional platform. Feedback gathered via a structured survey was analyzed descriptively.ResultsParticipants reported improvements in the perceived usability, empathy, and query resolution, and the proportion of promoters rose from 23% to 51%. Faculty valued accurate, contextually relevant responses requiring minimal rephrasing, direct links to source documents, and integration of P&T and human-resource information; approximately half nonetheless remained passively satisfied, indicating a preference for personalization such as CV-based gap analysis.DiscussionThese preliminary findings serve as proof of concept and offer a transferable, platform-agnostic framework for developing and evaluating AI policy-guidance tools in academic medical centers, which merits confirmation with validated instruments and objective performance measures.
Artificial intelligence (AI) for Science is shifting from systems that execute scientific tasks to systems that judge and generate research questions. Recognition-trained systems, including Reinforcement Learning from Community Feedback (RLCF), convert citations, review scores, or preference signals into supervisory signals for future research. This Perspective argues that such systems risk optimizing the wrong evaluative object: recognition is an institutional compression of scientific value, recording what a community has already rewarded, whereas scientific judgment must also preserve interpretive space, the socially available future of a claim before institutional recognition stabilizes a dominant appraisal of its value. Taste is therefore recast as one mode of appraising interpretive space, and answerability as the governance principle that keeps it open. The article proposes an interpretive layer above reward models as a research agenda for preserving interpretive space rather than merely predicting recognized success.
