CFIHL: a variety of chlorophyll a fluorescence transient image datasets of hydroponic lettuce

healthy Arabidopsis and cannot provide a basis for disease diagnosis. Therefore, Rousseau et al. further used Arabidopsis plants inoculated with the DC3000 bacterial strain as the research object, performing chlorophyll fluorescence imaging on each culture dish every two days to obtain TIFF images of Fo, Fm, and Fv/Fm for judging the status of diseased Arabidopsis. In addition, Pavicic et al. (2021) proposed an annotated dataset of diseased Arabidopsis fluorescence images and a threshold segmentation algorithm to segment the normal and diseased leaf areas in the images for disease diagnosis . Currently, there is a lack of similar chlorophyll fluorescence datasets for hydroponic lettuce research, which mainly uses RGB image data to evaluate lettuce growth status. Islam et al. (2024) used an improved Mask R-CNN method to detect and segment lettuce seedlings from the background of the seedling growth tray and estimate seedling growth . Abidi et al. (2024) collected an RGB image dataset of nutrient deficiencies in hydroponic lettuce and proposed a deep learning image processing method to classify these deficiencies.However, RGB images can only capture superficial color and texture information, making it difficult to obtain internal physiological parameters of lettuce leaves and accurately judge lettuce growth status.Therefore, this paper collects a ChlF dataset of hydroponic plants and tests region of interest extraction algorithms for fluorescence images, providing data for accurately estimating lettuce seedling status using ChlF technology.Hydroponically grown lettuce often uses sponges for anchorage, making it more challenging to extract the fluorescent area compared to plants grown in substrates, due to the common presence of microbes or green algae within the cultivation sponges (Weizhong, 2023). During the capture of ChlF systems, the fluorescence from microbes or green algae present in the cultivation sponges is also captured, resulting in minimal differences between them in grayscale images taken under blue light, affecting the extraction of areas of interest in the ChlF images of hydroponic lettuce seedlings. In microbial fluorescence images, such as that of Chlamydomonas reinhardtii, exposure to excitation light produces a lower level of Non-Photochemical Quenching (NPQ). In the chloroplasts of algae like Chlorella or Chlamydomonas, effective carbohydrate metabolism decomposition leads to a much reduced redox balance compared to plant chloroplasts (Alric, 2010). Differences exist in the fluorescence parameters under ChlF induction kinetics between lettuce and green algae (Papageorgiou et al., 2007).However, the threshold segmentation algorithm cannot determine the category of the area. In practice, the grayscale of the sponge block of green algae in the initial fluorescence image is very similar to that of the seedling canopy, leading to mis-segmentation of areas with similar grayscales. Given the excellent feature extraction and generalization performance of deep learning algorithms in image recognition, a deep learning semantic segmentation algorithm is considered for extracting the seedling canopy area from a single fluorescence image. Testing has shown that fusing multiple fluorescence images can improve the extraction accuracy of the seedling canopy area, effectively reduce background noise interference, and more accurately extract the fluorescence values. Image fusion, by integrating information from multiple sources with complementary datasets to enhance a single image (Qu et al., 2024), makes the generated images more detailed and reliable through the combination of key information from multiple source images (Karim et al., 2023). At this stage, it has been proposed to introduce accuracy of advanced visual tasks such as detection, recognition, and segmentation into the design of the loss function during the fusion phase (Jiayi, 2023;Liu et al., 2023), guiding the fusion process from a decision-making level to achieve better segmentation results (James and Dasarathy, 2014). This article introduces the Hydroponic Lettuce Seedling ChlF Dataset. The dataset comprises various transient images of ChlF from different cultivation states, with annotations marking the positions of the entire lettuce canopy projection. The aim is to accelerate the development of algorithms for extracting ChlF areas in hydroponically grown lettuce seedlings. The data for this article is publicly released at https://github.com/Yiyu-Jiang/CFIHL. For user convenience, the original .igr data files (which can be The primary contributions of this dataset are summarized as follows:(1) The CF-imager has provided a ChlF image set of the whole hydroponic lettuce seedlings, with the dataset being approximately 1.3 G in size. The dataset includes more than 300 sets of images, each set containing ten transient ChlF images. The dataset offers a variety of physiological images of plants and annotations for individual seedlings.(2) Compared to the existing lettuce seedling datasets, this paper has conducted tests with multiple image algorithms, comparing threshold segmentation and deep learning algorithms, as well as the segmentation effects of images before and after fusion. The fused image can extract the ChlF value of lettuce more accurately, and can further accurately obtain parameters such as photochemical quantum yield, thereby reflecting the photosynthetic utilization efficiency and lettuce yield.Lettuce was cultivated in the artificial light plant factory of China Agricultural University in 2022 and 2023. During the cultivation of hydroponic lettuce, we used an LED lamp (WR-16W) with an R ratio of 1.2, a light intensity of 200 µmolm -2 s -1 , and a photoperiod of 16 hd -1 . The temperature during the light period was 22±1 °C, with a relative humidity of 70±5% and a CO 2 concentration of 800±50µmolmol -1. During the dark period, the temperature was 18±1°C, the relative humidity was 65±10%, and the CO 2 concentration was not controlled. The seedling raising period for hydroponic lettuce was 20 days. The nutrient solution used was the Yamazaki lettuce formulaYan (2019), with an EC of 1.0-1.2 mScm -1 and a pH adjusted to 6.0-6.5 mol/L.In this paper, the seedlings grown for about 21 days were placed in the CF-Imager(Technologica, UK) instrument, then photographed, as shown in Fig. 1A. The CF-Imager system uses a high-performance AVT Stingray SXGA+ low-noise line scan CCD camera with a resolution of 1392 × 1040. When collecting ChlF images, the camera needs to respond quickly, which sacrifices some resolution. Therefore, the image resolution was set to 696 × 519. The imaging frequency is 50 FPS, and the readout noise is less than 12e. This ChlF imaging system used 470 nm blue light as the excitation light. Based on the principles of slow fluorescence kinetics, the detection light was set to 5 µmolm -2 s -1 , the actinic light to 200 µmolm -2 s -1 , and the saturating pulse light to 6000 µmolm -2 s -1 with a duration of 800 ms, establishing the excitation process (Daquan, 2002).The measurement principle is illustrated in Fig. 1B. After 30 minutes in the dark, a fully dark-adapted leaf was exposed to measuring light. Within 1-2 minutes, as the fluorescence level stabilized, the minimum initial fluorescence Fo image was obtained. Then, a saturating pulse light was applied and turned off after one pulse to obtain the maximum fluorescence image Fm. Therefore, the fluorescence parameter image Fv/Fm (Fv=Fm-Fo) can be calculated, representing the potential photochemical efficiency of PSII.Secondly, the actinic light of 200 µmolm -2 s -1 was turned on to induce leaf photosynthesis. After a few minutes, the leaf's photosynthetic rate reached a steady state, and the steady-state fluorescence(Fs) image was obtained. At this point, a pulse of saturating light was applied and then turned off to obtain the maximum fluorescence image under light Fm'. The actual quantum efficiency of PSII(Φ PSII ) and the non-photochemical quenching coefficient (NPQ), defined as (Fm -Fm')/Fm', were then calculated.Using the method described above to stimulate fluorescence in vegetables, various ChlF transient images were obtained, as shown in Fig. 1C, all with a resolution of 696 × 519. The ChlF images are essentially grayscale images with a single channel. For intuitive description, this paper maps the grayscale values of some ChlF images to the RGB color space (Oxborough, 2004), rendering them as pseudocolor images. The main ChlF parameters and descriptions are shown in Table 1: Table 1. Commonly abbreviated and described ChlF parameters.This list is only used to identify the most common parameters, as detailed in the review (Murchie and Lawson, 2013) This study uses Precision, Recall and intersection over union (IoU) as evaluation indicators for training target detection models. The range of these three evaluation metrics is [0, 1], the detailed descriptions are provided in Table 2. In Precision indicates the percentage of correctly identified pixels among the total number of identified pixels.RecallTP TP+FN × 100%Recall indicates the percentage of correctly identified pixels that meet the requirements to the total number of pixels in the test set.IoUTP-FP TP+FN × 100%IoU (Intersection over Union) is the ratio of the number of elements in the intersection of two sets to the number of elements in their union.All models in this study were trained on a single GPU. The hardware configuration for training and testing the network models includes an i9-12900K CPU, NVIDIA GeForce RTX 3070 GPU, and 32GB DDR4 RAM. The dataset was randomly divided into training, test, and validation sets in an 8:1:1 ratio. The input size of each fluorescence image is 696 × 519 pixels.The DeepLabv3+ algorithm (YU et al., 2024) sets the number of iterations (Epochs) to 300 times, with an initial learning rate of 0.001 and a batch size of 4. SGD is used as the optimizer, with a weight decay of 0.0001 and a momentum of 0.9. The training weights are verified and saved every 5 iterations.In this paper, the automatic threshold segmentation method (Zhao et al., 2023) The image segmentation results indicator results are shown in Table 3. Among them, the threshold segmentation algorithm has the best segmentation effect for the Fm parameterized image with an IoU of 84.89%, but there is still a 15% error with the actual canopy. The deep learning DeepLabv3+ network has higher segmentation accuracy in most images than the threshold segmentation method, and the best segmentation effect is for the Igi image, with an IoU of 87.27%, an increase of about 2%.Automatic threshold segmentation can roughly distinguish the seedling canopy from the background. However, the presence of green algae can cause the background threshold to be similar to that of the seedling, leading to incorrect segmentation. In contrast, deep learning methods automatically extract features and are wellsuited for segmenting complex scenes with similar pixels. Consequently, deep learning has shown improved segmentation effectiveness compared to simple grayscale threshold segmentation.This study uses the DeepLabv3+ algorithm to segment the seedling canopy area of fluorescence images.The Precision, Recall, and IoU are slightly improved compared to threshold segmentation, but the IoU is still below 90%. Therefore, it is necessary to consider other segmentation methods to improve the accuracy The fusion results of Igi and Φ PSII for the same seedling using four popular fusion algorithms (Zhang et al., 2020) and the SeAFusion algorithm are shown in Fig. 3B. The segmentation visualization results of the DeepLabv3+ network before and after the fusion of three seedlings are shown in Fig. 3C. This paper uses indicators such as entropy (EN), standard deviation (SD), spatial frequency (SF) and sum of difference correlations (SCD) to characterize the effect of image fusion (Zhang et al., 2019). The higher EN and SD indicate that the method can better preserve the source image. These images contain a large amount of information and have high pixel contrast (Liu et al., 2023).The comparison results of each fusion algorithm are shown in Table 4. Among them, the fusion image texture information obtained by the CNN, IFEVIP, LatLRR, and MST_SR algorithms has a greater loss. The fused image obtained by the SeAFusion algorithm not only ensures the contrast between the seedling canopy and the sponge block, but also fully retains the seedling texture. Combined with the visualization results in Fig. 3B, the SeAFusion algorithm in Table 4 has higher AG, SD, and EN parameter values and clearer edges compared to other fusion algorithms from a visual perspective.Finally, to verify that the fused image segmentation effect is better, this paper uses the Deeplabv3+ algorithm to compare the segmentation results of the fused images of the two fusion methods (Igi is fused with Φ PSII and Fm images respectively), and the results are shown in Table 5. Compared with the segmentation results in Table 3, the fusion Precision, Recall and IoU ratio parameters are increased by 0.25%, 0.35% and 0.56% respectively, which is helpful for the accurate extraction of fluorescent areas of interest such as the lettuce canopy. Among them, the fluorescent image segmentation results of three seedlings are shown in Fig. 3C. There are many wrongly segmented areas (green and magenta areas) in the seedling canopy segmentation using the Deeplabv3+ algorithm without image fusion, while the wrongly segmented areas (green and magenta areas) are significantly reduced after fusing Igi and Φ PSII . This is because the SeAFusion algorithm transmits the semantic information required for high-level visual tasks (segmentation) back to the fusion network, thereby enabling the fusion network to effectively retain the semantic information in the source image. Wang et al.However, due to limited information in a single image, the paper also considers the fluorescence value differences between the seedling canopy and the background, adopting canopy segmentation after image fusion. Further testing of multiple image fusion algorithms revealed that the fused images provided better segmentation of the lettuce canopy. Thus, using complementary information from different fluorescence images improves the segmentation of the seedling canopy's edge contours.There are some limitations in this study. For example, only some image fusion methods were tested during image fusion, and all fusion combinations were not enumerated due to the limited length of the article. Only two fluorescence parameter images were considered as input, and multiple image inputs were not considered. The fusion algorithm tested in this study is designed for merging infrared and visible light.It is recommended to design a fusion method suitable for the fusion of multiple fluorescence images in the future to help improve the acquisition of ChlF values. In addition, for the current dataset, future research should further expand the sample range and explore applications in different objects and scenarios such as plant mesophyll and vein segmentation, and segmentation of different types of cells in fluorescence images, in order to prove that the fusion of different fluorescence images has a wide range of applications and is worthy of in-depth research.