
Herbicide field trials require accurate identification of plant species and assessment of herbicide-induced damage across diverse environments. While general-purpose vision foundation models have shown promising results in complex visual domains, their performance can be limited in agriculture, where fine-grained distinctions between species and damage types are critical.In this work, we adapt DINOv2, a general-purpose vision foundation model, to herbicide trial characterization. Using self-supervised learning on 6 million image tiles from 340,554 herbicide-trial images spanning multiple years, locations, sensors, and crop species, the model learns representations optimized for herbicide trial images.Our domain-specific model significantly outperforms DINOv2 in both species identification (F1 score from 0.91 to 0.94) and damage classification (from 0.26 to 0.33) under in-domain conditions. Under unseen conditions (new locations and times), it achieves greater gains (species identification from 0.56 to 0.66; damage classification from 0.17 to 0.27), although absolute damage performance remains relatively low. In domain-shift scenarios, such as drone imagery, species classification improves from 0.49 to 0.60.We show that domain-specific pretraining enhances segmentation accuracy, particularly in low-annotation regimes. An annotation-efficiency analysis reveals that, under unseen conditions, the domain-specific model achieves 5.4% higher F1 score than DINOv2, while using 80% fewer labeled samples.On the public PhenoBench benchmark, which addresses a simpler crop/weed/soil segmentation task, a modest global F1 gain (0.736 to 0.742) indicates more limited evidence for cross-task transfer.These results demonstrate that large-scale, domain-specific self-supervised pretraining significantly improves in-domain performance and robustness under domain shift, while reducing manual annotation efforts, and providing a foundation for scalable and automated herbicide trial analysis.
Continuous biothermal monitoring is valuable for evaluating dairy calf health, but conventional methods are usually labor-intensive or invasive. This study evaluated a minimally invasive subcutaneous implant (SCI) system and assessed a LightGBM regression model to mitigate peripheral environmental thermal coupling. In a two-farm trial (n = 720 calves), SCIs were deployed under three presurgical skin preparation protocols. Telemetry was evaluated across development (n = 327 calves; 164,383 hourly sliding-window prediction epochs) and out-of-time (OOT) evaluation cohorts (n = 123 calves; 82,724 epochs; 100 novel, 23 overlapping), separated by an empirical washout buffer (≥ 532 h) to prevent feature leakage. Early implantation failure was low at 1.39% (10/720), with increasing procedural sequence associated with lower failure odds (p = 0.007), whereas skin preparation protocol revealed no statistically detectable difference in 60-day retention (p = 0.77). System performance was constrained by a 9.3% hardware failure rate (67/720; predominantly pre-insertion initialization failures). Although active devices achieved 90.5% telemetry integrity, intention-to-treat data completeness was 82.7%. Raw subcutaneous temperature showed marked disagreement with the gastrointestinal reference (TC) (RMSE = 1.094°C, CCC = 0.16). The LightGBM model reduced ambient dependence, decreasing error to RMSE = 0.403°C on matched OOT epochs, though concordance remained modest (CCC = 0.35). The model systematically underestimated hyperthermic reference temperatures (TC > 39.5°C), dampening peak febrile excursions. While subcutaneous implants demonstrate procedural feasibility, commercial implementation requires eliminating electronic infant mortality, accommodating telemetry gaps, and resolving hyperthermic attenuation.
Macronutrient deficiencies, including nitrogen (N), phosphorus (P), and potassium (K), are a major cause of yield loss and reduced crop quality in precision agriculture. Although deep learning models have demonstrated strong performance in visual plant disease and nutrient deficiency recognition, their black-box nature limits interpretability, trust, and practical adoption by agronomists and farmers. In this work, we propose a sign-guided neurosymbolic framework for coffee leaf nutrient deficiency diagnosis that explicitly integrates agronomic knowledge with deep visual perception. The proposed approach introduces a sign-guided learning paradigm in which eight agronomically meaningful visual symptom indicators are first predicted from coffee leaf images using a ResNet-50 backbone. These indicators are then processed by a differentiable fuzzy rule engine that encodes expert-defined nutrient–symptom relationships to infer deficiencies. Unlike conventional CNNs, the model provides both class predictions and interpretable intermediate reasoning through sign activation scores and rule-based decision traces. The framework is further extended with a retrieval-augmented generation (RAG) module to provide knowledge-grounded explanations and remedy suggestions based on curated agronomic information. The framework is evaluated on a combined coffee leaf dataset constructed from the CoLeaf and BRACOL benchmarks. Experimental results demonstrate strong performance, achieving an accuracy of 0.88, F1-score of 0.87, AUC-ROC of 0.93, and PR-AUC of 0.91. Qualitative analysis using Grad-CAM and sign-activation visualizations demonstrates that the proposed approach provides semantically aligned explanations beyond conventional visual attribution methods. The proposed framework establishes an interpretable decision-support approach for coffee nutrient deficiency analysis, while further validation across diverse crops and field conditions remains necessary.
Despite rising labor costs and growing labor shortages, adoption of machine harvesting in fresh market blueberry production has been limited by issues such as high rates of bruising and the significant loss of fruits on the ground. Machine settings strongly affect fruit quality and harvesting performance, yet they are manually configured based on operator experience, with no real-time performance feedback. Towards addressing these limitations, this paper presents a sensing proof-of-concept for analyzing berry flow through the harvester by estimating the pixel-wise proportion of blue vs green berries and volumetric flow rate from overhead RGB-D images of the harvester's conveyor. An automated labeling pipeline combining a domain-specific YOLOv8 patch detector with SAM2-based mask generation was developed to generate a realistic large-scale berry dataset from unlabeled images captured during commercial over-the-row blueberry harvesting operations. Compared to a baseline SegFormer model trained on a smaller dataset of manually annotated images, our fine-tuned SegFormer-trained and validated on the large-scale automatically generated dataset-achieves superior performance, with green berry IoU improving from 0.377 to 0.817 and recall increasing from 0.414 to 0.911. A scene-level YOLOv8-l detection model trained and evaluated on the same dataset achieves mAP(50-95) of 0.966 and 0.898 for blue and green berries, respectively. Both models' performance is further confirmed by an independent evaluation on a manually annotated subset of the validation images. We also compared two approaches for predicting harvest yields across five machine-setting trials: a depth-based method using the RGB-D sensor and a detection-based method using only RGB images, achieving R2 = 0.989 and R2 = 0.894, respectively, against ground-truth load cell measurements from these trials. To support future research, we publicly release our datasets and models.
To address the challenges of individual tray traceability and inefficient group-level growth monitoring in highdensity, severely occluded vertical rice seedling cultivation environments, this study proposes an end-edge-cloud collaborative monitoring framework that integrates depth-prior localization with a multi-task Vision Transformer (ViT). At the perception front-end, the Depth Anything v2 (DAv2) monocular depth estimation model is employed to achieve pixel-level basket segmentation using depth geometric priors, fundamentally overcoming the spatial redundancy inherent in traditional 2D detection boxes. At the diagnosis back-end, a Multi-Receptive Field Feature Enhancement (MRFFE) module is designed to decouple gradient competition among three concurrent tasks-leaf age, leaf color, and leaf disease-through multi-branch asymmetric and dilated convolutions, raising the disease identification F1-score from 90.01 % for the vanilla ViT baseline to 94.78 % for the same backbone with MRFFE, and to 96.85 % for the final DINOv2-base-reg-MRFFE model. At the deployment level, a three-tier collaborative architecture cascading from end-side tray cropping to edge-side localization and cloudside diagnosis is established: the end side runs a lightweight YOLOv11n model for real-time tray detection and cropping; the edge side performs depth segmentation and tag OCR; and the cloud side hosts the DINOv2-basereg-MRFFE multi-task diagnosis model (quantized to 98 MB, 64.5 FPS). INT8 quantization compresses the core model by 90.4 % (1022.05 MB -> 98.04 MB) with only 0.22 percentage points of accuracy degradation. Experimental results demonstrate that the proposed DINOv2-base-reg-MRFFE achieves a mean F1-score of 97.28 %, substantially outperforming baselines such as YOLOv11-MTL (58.12 %). Collectively, this study explores a shift in the operating unit, from individual fine-grained segmentation to group-unit perception, offering a validated prototype and technical basis for real-time unmanned management in a vertical rice seedling cultivation setting.