Transplanting seedling survival state monitoring method and system based on image analysis

Abstract

The application discloses a transplanted seedling survival state monitoring method and system based on image analysis, and belongs to the technical field of computer vision and intelligent forestry monitoring, and the method comprises the following steps: collecting a transplanted seedling canopy image and obtaining a canopy region image through illumination correction and background separation; performing leaf instance segmentation through a leaf segmentation network; calculating a leaf greenness index, a yellowing proportion and a browning area proportion in a multi-color space to obtain a color stress score; extracting a leaf curling degree, a wilting drooping angle and a leaf area change rate to obtain a morphological stress comprehensive degree; fusing color and morphological features based on a Bayesian inference framework to output a survival probability score and a health grade classification, and combining a time sequence recovery trend to dynamically correct.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary technical field of computer vision and intelligent forestry monitoring, specifically relating to a method and system for monitoring the survival status of transplanted seedlings based on image analysis. Background Technology

[0002] Tree transplantation plays an irreplaceable role in urban greening, ecological restoration, and landscape creation. However, during the adaptation process from their native environment to their new planting environment, transplanted seedlings often face the combined effects of multiple adverse factors, such as root damage, water stress, and nutrient supply interruption, leading to a high mortality risk in the early stages of transplantation. Assessing the survival status of transplanted seedlings has always been a core aspect of forestry project management, as the assessment results directly affect the project's acceptance conclusion and the formulation of subsequent maintenance strategies. Traditional methods for determining seedling survival rates mainly rely on manual on-site inspections. Technicians use the naked eye to observe external characteristics such as leaf color, branch elasticity, and new bud sprouting to comprehensively judge whether seedlings have survived. However, this manual inspection method has inherent drawbacks such as long inspection cycles, strong subjectivity, and low efficiency. Especially in large-scale transplanted forest scenarios, inspecting each seedling individually consumes a significant amount of manpower and time, and subtle signs of stress, such as slight yellowing of leaves or slight curling of leaf margins, are easily overlooked. By the time symptoms become apparent, the optimal window for maintenance intervention has often passed.

[0003] With the development of computer vision and deep learning technologies, image-based plant health monitoring methods have gradually become a research hotspot. Existing technologies include numerous schemes for plant disease detection, such as using convolutional neural networks to classify and identify leaf lesions, or analyzing the spectral reflectance characteristics of leaves through hyperspectral imaging to detect disease infection early. However, these schemes mainly focus on identifying disease symptoms caused by specific pathogens, and their technical approach revolves around classifying and training based on the morphological, texture, and color features of lesions, limiting their applicability to matching and identifying known disease types. Monitoring the survival status of transplanted seedlings differs fundamentally from disease detection: transplant stress is not caused by specific pathogens, but rather by a systemic physiological decline process resulting from root damage. This manifests as progressive chlorosis of leaves, gradual wilting and curling of leaves, and a comprehensive decline in overall canopy vitality—all comprehensive degenerative characteristics. The gradual and comprehensive nature of these characteristics makes it difficult to obtain accurate survival status results by directly applying disease detection models.

[0004] Chinese patent CN108492365A discloses a color-grading-based adaptive mapping visualization simulation method for leaves. This method enhances leaf texture images through color enhancement, color space conversion, and color weight calculation, adaptively attaching leaf textures with different color weights to corresponding branches to achieve a highly realistic visualization effect for 3D forest models. This method provides a leaf classification approach based on color weights, offering a reference for quantifying leaf color features. However, the core purpose of this method is 3D visualization modeling, not monitoring and evaluating the physiological health status of real seedlings. Specifically, its color weight calculation is based solely on a simple weighted ratio of RGB three-channel pixel values, without involving multi-color space feature analysis closely related to plant physiological states; its leaf classification criteria only distinguish the color differences between sun-facing and shaded sides, without establishing a quantitative evaluation system related to stress levels; furthermore, this method completely lacks leaf morphology feature analysis capabilities, failing to capture key morphological signs common in transplanted seedlings such as leaf curling, wilting, and drooping, and lacks the ability for multi-time-point dynamic tracking and survival probability prediction.

[0005] Therefore, there is an urgent need in existing technologies for an intelligent method specifically designed for monitoring the survival status of transplanted seedlings. This method needs to be able to simultaneously analyze leaf color and morphological characteristics, establish a quantitative assessment index system correlated with the degree of transplanting stress, and track the recovery trend of seedlings through longitudinal comparisons at multiple time points. This would provide objective, efficient, and accurate technical support for the maintenance management and survival rate acceptance of transplanted seedlings. Specifically, in terms of color feature analysis, it is necessary to overcome the limitations of a single color space and establish quantitative indicators closely related to the physiological degradation process of plants in multiple color spaces. In terms of morphological feature analysis, it is necessary to introduce quantitative descriptive indicators of leaf geometric deformation to capture early signals of water stress. In terms of assessment and decision-making, it is necessary to construct an intelligent assessment model that can integrate multidimensional heterogeneous features and take into account dynamic changes over time, thereby realizing a technological shift from post-hoc judgment to process monitoring of the survival status of transplanted seedlings. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for monitoring the survival status of transplanted seedlings based on image analysis, so as to solve the above-mentioned technical problems.

[0007] To achieve the above objectives, this invention provides a method for monitoring the survival status of transplanted seedlings based on image analysis, the method comprising the following steps:

[0008] Step S1, Image Acquisition and Preprocessing: Periodically acquire images of the canopy of transplanted seedlings using an image acquisition device. Perform illumination correction processing on the acquired canopy images to eliminate the interference of ambient light changes on image colors. Perform background separation processing on the illumination-corrected images to extract the canopy area image. The illumination correction processing includes white balance correction based on the gray world hypothesis and color calibration based on a reference color chart.

[0009] Step S2, Leaf Instance Segmentation Step: Input the canopy region image obtained in Step S1 into the leaf segmentation network, perform instance segmentation processing on each leaf in the canopy region image, and output the leaf mask and leaf position coordinates corresponding to each leaf. The leaf segmentation network is a deep convolutional neural network based on an encoder-decoder architecture, and combines an attention mechanism to enhance the leaf edge features.

[0010] Step S3, color feature analysis step: Convert the leaf regions corresponding to each leaf mask obtained in step S2 to multiple color spaces respectively. Calculate the green retention index, yellowing ratio and browning area ratio of the leaves in each color space, and obtain the comprehensive score of leaf color stress based on the weighted fusion of each color index. Among them, the multiple color spaces include HSV color space and CIELAB color space.

[0011] Step S4, Morphological Feature Analysis: Based on the leaf masks obtained in Step S2, extract the degree of leaf curling, wilting and drooping angle, and leaf area change rate. Then, perform nonlinear fusion of the degree of leaf curling, wilting and drooping angle, and leaf area change rate to obtain the comprehensive morphological stress degree. The leaf area change rate is obtained by comparing the leaf area at the current collection time point with that at the previous collection time point.

[0012] Step S5, survival status assessment step: Input the leaf color stress comprehensive score obtained in step S3 and the morphological stress comprehensive score obtained in step S4 into the survival probability assessment model, and output the survival probability score and health level classification of the transplanted seedlings. The survival probability assessment model is based on the Bayesian posterior inference framework, which integrates color features and morphological features, and dynamically corrects them by combining the temporal recovery trend of multi-time point images. The health level classification result is fed back to step S2 to adaptively adjust the segmentation sensitivity parameters of the leaf segmentation network.

[0013] This invention also provides an image analysis-based system for monitoring the survival status of transplanted seedlings, comprising: an image acquisition and preprocessing module for periodically acquiring images of the canopy of transplanted seedlings using an image acquisition device, performing illumination correction and background separation processing on the acquired canopy images to extract canopy region images; a leaf instance segmentation module for segmenting leaf instances in the canopy region images, outputting the leaf mask and leaf position coordinates corresponding to each leaf; a color feature analysis module for calculating leaf color indices in multiple color spaces and obtaining a comprehensive score of leaf color stress; a morphological feature analysis module for extracting leaf morphological indices and obtaining a comprehensive degree of morphological stress; and a survival status assessment module for fusing color features and morphological features based on a Bayesian posterior inference framework to output a survival probability score and health level classification, and feeding the results back to the leaf instance segmentation module.

[0014] The beneficial effects of this invention are as follows: First, by analyzing the leaf green retention index, yellowing ratio, and browning area proportion through multi-color space fusion, a refined quantitative assessment of the color stress state of transplanted seedlings is achieved. This overcomes the limitations of existing technologies that rely solely on a single color space, making it difficult to accurately distinguish different stress stages. In particular, the green retention index defined in the CIELAB color space exhibits excellent sensitivity to gradual changes in chlorophyll content. Second, by introducing morphological features such as leaf curling degree, wilting and drooping angle, and leaf area change rate, and employing a nonlinear fusion strategy based on a combination of Sigmoid mapping and weighted geometric mean, the lack of morphological analysis dimensions in existing technologies is compensated for. This is especially true for early wilting symptoms caused by water stress, and the nonlinear fusion method can produce a synergistic amplification effect when multiple morphological indicators simultaneously show stress signals. Third, the survival probability assessment model based on a Bayesian posterior inference framework simultaneously integrates information from both color and morphology dimensions, using a Beta distribution for likelihood probability parameterization modeling. Compared to single-dimensional judgment methods, this approach yields more accurate and robust assessment results. Finally, the multi-time-point longitudinal tracking and time-series recovery trend dynamic correction mechanism enables the system to distinguish between temporary stress response and irreversible death process, significantly reducing the misjudgment rate of survival determination. Furthermore, the system adjusts the leaf segmentation sensitivity parameter and color index weight coefficient in reverse through the health level classification results, forming a complete closed-loop collaborative optimization mechanism from front-end data acquisition to back-end decision-making and then to front-end parameter optimization. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the image analysis-based method for monitoring the survival status of transplanted seedlings in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of the architecture of the transplanted seedling survival status monitoring system based on image analysis in an embodiment of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are only used to illustrate the technical solution of the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0018] like Figure 1 As shown in the figure, this invention provides a method for monitoring the survival status of transplanted seedlings based on image analysis. This method achieves intelligent, refined, and dynamic monitoring of the survival status of transplanted seedlings through deep coupling and closed-loop collaboration of five core steps: image acquisition and preprocessing, leaf instance segmentation, color feature analysis, morphological feature analysis, and survival status assessment. In one embodiment of this invention, the entire method is driven by image data, with the output of the previous step serving as the key input for the next step. Simultaneously, the evaluation results of subsequent steps can inversely influence the processing parameters of the preceding steps, thus forming an adaptive optimization closed-loop system. Each step is described in detail below.

[0019] Step S1, Image Acquisition and Preprocessing. The main purpose of this step is to acquire high-quality images of the transplanted seedling canopy and convert the original acquired images into standardized canopy area images suitable for subsequent feature analysis through illumination correction and background separation. Preferably, the image acquisition device can be configured in two ways: one is a mobile acquisition method, in which inspectors use a handheld mobile terminal equipped with a high-resolution camera to collect images of the seedling canopy one by one or row by row along a preset route in the transplanting area. In one embodiment of the present invention, the resolution of the mobile terminal camera is not less than 12 million pixels, the acquisition distance is controlled between 0.5m and 2.0m, and the acquisition angle is 30° to 60° from above to ensure effective coverage of the canopy area; the other is a fixed monitoring method, in which fixed network monitoring cameras are deployed in the transplanting area to automatically collect images of the seedling canopy at preset time intervals. Preferably, the acquisition interval of the fixed monitoring cameras is 1 to 3 times per day, and the acquisition time is selected between 9:00 and 11:00 am when the lighting conditions are relatively stable to reduce the impact of changes in the sun's altitude angle on the image color.

[0020] After acquiring the original canopy image, illumination correction is performed first. Since transplanted seedlings are typically in outdoor natural light environments, images acquired at different times and under different weather conditions exhibit significant differences in brightness and color temperature. Without correction, this will severely affect the accuracy and consistency of subsequent color feature analysis. In one embodiment of this invention, illumination correction includes two cascaded sub-steps: the first sub-step is white balance correction based on the gray-world assumption. This method assumes that the average value of all colors in a natural scene tends towards gray, and calculates the average value of the RGB three channels accordingly, adjusting the channel gain. Specifically, for each pixel in the original image... Corrected pixel values Calculate as follows: ,in, , , These are the average values ​​of all pixels in the R, G, and B channels of the original image, ranging from 0 to 255, and are dimensionless. It is the arithmetic mean of the three channel values, i.e. Dimensionless. This white balance correction can eliminate the overall color cast caused by changes in the color temperature of the light source, ensuring that images acquired at different times have a basically consistent color reference.

[0021] The second sub-step is color calibration based on a reference color chart. Preferably, during each image acquisition, a reference color chart containing at least 18 standard color patches is placed near the seedling canopy, where each color patch on the standard color chart has a known standard color value. After image acquisition, the color chart area is automatically identified, and the actual acquired color value of each color patch in the image is extracted. Compare it with the corresponding standard color value For comparison, among them Number the color blocks. Fitting a value using the least squares method Color correction matrix This minimizes the mean square error between the corrected color values ​​and the standard color values. The color-calibrated image maintains highly consistent color response characteristics under different environmental conditions, which is crucial for subsequent cross-time point color feature comparison and analysis.

[0022] After illumination correction, background separation is performed to extract the canopy region of the seedlings. In one embodiment of the present invention, background separation employs a dual-channel strategy that combines a vegetation segmentation method based on the supergreen index with deep learning semantic segmentation. Specifically, the first channel is a fast segmentation channel based on the supergreen index, which calculates the supergreen index for the illumination-corrected image. : ,in, , , These are the normalized red, green, and blue channel values, ranging from 0 to 1, and are dimensionless. Super Green Index The value ranges from -1 to 1, with a larger value indicating more prominent green vegetation features. Preferably, a segmentation threshold is set. In one embodiment of the present invention The value is 0.1, when a certain pixel's Value greater than Pixels are classified as vegetation area pixels if they are not in the vegetation area, otherwise as background pixels. After segmentation, morphological closing operations are used to fill small holes inside the vegetation area. The closing operation uses a circular kernel with a radius of 5 pixels as the structuring element. Connected component analysis is then used to remove noise areas smaller than a preset minimum area threshold (preferably 500 pixels), ultimately obtaining a complete canopy area image. It is worth noting that when transplanted seedlings suffer severe browning or dieback, their leaf color may be close to the background soil color. In this case, relying solely on the supergreen index may miss some withered leaves. To address this, this invention further combines deep learning semantic segmentation as a second auxiliary method. Based on the supergreen index segmentation results, regions with low green index but still belonging to the seedling canopy are extracted. The segmentation results of the two channels are merged through a logical OR operation to ensure that withered leaves can also be included in subsequent analysis. Preferably, the deep learning semantic segmentation model uses the lightweight MobileNetV3 as the backbone network to ensure that the inference speed on edge computing nodes meets the real-time processing requirements.

[0023] Step S2, Leaf Instance Segmentation. This step receives the canopy region image output from Step S1 as input, and segments each leaf in the canopy one by one using a leaf segmentation network, outputting the corresponding binary mask and centroid coordinates for each leaf. Unlike semantic segmentation, which only distinguishes between foreground and background, instance segmentation needs to distinguish overlapping and occluded leaves in the same canopy as independent individuals. This is a necessary prerequisite for subsequent refined color and morphological analysis of individual leaves.

[0024] In one embodiment of the present invention, the leaf segmentation network is implemented using a deep convolutional neural network based on an encoder-decoder architecture. The encoder part uses ResNet-50 as the backbone network, extracting multi-scale feature maps through successive convolutional and pooling layers, wherein the convolutional kernel sizes are sequentially [sizes to be filled in]. , The pooling layer step size is 2, thus obtaining [data at] the four scale levels respectively. , , , The feature map has a high resolution. The decoder recovers the spatial resolution through layer-by-layer bilinear interpolation upsampling, and fuses the feature maps of the corresponding encoder levels at each upsampling level through skip connections, so that high-level semantic information and low-level spatial detail information can be effectively integrated.

[0025] Preferably, an attention mechanism is introduced between the last level of the encoder and each upsampling level of the decoder to enhance the expressive power of blade edge features. In one embodiment of the invention, the attention mechanism employs a cascaded combination of channel attention (SE module) and spatial attention (CBAM spatial branch). The channel attention module first performs global average pooling and global max pooling operations on the feature map, inputs the pooling results into two fully connected layers (with an intermediate dimension compression ratio of 16), and then generates a channel attention weight vector through a Sigmoid activation function. This weight vector is multiplied channel-by-channel with the original feature map to highlight important channel features related to the blade. Subsequently, the spatial attention module takes the average and maximum values ​​along the channel dimension of the channel-weighted feature map, concatenates the two results, and then... The convolutional kernel performs convolution and generates a spatial attention map through a sigmoid activation function. This spatial attention map is multiplied pixel-by-pixel with the input feature map to enhance the feature responses of leaf edges and contour regions. Through this concatenated attention mechanism, the leaf segmentation network can more accurately locate leaf boundaries when faced with complex scenes where leaves occlude each other and leaves intersect with branches.

[0026] At the network output, the leaf segmentation network employs a multi-task output head design. Preferably, the network simultaneously outputs a semantic segmentation branch and an instance center offset branch: the semantic segmentation branch predicts the probability that each pixel belongs to the leaf foreground or background, and the instance center offset branch predicts the two-dimensional offset vector from each foreground pixel to the center of its corresponding leaf instance. In the post-processing stage, the leaf foreground region is first obtained based on the semantic segmentation results. Then, through cluster analysis of the center offset vectors (mean-shift clustering is used in one embodiment of this invention, with the bandwidth parameter set to 15 pixels), the foreground pixels are assigned to each leaf instance, and finally, a binary mask for each leaf is output. and the corresponding centroid position coordinates ,in Number the blades.

[0027] It is worth emphasizing that the segmentation sensitivity parameters of the leaf segmentation network in this invention have adaptive adjustment capabilities. Specifically, the health level classification result output in step S5 is fed back to this step. When the seedling is assessed as moderate or severe stress, it indicates that the seedling leaves may have severely withered or fallen off. At this time, the contrast between the leaves and the background decreases, increasing the segmentation difficulty. In one embodiment of this invention, after receiving the moderate or severe stress level signal, the system automatically lowers the foreground determination probability threshold of the semantic segmentation branch from the default 0.5 to 0.3, and expands the bandwidth parameter of the instance center offset clustering from 15 pixels to 25 pixels, thereby improving the detection sensitivity of degraded leaves. This closed-loop feedback mechanism ensures that reliable leaf segmentation results can be obtained at different stress stages of the seedling.

[0028] In one embodiment of the present invention, the leaf segmentation network is trained using a transplanted seedling leaf segmentation dataset containing 5000 labeled images, with pixel-level instance annotation for each leaf. Data augmentation strategies are employed during training, including random horizontal flipping and random rotation (within a certain angle range). to Random brightness adjustment (ranging from 0.7 to 1.3 times the original brightness) and random cropping were used. Training employed the Adam optimizer with an initial learning rate of [missing information]. The learning rate decreases to 0.5 times its original value every 30 training rounds, with a total of 120 training rounds.

[0029] Step S3, Color Feature Analysis. This step receives the blade masks output from step S2. For each leaf, a refined color feature extraction and quantitative analysis was performed. Color features are one of the most intuitive indicators reflecting the physiological state of transplanted seedlings. Leaf chlorosis, yellowing, and browning correspond to different degrees of stress response. Therefore, establishing a scientific and reasonable color quantitative index system is of great significance for accurately assessing the survival status of seedlings.

[0030] In one embodiment of the present invention, each blade is first masked. The corresponding leaf region pixels were converted from the RGB color space to the HSV and CIELAB color spaces, respectively. These two color spaces were chosen because: the HSV color space decouples color information into three independent channels: hue (H), saturation (S), and lightness (V). The hue channel directly reflects the transition of leaf color from green to yellow and brown, and its sensitivity to changes in light intensity is lower than that of the RGB color space, making it suitable for leaf color analysis under outdoor natural light conditions. The CIELAB color space exhibits perceptual uniformity, with the a channel representing the change from green to red and the b channel representing the change from blue to yellow. The combination of these two spaces can precisely distinguish the color differences between healthy green leaves and leaves subjected to varying degrees of stress. The RGB to HSV conversion was implemented using a standard algorithm, while the RGB to CIELAB conversion was completed via the XYZ intermediate color space. The XYZ to LAB conversion used a D65 standard light source as the white point reference. By simultaneously extracting features from these two complementary color spaces, the leaf color changes caused by transplant stress can be comprehensively captured from different dimensions of color perception, exhibiting higher robustness and discriminative power compared to analysis in only a single color space.

[0031] Based on the HSV and CIELAB dual color spaces, this invention defines and calculates the following three color indices:

[0032] First, the Green Retention Index (LGRI). This index quantifies the degree to which leaves retain their green color after transplanting and is an indirect indicator reflecting chlorophyll content and photosynthetic activity. In one embodiment of the present invention, the formula for calculating the Green Retention Index is: ,in, For the first The green retention index of a leaf has a value ranging from 0 to 1 and is dimensionless. The closer the value is to 1, the better the green retention of the leaf. For the first blade mask The total number of pixels contained; A single pixel within the leaf mask; For this pixel in the CIELAB color space Channel value, ranging from -128 to 127, dimensionless, negative value represents green direction, positive value represents red direction; For this pixel Channel value, ranging from -128 to 127, dimensionless, positive value represents the yellow direction; To prevent small constants with denominators of zero, it is preferable to take a value of [value missing]. When the leaves are a healthy green, The value is relatively negative, and the calculation result approaches 1; when the leaves turn yellow or brown, The calculated result decreases as the value approaches 0 or becomes positive. Compared to the traditional Normalized Difference Vegetation Index (NDVI), this index is more suitable for the analysis of near-field visible light images and has higher sensitivity to gradual changes in leaf color.

[0033] Second, the proportion of yellowing ( This indicator is used to quantify the area of ​​leaf regions exhibiting yellowing symptoms. In one embodiment of the present invention, the yellowing hue range is defined in the HSV color space as follows: saturation range is The brightness range is Pixels that meet the above three conditions are identified as yellowed pixels. The formula for calculating the yellowing ratio is: ,in, For the first The yellowing rate of each leaf is dimensionless, ranging from 0 to 1. For the first The number of pixels in the leaf blade that meet the yellowing criteria; This represents the total number of pixels on the leaf blade. In one embodiment of the present invention, when... At that time, it was believed that the leaf showed significant yellowing symptoms.

[0034] Third, the percentage of browning area ( This indicator is used to quantify the percentage of leaf area showing browning or scorching symptoms. In one embodiment of the present invention, the browning determination criterion is defined in the CIELAB color space as: brightness value. (Values ​​range from 0 to 100, dimensionless) and chromaticity values (in (Values ​​range from 0 to 180, dimensionless). The formula for calculating the percentage of browning area is: ,in, For the first The percentage of browned area on each leaf, ranging from 0 to 1, is dimensionless; The number of pixels required to meet the browning criteria.

[0035] After obtaining the above three color indicators, the comprehensive score of leaf color stress is calculated by weighted fusion. In one embodiment of the present invention, the fusion formula is: ,in, For the first The color stress score of each leaf is dimensionless and ranges from 0 to 1. The larger the value, the more severe the color stress. , , These are the weighting coefficients for the green retention index, yellowing rate, and browning area percentage, respectively, satisfying... Preferably, , , This weighting reflects the decisive role of green retention capacity in survival assessment, while also taking into account the auxiliary reference value of yellowing and browning indicators. For the comprehensive score of canopy color stress for a single seedling, all leaves of that seedling are used. The area-weighted average was used as the overall color stress score for the seedlings. .

[0036] Step S4, morphological feature analysis step. This step also receives the blade masks output from step S2. As input, leaf deformation features closely related to transplanting stress are extracted from the morphological geometry dimension. Unlike color features, which focus on reflecting chlorophyll content and photosynthetic system status, morphological features mainly reflect leaf water status and changes in cell turgor pressure, and have higher detection sensitivity for early wilting symptoms caused by water stress in the early stages of transplanting. In one embodiment of the present invention, the following three morphological indicators are extracted: First, the degree of leaf curling ( When transplanted seedlings are in a water deficit, their leaves curl inward to reduce transpiration area; the greater the curling, the more severe the water stress. In one embodiment of the present invention, the degree of leaf curling is characterized by the ratio between the actual area of ​​the leaf eccentricity and its minimum circumscribed hull area. ,in, For the first The degree of leaf curling ranges from 0 to 1, is dimensionless, and the larger the value, the more severe the leaf curling. For blade masking The actual pixel area, in pixels. ; This represents the area of ​​the minimum circumscribed convex hull of the blade mask, in pixels. When the blade is fully extended and its edges are smooth, the area of ​​the blade's masking surface is close to its convex hull area. Approaching 0; when the blade curls, causing the edges to become concave or the area to shrink, The value increases. Preferably, the leaf mask is first smoothed by 5 pixels of Gaussian before calculating the convex hull to eliminate the influence of edge noise on the convex hull calculation. In one embodiment of the present invention, the degree of curling is divided into three intervals: This indicates the normal unfolded state. Indicates slight curling. It indicates obvious curling.

[0037] Second, the angle of wilting and drooping ( Transplanted seedlings exhibit noticeable drooping of leaves after losing turgor pressure; the greater the angle at which the midrib deviates from the horizontal plane, the more severe the wilting. In one embodiment of this invention, the wilting and drooping angle is obtained through the following steps: first, the leaves are masked. A morphological skeleton extraction operation is performed to obtain the skeletal lines of the leaf; then, straight line segment fitting (preferably using the least squares method) is performed on the skeletal lines to obtain the direction vector of the leaf midrib. Finally, calculate the direction vector and the horizontal direction vector of the image. The angle between them is used as the angle of wilting and drooping: ,in, For the first The drooping angle of a leaf blade ranges from 0° to 90°, and the unit is degrees (°). This is the dot product of two direction vectors. For canopy images acquired from a top-down view, the midrib of a healthy, unfolded leaf is roughly horizontal. The value is relatively small; while wilted and drooping leaves appear smaller and more elongated in top-view images, with their skeletal structure deviating from the horizontal plane. The value increases. In one embodiment of the present invention, when... The birds were then judged to be in a severely wilted state. It should be noted that for images captured by fixed monitoring cameras, due to the fixed shooting angle, adjustments need to be made based on the camera's installation angle. Perform geometric corrections to ensure that the wilting angles obtained by cameras at different installation positions are comparable.

[0038] Third, the rate of change in leaf area ( This indicator reflects the dynamic degradation or recovery trend of leaves by comparing the change in leaf area at the current collection time point with that at the previous collection time point. In one embodiment of the present invention, the formula for calculating the leaf area change rate is: ,

[0039] in, For the first The rate of change of the area of ​​each leaf blade ranges from -1 to positive infinity and is dimensionless. and These are the current data collection times. and the previous collection time point The area of ​​the leaf mask, in pixels. ; To prevent small constants with denominators of zero, the value is taken as... Positive values ​​indicate an increase in leaf area (new leaves unfolding or re-expanding), while negative values ​​indicate a decrease in leaf area (curling or shedding). To compare leaf areas across time points, the matching algorithm for leaf position coordinates in step S2 is needed to associate instances of the same leaf at different time points. Preferably, this invention uses a method based on nearest neighbor matching of position coordinates and joint judgment of leaf shape similarity for cross-time point leaf matching, wherein the position deviation threshold is set to 30 pixels and the shape similarity threshold (using Hu moment distance metric) is set to 0.15.

[0040] After obtaining the above three morphological indices, the morphological stress comprehensiveness is calculated through nonlinear fusion. This invention employs a nonlinear fusion strategy combining Sigmoid mapping and weighted geometric mean. Specifically, each morphological indicator is first mapped to a unified dimension of 0 to 1 using the Sigmoid function: ,in, This is the standard Sigmoid function; , , The mapping slope parameters for each index control the steepness of the Sigmoid curve. Preferably, the values ​​are 10, 8, and 12, which are dimensionless. , , Here, represents the center offset parameter for each indicator, corresponding to the critical stress threshold for each indicator. Preferably, these values ​​are 0.15, 0.3, and 0.1, respectively, and are dimensionless. Note In the calculation of We take the negative sign because a decrease in area (negative rate of change) corresponds to increased stress.

[0041] Then, the data is fused using a weighted geometric mean: ,in, For the first The morphological stress comprehensiveness of a leaf blade ranges from 0 to 1 and is dimensionless. The larger the value, the more severe the morphological stress. , , The fusion weight index for each indicator satisfies Preferably , , The key significance of using a weighted geometric mean instead of an arithmetic mean is that when the mapping value of any morphological indicator approaches 0 (i.e., in a normal state), the overall fusion result will be significantly lowered, thus avoiding the situation where an abnormally high value of a single indicator dominates the overall score. Conversely, when multiple indicators simultaneously show stress signals, the comprehensive effect produced by the geometric mean will be greater than the linear superposition of a single indicator, reflecting a synergistic effect. For the comprehensive stress degree of canopy morphology of a single seedling, all leaves are also taken. The area-weighted average value was used as the overall morphological stress score for the seedlings. .

[0042] Step S5, Survival Status Assessment. This step is the final decision-making stage of the entire monitoring method, receiving the overall color stress score of the seedlings output from Step S3. The overall morphological stress score of the seedlings output in step S4 As input, the survival probability assessment model outputs a survival probability score and health level classification for the transplanted seedling. In one embodiment of the present invention, the survival probability assessment model performs multi-source information fusion and decision-making based on a Bayesian posterior inference framework.

[0043] Specifically, color stress score and morphological stress score Considered as the actual survival status of the given seedlings ( Indicates survival. Given two observations (representing death), calculate the posterior probability of seedling survival using Bayes' theorem: ,in, The posterior probability of seedling survival under given color and morphology observations, i.e., the survival probability score, ranges from 0 to 1 and is dimensionless. The likelihood probability of observing the current color and morphological features in a living state; The prior probability of survival is set to 0.7 in one embodiment of the present invention based on historical statistical data; This is the normalization constant.

[0044] Preferably, assuming that color features and morphological features are conditionally independent under a given survival state, the likelihood probability is decomposed as follows: ,in, and Parametric modeling was performed using the Beta distribution. In one embodiment of the invention, the shape parameters of the Beta distribution were obtained by fitting parameters to a training set containing actual survival data of 2000 transplanted seedlings: for the color stress score distribution under survival conditions, the shape parameters... , Dimensionless (indicating that the color stress scores of surviving seedlings are concentrated in the lower value range); for the color stress score distribution in the dead state, the shape parameter... , (This indicates that the color stress scores of dead seedlings are concentrated in the higher value range.) Similarly, for the distribution of the overall morphological stress in the surviving state, the shape parameter... , For the distribution of morphological stress comprehensiveness in the death state, shape parameters , All the above shape parameters are dimensionless. The technical advantage of Beta distribution parameterized modeling is that, through flexible two-parameter shape control, it can accurately fit the distribution differences between surviving and dead seedlings in the color and morphological feature space of actual observation data, avoiding the fitting bias that may be introduced by the Gaussian distribution assumption. Moreover, the domain of the Beta distribution is naturally limited to the interval between 0 and 1, which perfectly matches the normalization index value range of this invention.

[0045] After obtaining the initial survival probability score at the current time point, this invention further introduces a time-series recovery trend dynamic correction mechanism. The core idea of ​​this mechanism is that the survival determination of transplanted seedlings should not be based solely on a snapshot at a single time point, but should also consider the trend of seedling status changes over time. If a seedling's current color and morphology indicators are poor, but it shows a continuous improvement trend recently, its true survival probability should be higher than the result calculated based solely on the current indicators; conversely, if the seedling indicators continue to deteriorate, its true survival probability should be lower than what the current indicators indicate.

[0046] In one embodiment of the present invention, the current collection time point is obtained. Previous consecutive Survival probability score sequence corresponding to each collection time point ,in The time window length is preferably... The recovery trend value was calculated using the exponentially weighted moving average method for this sequence. : ,in, This represents the recovery trend value at the current time point, ranging from 0 to 1, and is dimensionless. This is the attenuation coefficient, ranging from 0 to 1, dimensionless; a larger value indicates slower attenuation of historical data. Preferably, the attenuation coefficient... Based on the number of days after transplantation Perform adaptive adjustments: ,in, The initial attenuation coefficient is preferably 0.3, which is dimensionless. The stabilization period decay coefficient is preferably taken as 0.8, which is dimensionless. The number of days after transplantation is in days (d). The maximum number of days for attenuation adjustment is preferably 60 days. The rationale for this adaptive adjustment strategy is that the seedling condition fluctuates drastically in the early stages of transplantation, and therefore more reliance should be placed on recent data (which is less recent). As time goes by after transplanting, the condition of the seedlings tends to stabilize, and historical trends can be used as a reference (for larger trends). ).

[0047] Final revised survival probability score for: ,

[0048] in, The fusion weight for the score at the current time point is preferably 0.6, which is dimensionless. (Corrected) This is the final survival probability score of the transplanted seedling.

[0049] Based on the survival rate score, the seedlings are divided into the following four health levels: The corresponding health recovery level This corresponds to a mild stress level. This corresponds to a moderate level of stress. This corresponds to a severe stress level. Among them, , , The preset threshold for classifying levels is preferably... , , Dimensionless. This four-level classification system corresponds to common forestry seedling maintenance and management levels. Seedlings in the healthy recovery level can be maintained using routine maintenance strategies. Seedlings in the mild stress level are advised to increase watering frequency. Seedlings in the moderate stress level require comprehensive intervention measures including shading and watering. Seedlings in the severe stress level need to be assessed to determine whether replanting or replacement is necessary.

[0050] As mentioned earlier, the health level classification results will be fed back to the leaf segmentation network in step S2, forming a closed-loop collaborative optimization mechanism. This feedback mechanism is reflected not only in the adaptive adjustment of the segmentation sensitivity parameter, but also in the dynamic adjustment of the color index weight: when seedlings are assessed as moderate or severe stress, the system automatically increases the weight of the browning area proportion. Reduce the index weight to 0.35 and correspondingly decrease the green index weight. The value was reduced to 0.40 to focus more on irreversible damage indicators in severe degradation scenarios, thereby further improving the accuracy and robustness of the assessment.

[0051] like Figure 2 As shown, this embodiment of the invention also provides an image analysis-based seedling survival status monitoring system. This system corresponds one-to-one with the method steps in the above method embodiments, including an image acquisition and preprocessing module, a leaf instance segmentation module, a color feature analysis module, a morphological feature analysis module, and a survival status assessment module. The modules are connected through a data bus to form a closed-loop data flow architecture that combines feedforward and feedback.

[0052] The image acquisition and preprocessing module includes an image acquisition subunit and an image preprocessing subunit. The image acquisition subunit is responsible for managing and scheduling image acquisition devices, supporting flexible switching between mobile acquisition mode and fixed monitoring mode. In one embodiment of the present invention, the image acquisition subunit maintains a device registry, recording parameter information such as device identifier, installation location, shooting angle, and acquisition schedule for each acquisition device. The image preprocessing subunit performs illumination correction and background separation processing in step S1 of the above method embodiment, specifically including three processing components: a white balance correction engine, a color card recognition and calibration engine, and a vegetation segmentation engine. The white balance correction engine automatically calculates channel gain and performs white balance correction based on the gray-scale world assumption; the color card recognition and calibration engine automatically locates the standard color card region in the image through template matching, extracts the acquisition value of each color block, and fits a color correction matrix; the vegetation segmentation engine extracts the canopy region through supergreen index calculation combined with deep learning semantic segmentation. The preprocessed canopy region image and corresponding metadata (including acquisition timestamp, device identifier, and GPS location coordinates) are transmitted together to the leaf instance segmentation module.

[0053] The blade instance segmentation module carries the blade segmentation network and post-processing logic in step S2 of the above method embodiment. This module includes two core components: a network inference engine and an instance post-processing engine. The network inference engine loads the pre-trained blade segmentation network model weight file and performs forward inference calculations, outputting a semantic segmentation probability map and an instance center offset vector map. The instance post-processing engine generates a binary mask and centroid coordinates for each blade instance based on the semantic segmentation probability map and the offset vector map through mean-shift clustering. Preferably, this module also includes a segmentation sensitivity adaptive adjustment interface, used to receive the health level signal fed back by the survival status assessment module and dynamically adjust the foreground determination probability threshold and clustering bandwidth parameters accordingly.

[0054] The color feature analysis module carries the color space conversion and color index calculation logic in step S3 of the above method embodiment. This module receives leaf masks output by the leaf instance segmentation module and sequentially performs HSV conversion, CIELAB conversion, green retention index calculation, yellowing ratio calculation, browning area ratio calculation, and weighted fusion of color stress comprehensive scores. In one embodiment of the present invention, this module internally maintains a color parameter configuration table, storing configurable parameters such as yellowing hue range, browning brightness threshold, browning chromaticity threshold, and weight coefficients for each index, supporting personalized parameter settings based on the color characteristics of different seedling species.

[0055] The morphological feature analysis module carries out the morphological index extraction and nonlinear fusion logic in step S4 of the above method embodiment. This module receives the leaf masks and position coordinates output by the leaf instance segmentation module, and sequentially performs processing steps such as convex hull calculation, skeleton extraction, midrib direction fitting, cross-time point leaf matching, and area change rate calculation. It then outputs the morphological stress comprehensive degree through fusion of Sigmoid mapping and weighted geometric mean. Preferably, this module includes a leaf tracking database to store leaf instance information of each seedling at multiple consecutive collection time points, providing a data foundation for cross-time point leaf matching and area change rate calculation.

[0056] The survival status assessment module carries the Bayesian posterior inference and temporal recovery trend correction logic in step S5 of the above method embodiment. This module receives the color stress comprehensive score output by the color feature analysis module and the morphological stress comprehensive degree output by the morphological feature analysis module. Through Bayesian posterior probability calculation and dynamic correction of the temporal recovery trend, it outputs the final survival probability score and health level classification. In one embodiment of the invention, this module internally maintains a temporal score database to store the survival probability scores of each seedling at multiple consecutive collection time points, providing historical data support for the calculation of the exponentially weighted moving average trend. This module is also responsible for sending health level feedback signals to the leaf instance segmentation module and weight dynamic adjustment signals to the color feature analysis module, forming a system-level closed-loop collaborative optimization mechanism. Furthermore, the output of the survival status assessment module also provides seedling survival probability scores, health level classifications, and recovery trend reports to external systems through a standard data interface to support transplanted seedling maintenance management decisions and survival rate acceptance statistics.

[0057] Preferably, the system of the present invention also includes a data management and visualization platform, which runs on the server side and displays the monitoring results to users through a web interface or mobile terminal application. The data management and visualization platform includes two core components: a spatial data management engine and a report generation engine. The spatial data management engine, based on a geographic information system framework, associates the GPS location coordinates of each transplanted seedling with its corresponding survival probability score and health level information, marking the health status of each seedling on an electronic map with different colors: green represents the health recovery level, yellow represents the mild stress level, orange represents the moderate stress level, and red represents the severe stress level. The report generation engine can automatically summarize and statistically analyze the overall survival rate of transplanted seedlings, the distribution of seedlings of each level, and the trend curve of survival rate changes over time according to the user-specified time range and spatial area, and output electronic reports conforming to the forestry engineering acceptance standard format.

[0058] In one embodiment of the present invention, the overall data flow architecture of the system is as follows: the image acquisition device uploads the original canopy image to the edge computing node via a wireless network. The edge computing node performs computationally intensive tasks such as image preprocessing and leaf instance segmentation, compresses the segmentation results, and uploads them to the cloud server. The cloud server performs decision-making tasks such as color feature analysis, morphological feature analysis, and survival status assessment, and stores the final assessment results in the database. This edge-cloud collaborative computing architecture fully utilizes the low-latency processing capabilities of the edge computing node and leverages the advantages of the cloud server in data storage and complex model inference, meeting the parallel monitoring needs of hundreds to thousands of seedlings in large-scale transplantation projects. Preferably, the edge computing node is implemented using an industrial control computer equipped with an embedded GPU, and a single node can support the real-time processing of 8 monitoring cameras.

[0059] To verify the effectiveness of the method of this invention, a test experiment was conducted in an actual urban greening transplantation project. The experimental site was located in a newly built urban park in North China. The transplanted tree species included three common urban greening tree species: ginkgo, Chinese scholar tree, and ash, totaling 300 transplanted seedlings (100 seedlings of each species). The diameter at breast height of the seedlings ranged from 8cm to 15cm, and the transplanting time was mid-April in spring.

[0060] In one embodiment of the present invention, data acquisition is performed using fixed monitoring cameras. Twelve network monitoring cameras are deployed in the transplanted area. Each camera has a resolution of 16 megapixels, a focal length of 16mm, an installation height of 3.5m, and a viewing angle of 45°. Canopy images are automatically acquired once daily at 10:00 AM, with continuous monitoring for 90 days. Simultaneously, experienced forestry technicians conduct weekly manual inspections as a baseline for comparison.

[0061] Experimental results show that, in the survival assessment on day 30 post-transplantation, the survival assessment accuracy of the method of this invention reached 92.3%, with a precision rate of 95.1% for healthy recovery seedlings and a recall rate of 93.7% for severely stressed seedlings. In contrast, the survival assessment accuracy using a single-dimensional assessment method based solely on color features was only 83.6%, indicating that the dual-dimensional fusion strategy of color and morphology improved accuracy by approximately 8.7 percentage points compared to the single-color-dimensional method. Furthermore, compared to the single-dimensional assessment method based solely on morphological features (accuracy 86.2%), the dual-dimensional fusion method of this invention also showed significant advantages, improving accuracy by approximately 6.1 percentage points, verifying the information complementarity between color and morphological features, and that their fusion can produce a synergistic enhancement effect greater than the sum of its parts (1+1>2).

[0062] Regarding leaf segmentation accuracy, the encoder-decoder network employing this invention, which combines channel attention and spatial attention in a concatenated mechanism, achieved an average intersection-over-union (mIoU) of 82.6% on the transplanted seedling leaf segmentation test set, representing a 6.3 percentage point improvement compared to the baseline network without the attention mechanism (mIoU of 76.3%). The improvement brought by the attention mechanism was particularly significant in segmentation accuracy in overlapping areas between leaves and in areas where leaves intersect with branches, with the F1 score of boundary pixels increasing from 0.69 to 0.78.

[0063] Furthermore, during the early stress detection window from day 15 to day 25 post-transplantation, the method of this invention successfully identified 47 seedlings exhibiting early signs of wilting through the morphological feature analysis module. Of these, 38 recovered after timely watering intervention, while manual inspection only identified 21 seedlings during the same period. This indicates that the detection sensitivity of the method of this invention for early signs of stress is approximately 2.2 times that of manual inspection. From a time efficiency perspective, with a deployment of 12 monitoring cameras, the system of this invention completes the full-process monitoring and evaluation of 300 seedlings in approximately 25 minutes, while manual inspection of 300 seedlings would take approximately 2 working days. The monitoring efficiency of the system is approximately 38 times that of the manual method.

[0064] The time-series recovery trend dynamic correction mechanism reduced the misjudgment rate of survival assessment from 12.1% to 7.7%, especially for seedlings that recovered spontaneously after experiencing brief stress, effectively avoiding the misjudgment of temporary stress as irreversible death. Preferably, the time-series correction had the most significant reduction in misjudgment rate in Ginkgo biloba (from 14.3% to 6.8%), which is consistent with the biological characteristics of Ginkgo biloba, where leaves are prone to temporary yellowing in the early stages of transplantation but have a strong ability to recover later. The closed-loop feedback mechanism also played an important role in the monitoring of severely stressed seedlings. By adaptively lowering the segmentation threshold, the detection rate of withered leaves increased from 81.4% to 91.2%, further ensuring the continuity and reliability of monitoring and assessment throughout the entire stress stage.

[0065] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for monitoring the survival status of transplanted seedlings based on image analysis, characterized in that, Includes the following steps: Step S1, Image Acquisition and Preprocessing: Periodically acquire images of the canopy of transplanted seedlings using an image acquisition device. Perform illumination correction processing on the acquired canopy images to eliminate the interference of ambient light changes on image colors. Perform background separation processing on the illumination-corrected images to extract the canopy area image. The illumination correction processing includes white balance correction based on the gray world hypothesis and color calibration based on a reference color chart. Step S2, Leaf Instance Segmentation Step: Input the canopy region image obtained in step S1 into the leaf segmentation network, perform instance segmentation processing on each leaf in the canopy region image, and output the leaf mask and leaf position coordinates corresponding to each leaf. The leaf segmentation network is a deep convolutional neural network based on an encoder-decoder architecture, and combines an attention mechanism to enhance the leaf edge features. Step S3, color feature analysis step: Convert the leaf regions corresponding to each leaf mask obtained in step S2 to multiple color spaces respectively. Calculate the green retention index, yellowing ratio and browning area ratio of the leaves in each color space. Obtain the comprehensive score of leaf color stress based on the weighted fusion of each color index. Among them, the multiple color spaces include HSV color space and CIELAB color space. Step S4, Morphological Feature Analysis Step: Based on the leaf masks obtained in Step S2, extract the degree of leaf curling, wilting and drooping angle and leaf area change rate, and nonlinearly fuse the degree of leaf curling, wilting and drooping angle and leaf area change rate to obtain the comprehensive degree of morphological stress. Among them, the leaf area change rate is obtained by comparing the leaf area at the current collection time point with the leaf area at the previous collection time point. Step S5, survival status assessment step: Input the leaf color stress comprehensive score obtained in step S3 and the morphological stress comprehensive score obtained in step S4 into the survival probability assessment model, and output the survival probability score and health level classification of the transplanted seedlings.

2. The method according to claim 1, characterized in that, In step S1, the illumination correction process specifically includes: simultaneously capturing a standard color chart containing known color values ​​when acquiring images; constructing a color correction matrix based on the deviation between the actual acquired value and the standard value of the standard color chart; and using the color correction matrix to perform color correction on each pixel in the canopy image, wherein the standard color chart contains at least 18 color blocks.

3. The method according to claim 1, characterized in that, In step S3, the green retention index of the leaf is obtained by averaging the ratio of the negative value of the a* channel value of each pixel in the CIELAB color space to the magnitude of the pixel's chromaticity vector. The yellowing ratio is the ratio of the number of pixels in the leaf area whose hue angle is within a preset yellowing hue range to the total number of pixels in the leaf. The browning area ratio is the ratio of the area of ​​pixels in the leaf area whose brightness value is lower than a preset browning brightness threshold and whose chromaticity value is lower than a preset browning chromaticity threshold to the total area of ​​the leaf.

4. The method according to claim 1, characterized in that, In step S4, the degree of leaf curling is characterized by the ratio between the actual area of ​​the leaf mask and its minimum circumscribed convex hull area, and the wilting and drooping angle is characterized by the angle between the leaf midrib direction and the horizontal direction. The leaf midrib direction is obtained by fitting the leaf mask after performing skeleton extraction.

5. The method according to claim 1, characterized in that, In step S3, the comprehensive score of leaf color stress is calculated as follows: after normalizing the leaf green retention index, yellowing ratio and browning area ratio respectively, the scores are weighted and summed using preset color weight coefficients, wherein the weight coefficient of the green retention index is greater than the weight coefficient of the yellowing ratio, and the weight coefficient of the yellowing ratio is greater than the weight coefficient of the browning area ratio.

6. The method according to claim 1, characterized in that, In step S4, the nonlinear fusion method of the morphological stress comprehensive degree is as follows: after mapping the degree of curling, wilting and drooping angle and the leaf area change rate to a unified dimension through the Sigmoid function, the weighted geometric mean is used for fusion, thereby producing an amplification effect when any morphological index reaches the severe stress level.

7. The method according to claim 1, characterized in that, In step S5, the survival probability assessment model integrates color and morphological features based on a Bayesian posterior inference framework and dynamically corrects the model by combining the temporal recovery trend of multi-time point images. The health level classification result is fed back to step S2 to adaptively adjust the segmentation sensitivity parameters of the leaf segmentation network. The dynamic correction of the temporal recovery trend specifically includes: obtaining the survival probability score sequence corresponding to multiple consecutive acquisition time points before the current acquisition time point; calculating the recovery trend value of the survival probability score sequence using an exponentially weighted moving average method; and using the recovery trend value to perform weighted correction on the survival probability score of the current acquisition time point. The attenuation coefficient in the exponentially weighted moving average method is adaptively adjusted according to the number of days after transplantation.

8. The method according to claim 7, characterized in that, In step S2, the encoder of the leaf segmentation network uses ResNet series networks as the backbone network to extract multi-scale feature maps. The decoder recovers spatial resolution by upsampling layer by layer and fusing the feature maps of the corresponding layers of the encoder. The attention mechanism is a cascaded combination of channel attention and spatial attention.

9. The method according to claim 8, characterized in that, In step S5, the health level classification includes dividing the transplanted seedlings into four levels: healthy recovery, mild stress, moderate stress, and severe stress. The survival probability score is greater than the first threshold, which corresponds to the healthy recovery level. The survival probability score is between the first and second thresholds, which corresponds to the mild stress level. The survival probability score is between the second and third thresholds, which corresponds to the moderate stress level. The survival probability score is less than the third threshold, which corresponds to the severe stress level.

10. A transplant seedling survival status monitoring system based on image analysis, used to implement the method described in any one of claims 1-9, characterized in that, include: The image acquisition and preprocessing module is used to periodically acquire images of the canopy of transplanted seedlings through an image acquisition device, perform illumination correction processing on the acquired canopy images to eliminate the interference of changes in ambient light on image colors, and perform background separation processing on the illumination-corrected images to extract the canopy area image. The leaf instance segmentation module is used to input the canopy region image output by the image acquisition and preprocessing module into the leaf segmentation network, perform instance segmentation processing on each leaf in the canopy region image, and output the leaf mask and leaf position coordinates corresponding to each leaf. The color feature analysis module is used to convert the leaf regions corresponding to each leaf mask output by the leaf instance segmentation module to multiple color spaces, calculate the green retention index, yellowing ratio and browning area ratio of the leaf, and obtain the comprehensive score of leaf color stress based on the weighted fusion of each color index. The morphological feature analysis module is used to extract the degree of leaf curling, wilting and drooping angle and leaf area change rate based on the leaf mask output by the leaf instance segmentation module, and to nonlinearly fuse the degree of leaf curling, wilting and drooping angle and leaf area change rate to obtain the morphological stress comprehensive degree. The survival status assessment module is used to input the leaf color stress comprehensive score output by the color feature analysis module and the morphological stress comprehensive degree output by the morphological feature analysis module into the survival probability assessment model, output the survival probability score and health level classification of the transplanted seedlings, and feed back the health level classification results to the leaf instance segmentation module to adaptively adjust the segmentation sensitivity parameters of the leaf segmentation network.

Images