Monitoring nitrogen nutrition in hydroponic lettuce using chlorophyll fluorescence imaging and improved YOLOv10

Accurate and timely diagnosis of nitrogen nutrition is often required to improve nutrient-use efficiency for the crop quality in plant factories. Hydroponic leaf lettuce has a short growth cycle with rapid response to nitrogen availability, leading to strong effects on leaf development, photosynthetic activity, biomass accumulation, and commercial quality. However, conventional diagnostic approaches with destructive chemical analysis cannot fully meet the large-scale production in recent years, due to time-consuming unsuitable for continuous monitoring. While conventional chlorophyll-meter measurements are easily affected by leaf structure and environment. It is required to capture early physiological properties. Chlorophyll fluorescence imaging can provide a rapid and non-destructive approach to characterize photosynthetic responses and nitrogen physiological variation before visible symptoms. Nevertheless, nitrogen fluorescence phenotypes are generally weak, spatially heterogeneous, and strongly coupled with irregular leaf morphology. Accurate diagnosis and real-time deployment are confined to the dense leaf distribution, overlapping boundaries, large variations in leaf size, and limited computing resources at the edge in hydroponic production systems. In this study, a lightweight object detection model was developed, termed YOLOv10n-SL, for rapid and non-destructive diagnosis of nitrogen nutrition status in hydroponic leaf lettuce. Chlorophyll fluorescence imaging was also integrated with an improved YOLOv10n framework. Individual leaves were simultaneously located to classify their nitrogen nutrition status, thereby providing both spatial and categorical information in a unified detection framework. Three targeted improvements were introduced into the original YOLOv10n model. (1) An RFAVanillaNet module was incorporated with receptive field attention convolution into the backbone network. Compared with conventional lightweight convolutional structures with fixed receptive fields, RFAVanillaNet adaptively re-weighted local receptive-field information to strengthen the response to informative spatial regions. There were improved representation of irregular leaf boundaries, local fluorescence variation, and early weak nitrogen phenotypic differences, while redundant computation was reduced to preserve a lightweight network architecture suitable for edge deployment. (2) An efficient channel attention-guided bidirectional feature pyramid network, termed ECA-BiFPN, was constructed in the neck network to improve multi-scale feature fusion. The efficient channel attention mechanism enhanced informative channel responses with nitrogen status and high dimensionality reduction, whereas the bidirectional fusion pathway promoted interaction between high-level semantic information and low-level spatial details. In addition, a high-resolution detection branch was added to preserve finer spatial information for the perception of small leaves and fine-grained weak phenotypic features under dense and overlapping. (3) The minimum point distance intersection over union loss function was adopted to optimize bounding-box regression. MPDIoU directly constrained the distances between corner points of predicted and ground-truth bounding boxes, instead of the conventional IoU-based loss functions that relied mainly on overlap, center distance, or aspect-ratio constraints. More effective geometric supervision was improved for irregular, overlapping, and morphologically diverse lettuce leaves for the localization precision and regression stability. Ablation experiments were conducted to evaluate the contribution rates of each modification. Comparative experiments were performed with mainstream object detection models during training. The results showed that YOLOv10n-SL achieved a precision of 76.7%, a recall of 75.3%, and a mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5) of 76.4%. The precision, recall, and mAP@0.5 increased by 5.3, 4.5, and 4.1 percentage points, respectively, compared with the original YOLOv10n model. As such, RFAVanillaNet reduced computational complexity for the weak-feature representation. While ECA-BiFPN and high-resolution branch improved multi-scale and fine-detail perception. MPDIoU strengthened boundary localization under complex leaf arrangements. The better balance between detection performance and computational efficiency was achieved, compared with the baseline model. Overall, the chlorophyll fluorescence imaging and YOLOv10n-SL were integrated to provide an effective for rapid and non-destructive diagnosis of nitrogen nutrition status in hydroponic leaf lettuce. The approach can also offer technical support for online crop monitoring, precise nutrient-solution regulation, and intelligent production in plant factory.