Design and Evaluation of a YOLOv8-Based Strawberry Growth-Stage Recognition and Stage-Aware Nutrient-Control Prototype

I. INTRODUCTION Image-based sensing is increasingly used in smart farms to observe crop status non-destructively and to connect plant information with irrigation, nutrient, and environmental management. Strawberry plants pass through continuously changing vegetative and reproductive states during a long cultivation period. Emergence of flowers, fruit set, fruit enlargement, and harvest readiness each imply different management requirements. Consequently, selecting a management profile from the observed crop state can be more responsive than applying a fixed schedule based only on calendar date or elapsed days. Recent strawberry vision studies have demonstrated YOLO-based ripeness detection, flower and fruit detection, and multi-stage growth monitoring. Yang et al. proposed a YOLOv8-based ripeness detector designed for illumination variation and occlusion [3]. Bai et al. addressed small flower and fruit objects on strawberry seedlings using an improved YOLO architecture [4]. Park et al. compared deep-learning detectors for seven strawberry growth stages and emphasized the importance of acquisition timing and data splitting for generalization [5]. These studies establish the feasibility of image-based growth assessment, but fewer reports evaluate the entire pathway from visual stage inference to control-profile selection, communication, and device-command generation. Nutrient management also benefits from stage-aware operation. Stage-specific NPK uptake has been used to develop fertigation guides for greenhouse strawberries [6], while closed-loop hydroponic studies show that ion correction and nutrient composition influence nutrient-use efficiency and normal crop growth [7]. Hydroponic nutrient formulations can also affect strawberry growth and yield [8]. These findings motivate stage-aware nutrient control, but they do not by themselves demonstrate that the prototype evaluated here increases yield or quality. This paper reports an industry–university field-practice prototype built around two SAMES patent-application concepts. The first concerns stable nutrient mixing and delivery using a tank, pump, Venturi structure, filtration, and pressure reduction [9]. The second concerns image-based growth-stage estimation and stage-specific environmental control, explicitly allowing a YOLO-based AI module [10]. The contribution of this study is therefore system-oriented rather than a new detector architecture: (1) a six-stage YOLOv8s model is evaluated on field-acquired strawberry images, (2) its output is linked to stage profiles and nutrient-control commands, and (3) the integrated latency and command-application reliability are quantified while clearly separating technical feasibility from unmeasured agronomic productivity effects.

A. Vision-Based Strawberry Monitoring YOLO detectors predict object locations and classes in a single real-time inference pipeline and have evolved through multiple generations with different accuracy–latency trade-offs [1], [2]. Strawberry images are challenging because leaves and stems occlude targets, flowers and young fruit can be small, illumination changes within greenhouses, and adjacent biological stages have gradual rather than discrete visual boundaries. For deployment-oriented evaluation, overall mAP should therefore be considered together with per-stage precision/recall, independent data splitting, and inference latency. B. Stage-Aware Nutrient Management and System Context Lee et al. developed a greenhouse strawberry fertigation guide from growth-stage-specific NPK uptake, supporting the principle that nutrient management can vary with plant development [6]. Lim et al. showed that nutrient correction in closed-loop strawberry hydroponics affects nutrient balance and resource-use efficiency [7]. Sahoo et al. likewise reported that hydroponic nutrient formulations can substantially alter plant performance [8]. The present study does not import the optimal nutrient concentrations reported in those studies; it evaluates the information and control structure needed to select a predefined profile after visual stage recognition. The 2023 SAMES patent application describes the lower actuation layer: a mixed-solution tank, supply pump, filter, Venturi unit, and pressure-reducing valve support stable delivery to multiple cultivation zones [9]. The 2024 application describes the upper perception/decision layer: crop images are processed by a deep-learning module to estimate growth stage, and the corresponding environment settings are used to control nutrient mixing and supply, with optional temperature, humidity, circulation, and lighting control [10]. The field-practice prototype links these two concepts through a stage-to-profile mapping and REST-based profile service, as summarized in Fig. 1.

A. Dataset and Growth-Stage Labels According to the development report, 1,800 RGB images were acquired directly in a greenhouse strawberry cultivation area at different positions and times. Images could contain multiple plant objects. Bounding boxes and management-stage labels were assigned in Roboflow, with annotation guidance covering stage boundaries, partial occlusion, duplicate boxes, and separate review of ambiguous samples. The six operational classes were Nursery, Early Post-Transplant, Vegetative Growth, Flowering, Fruit Set/Enlargement, and Harvest. This operational six-class scheme is the one used for the reported experiments and is not claimed to be identical to every example stage name in the 2024 patent specification. TABLE I DATASET DISTRIBUTION AND MODEL-TRAINING CONFIGURATION Item Setting Images / split 1,800 total; 1,260 train / 360 val. / 180 test Class counts 280 / 305 / 315 / 295 / 300 / 305 Model YOLOv8s, pretrained weights Input / epochs / batch 640×640 / 100 / 16 Optimizer / initial LR AdamW / 0.001 Augmentation Horizontal flip, ±10° rotation, brightness/saturation Hardware NVIDIA RTX 3060 12 GB Best checkpoint Epoch 87 by validation mAP@0.5:0.95 B. Training and Evaluation YOLOv8s was fine-tuned from pretrained weights. The training, validation, and test split was 70%, 20%, and 10%, respectively. The test set was not used for checkpoint selection; the best checkpoint was selected using validation mAP@0.5:0.95. Detection performance was evaluated by precision, recall, F1-score, mAP@0.5, and mAP@0.5:0.95. The supplied report did not include raw predictions, class-wise AP, a confusion matrix, or complete epoch logs, so those quantities were not reconstructed in this manuscript. The report also does not state whether images were grouped by date, bed, camera position, or individual plant before splitting. This is an important limitation because visually similar frames across subsets may inflate generalization estimates. A future replication should use group-wise splitting or an independent cultivation season for external testing. C. Stage-to-Control Prototype The predicted stage was mapped to a locally stored environment profile or to a profile retrieved through a REST API. The system then generated device-control commands. The integrated evaluation treated recognition, communication, and bench actuation as separate targets. API round-trip latency and profile-application success were measured over 1,000 request/response cycles. Device-command reflection was tested 250 times on a bench, using ±5% of the commanded target as the acceptance tolerance. End-to-end latency included image input, inference, API interaction, and command generation, but not the hydraulic settling time required for actual flow, pressure, EC, or pH to reach a target in a cultivation bed. For the two binomial success proportions reported with explicit trial counts, 95% confidence intervals were calculated using the Wilson method. Confidence intervals for detector metrics were not recomputed because raw TP, FP, FN, and prediction files were not available.

A. Growth-Stage Recognition Performance On 180 test images, the recorded macro precision was 0.918, recall 0.885, and F1-score 0.901. mAP@0.5 was 0.936 and mAP@0.5:0.95 was 0.694. Flowering yielded the highest reported class-level F1-score (0.93), whereas Nursery had the lowest recall (0.86). The development report noted that visually distinctive flowers and fruit improved recognition, while neighboring stages such as early post-transplant and vegetative growth were more easily confused because their appearances change continuously. The 0.242 difference between mAP@0.5 and mAP@0.5:0.95 indicates that performance declines as localization requirements become stricter. Stage classification is the immediate control input in the current prototype, so correct class inference is more important than highly precise box geometry. However, stronger localization would become important if later versions estimate organ size, fruit count, or other spatial growth traits. B. Runtime and Integrated Control Performance The best validation checkpoint occurred at epoch 87 of 100. Final train and validation box losses were 0.62 and 0.78, respectively. Mean inference time was 18.7 ms/image and p95 inference time was 26.4 ms/image on the RTX 3060. The reciprocal of the mean corresponds to approximately 53.5 images/s under the reported single-image benchmark, although camera acquisition and system communication must be considered separately in deployment. Profile application succeeded in 992 of 1,000 requests; the eight failures were recorded as response timeouts. The resulting Wilson 95% confidence interval was 98.43–99.59%. In the bench test, 96.8% of 250 command trials were reflected within ±5% tolerance, corresponding to approximately 242 successful trials and a Wilson 95% interval of 93.81–98.37%. End-to-end p95 latency from image input to command generation was 258 ms. These results show sub-second command-generation capability, but they do not establish closed-loop hydraulic stability after valves and pumps operate.

A. Meaning of the Perception-to-Control Link The principal contribution is not a new YOLO architecture but the evaluation of a complete perception-to-control path in a single prototype. Existing strawberry vision work has focused mainly on ripeness, flower/fruit detection, or multi-stage monitoring [3]–[5]. Here, the model output is converted into a stage profile and then into a device command through a measurable communication path. Separating the perception/decision layer from the nutrient-delivery layer also supports modular replacement of the detector, profile service, or nutrient hardware without redesigning the entire system interface. B. Agronomic Interpretation and Study Limits Prior agronomic studies support the premise that nutrient requirements and nutrient-management strategies vary with strawberry growth and hydroponic conditions [6]–[8]. It is therefore reasonable to hypothesize that reliable growth-stage recognition can support more adaptive nutrient management. Nevertheless, this study did not compare crop yield or fruit quality between a fixed schedule and AI-driven stage control. The present results should therefore be interpreted as evidence of technical operability and responsiveness, not as proof of productivity improvement. A cultivation validation should compare at least three treatments under the same greenhouse conditions: (1) a fixed conventional schedule, (2) stage-specific profiles selected by a human manager, and (3) profiles selected automatically from YOLOv8 predictions. Outcomes should include marketable yield, fruit count, mean fruit weight, soluble solids (°Brix), defect rate, supply and drainage EC/pH, irrigation volume, and nutrient-use efficiency. Replicated beds or blocks are needed to link recognition performance to agronomic outcomes. Additional technical limitations also remain. The test set contained only 180 images, and independence by plant, bed, date, or camera position was not documented. Class-wise AP, confusion matrices, precision–recall curves, object-instance counts, and repeated-training variance were unavailable. Only YOLOv8s was tested, so accuracy–latency trade-offs against lighter models remain unknown. The integrated experiment measured API behavior and bench command reflection but not zone-to-zone flow uniformity, pressure variation, EC/pH settling, or actuation dynamics in production beds. Finally, stage boundaries may flicker between adjacent classes in video; temporal voting, confidence thresholds, hysteresis, or minimum-duration rules should be considered before fully autonomous greenhouse control. C. Deployment Safeguards and Cultivation Validation A practical autonomous controller should avoid changing profiles on every isolated frame. Because adjacent strawberry stages are visually continuous, deployment should combine a confidence threshold with temporal stabilization, such as majority voting over consecutive frames, a minimum stage-persistence time, or hysteresis between forward and backward stage transitions. Communication failures should also invoke a safe fallback profile rather than leaving a partially updated command state. Manual override and event logging are desirable so that growers can audit when a stage change caused a profile change. The agronomic benefit of the proposed workflow can be tested with a three-treatment greenhouse experiment that separates the benefit of stage-aware management from the benefit of automated visual recognition. A fixed-schedule treatment provides the operational baseline; a human-selected stage-profile treatment estimates the effect of stage-aware management without recognition error; and an automatically selected YOLOv8 treatment quantifies the additional effect of automation. The three treatments should use the same cultivar, planting date, substrate, greenhouse zone, and profile library Primary outcomes should include marketable yield, fruit number, mean fruit weight, soluble solids (°Brix), and defect rate. Process outcomes should include supplied and drained volume, EC/pH, nutrient-use efficiency, zone pressure/flow uniformity, and the frequency of stage or control overrides. Replicated cultivation beds should be treated as experimental units, and the analysis should account for repeated harvests over time. This design would directly test whether the fast and reliable command-generation performance reported here produces measurable cultivation benefits.

A YOLOv8s-based prototype was developed for six-stage strawberry growth recognition and stage-aware nutrient-control command generation using 1,800 greenhouse RGB images. The reported test performance was 0.918 macro precision, 0.885 recall, 0.901 F1-score, 0.936 mAP@0.5, and 0.694 mAP@0.5:0.95, with 18.7 ms/image mean inference time. The integrated prototype achieved 99.2% profile-application success, 96.8% command reflection within ±5% tolerance, and 258 ms p95 latency from image input to command generation. These results support the feasibility of connecting visual growth-stage estimation to nutrient-control workflows. Independent-season evaluation, hydraulic measurements, temporal stabilization, and controlled cultivation trials are required before claims regarding yield, quality, or nutrient-use improvement can be made.