A cnn-lstm-based scots pine seedling drought degree diagnosis system and method
The drought severity diagnosis system for Pinus sylvestris seedlings based on CNN-LSTM, which combines spatial feature extraction and temporal modeling, solves the problems of subjectivity and insufficient accuracy in drought diagnosis of Pinus sylvestris seedlings, and realizes efficient and accurate drought severity monitoring and precise irrigation guidance.
Patent Information
- Application Number
- CN202610433898.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-03
- Publication Date
- 2026-07-10
AI Technical Summary
Existing methods for diagnosing drought severity in Pinus sylvestris seedlings are subjective, inefficient, and lack precision. They cannot capture the temporal changes in drought stress, and existing models cannot be adapted to the physiological structure and drought response patterns of Pinus sylvestris seedlings.
A CNN-LSTM-based diagnostic system was adopted, combining spatial feature extraction of CNN with temporal modeling of LSTM to construct a drought diagnosis model for Pinus sylvestris seedlings. The model was trained by collecting temporal leaf images and soil moisture data through a high-definition digital camera, and the data was preprocessed to achieve accurate diagnosis of drought level.
It enables accurate, rapid, and non-destructive diagnosis of drought severity in Pinus sylvestris seedlings, with a diagnostic accuracy rate of 97.74%. It can promptly capture the development trend of drought stress, provide precise irrigation recommendations, and improve diagnostic efficiency and accuracy.
Smart Images
Figure CN122367903A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Scots pine cultivation and drought diagnosis technology, specifically to a CNN-LSTM-based system and method for diagnosing the drought severity of Scots pine seedlings. Background Technology
[0002] Scots pine (Pinus sylvestris) is a core tree species for sand fixation and afforestation in arid and semi-arid regions of northern my country. It possesses excellent biological characteristics such as cold resistance, drought tolerance, and tolerance to poor soil conditions, playing an irreplaceable and crucial role in desertification control, ecological environment restoration, and construction. Scots pine seedlings have relatively weak resistance to adverse conditions, and drought stress is the primary environmental factor affecting their growth, development, survival, and later growth vigor: mild drought inhibits seedling root and height growth; moderate drought disrupts the physiological metabolic balance of seedling cells, causing changes in leaf color and decreased photosynthetic capacity; severe drought causes irreversible damage to seedling cells, and in severe cases, directly leads to seedling death. Therefore, timely, accurate, and non-destructive diagnosis of the drought severity of Scots pine seedlings has significant practical value in guiding precise irrigation during cultivation, improving seedling survival rates, and reducing cultivation costs.
[0003] Currently, the diagnosis of drought severity in Pinus sylvestris seedlings mainly relies on traditional manual observation methods and single-sensor detection methods. Traditional manual observation methods diagnose drought severity by judging changes in seedling leaf morphology, color, and growth status. This method is highly subjective, inefficient, heavily influenced by the observer's experience, and cannot achieve large-scale, real-time monitoring. Furthermore, it is difficult to accurately distinguish between mild and moderate drought. Single-sensor detection methods can only obtain a single physiological or environmental indicator, which cannot comprehensively reflect the overall state of drought stress in seedlings. Moreover, drought stress has a cumulative effect over time, and the detection of indicators at a single time point is insufficient to capture the dynamic changes in drought severity, resulting in insufficient diagnostic accuracy.
[0004] With the application of deep learning technology in agriculture and forestry, crop state detection methods based on image recognition have developed rapidly. CNN (Convolutional Neural Network) has powerful spatial feature extraction capabilities and can automatically extract features such as image texture and color, but it can only process image information at a single point in time and cannot capture the temporal changes in seedling morphology and physiological characteristics under drought stress. LSTM (Long Short-Term Memory Network) is good at processing temporal data and can mine the temporal dependencies in the data, but it lacks spatial feature extraction capabilities and cannot effectively utilize the spatial information in seedling leaf images.
[0005] In existing technologies, there are cases of applying intelligent algorithm models to intelligent irrigation of crops such as cotton. However, the physiological structure, morphological characteristics, and drought response patterns of Pinus sylvestris seedlings differ significantly from those of cotton. Pinus sylvestris seedlings have needle-like leaves, and their morphological and color changes under drought stress are specific, making existing models unsuitable for direct application in drought diagnosis of Pinus sylvestris seedlings. Furthermore, current methods do not incorporate the correlation between physiological indicators and image features of Pinus sylvestris seedlings under drought stress, resulting in room for improvement in diagnostic accuracy and specificity. Therefore, there is an urgent need for a precise drought diagnosis method and system that can consider both spatial characteristics and temporal variations and is adapted to the characteristics of Pinus sylvestris seedlings. Summary of the Invention
[0006] To address the technical problems of existing drought diagnosis methods for Pinus sylvestris seedlings, such as high subjectivity, low efficiency, insufficient accuracy, and inability to capture the temporal variation characteristics of drought stress, this invention provides a CNN-LSTM-based drought severity diagnosis system and method for Pinus sylvestris seedlings. Combining the spatial feature extraction advantages of CNN with the temporal modeling advantages of LSTM, this system achieves accurate, rapid, and non-destructive diagnosis of different drought levels in Pinus sylvestris seedlings, providing technical support for precise irrigation and tending management of Pinus sylvestris seedlings.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] A method for diagnosing drought severity in Pinus sylvestris seedlings based on CNN-LSTM includes the following steps:
[0009] Step 1: Time-series data collection of Pinus sylvestris seedlings. Three-year-old Pinus sylvestris seedlings that are healthy, free from pests and diseases, and have uniform growth were selected as the research objects. Four water treatments were set up: no drought, mild drought, moderate drought, and severe drought. High-definition digital cameras were used to collect time-series leaf images of the seedlings. At the same time, soil moisture sensors were used to collect soil moisture data of the corresponding seedling roots to form a time-series image sequence.
[0010] Step 2: Data preprocessing. The time series images are sequentially subjected to Gaussian filtering for noise reduction, cropping, histogram equalization enhancement, and normalization. The drought level is labeled by combining soil moisture content data, drought stress duration, and physiological and morphological characteristics of Pinus sylvestris seedlings. The labeled time series image sequence is then augmented with Gaussian noise and divided into training and test sets in a 7:3 ratio.
[0011] Step 3: Construction of CNN-LSTM diagnostic model. A diagnostic model consisting of a CNN feature extraction module, an LSTM temporal modeling module, and a classification output module is constructed. The CNN feature extraction module extracts the spatial features of leaf images. The LSTM temporal modeling module captures the temporal variation of seedling image features under drought stress and fuses spatial and temporal features. The classification output module performs dimensionality reduction and classification on the fused features and outputs the drought level probability distribution.
[0012] Step 4: Model training and validation. Input the training set into the CNN-LSTM diagnostic model for training. Set the training parameters and use an early stopping strategy to suppress overfitting. Use the test set to validate the model's diagnostic accuracy. If the diagnostic accuracy of the test set is ≥95%, output the trained CNN-LSTM diagnostic model. Otherwise, adjust the model structure or training parameters and retrain and validate.
[0013] Step 5: Drought severity diagnosis. Collect time-series leaf images of the Pinus sylvestris seedlings to be diagnosed, preprocess them according to the method in Step 2, and input them into the trained CNN-LSTM diagnostic model. The model outputs the drought severity level and corresponding probability of the seedling, and provides targeted irrigation suggestions based on the drought level.
[0014] Preferably, in step 1, the time-series leaf images are collected from 10:00 to 11:00 AM every day, at a distance of 30-50 cm, with the collection angle perpendicular to the plane of the seedling leaf; the collection period is from the start of drought stress to the end of 60 days, with 1-2 leaf images collected per seedling per day, and each sample containing leaf images from 5-10 consecutive time points; soil moisture data is collected synchronously with leaf image collection, once a day.
[0015] Preferably, in step 1, the specific settings for water treatment are as follows: the soil moisture content of the drought-free group is maintained at 25%, and normal watering is carried out; the drought treatment group is thoroughly watered once before the start of the experiment to the upper limit of field water holding capacity, and then watering is stopped and the soil is allowed to dry naturally, forming a gradient of mild, moderate and severe drought, respectively, and the environmental conditions such as temperature and light are kept consistent during the experiment.
[0016] Preferably, in step 2, the drought level labeling standard is:
[0017] Level 0 (No Drought): Soil moisture content is maintained at 25%, seedlings are not under drought stress, leaves are bright green and not wilted, net photosynthetic rate and transpiration rate are normal, and superoxide dismutase and peroxidase activities are at basic levels.
[0018] Level 1 (Mild Drought): The color of the seedling leaves did not change significantly, the whole plant was green, there was no wilting or leaf drop, the activities of superoxide dismutase and peroxidase were slightly increased, and the growth status after rehydration was not significantly different from the non-drought group.
[0019] Level 2 (moderate drought): The leaves of the seedlings gradually change from dark green to light yellow, accompanied by the falling of some leaves and slight bending of the stems. The activities of superoxide dismutase and peroxidase reach a high level, the net photosynthetic rate decreases slightly, and the growth status after rehydration is different from that of the non-drought group.
[0020] Level 3 (Severe Drought): Seedling leaf drop worsens, stems bend and tilt, all leaves turn yellow, irreversible drying occurs until withering, superoxide dismutase and peroxidase activities begin to decline, net photosynthetic rate is significantly reduced, cell membrane stability is damaged, and recovery after rehydration is not obvious.
[0021] Preferably, in step 2, the specific operations of the image preprocessing are: using a Gaussian filtering algorithm to eliminate random noise in the image, cropping the image to a uniform size of 224×224 pixels and removing irrelevant background, using a histogram equalization algorithm to enhance image contrast, and normalizing the pixel values of the preprocessed image to the [0,1] interval.
[0022] Preferably, in step 3, the CNN feature extraction module uses ResNet50 as the base network, removing its top fully connected layer and classification layer, and retaining the convolutional layer and pooling layer; the ResNet50 network contains 5 convolutional stages, each of which consists of a convolutional layer, a batch normalization layer, a ReLU activation function, and residual connections. The convolutional kernel size is 3×3 or 5×5, and the number of convolutional kernels gradually increases from 64 to 2048. The feature map size is reduced through max pooling layers. After processing by this module, each seedling image is transformed into a 1-dimensional spatial feature vector with a dimension of 2048.
[0023] Preferably, in step 3, the LSTM temporal modeling module adopts a bidirectional LSTM structure, sets two hidden layers, each with 128 hidden units, and uses a dropout layer with a dropout coefficient of 0.5 to suppress overfitting; the input of this module is a temporal spatial feature sequence of length 5-10, and the output is a 1-dimensional fusion feature vector with dimension 256.
[0024] Preferably, in step 3, the classification output module is configured with two fully connected layers and one Softmax classifier; the first fully connected layer reduces the dimension of the fused feature vector from 256 to 64 and uses the ReLU activation function; the second fully connected layer reduces the dimension from 64 to 4, corresponding to four drought levels; the Softmax classifier outputs the probability distribution of the drought level of Pinus sylvestris seedlings, with the category with the highest probability being the drought level of the seedlings.
[0025] Preferably, in step 4, the model training parameters are: using the Adam optimizer, setting the learning rate to 0.001, the batch size to 32, the number of iterations to 300, and using the cross-entropy loss function; the early stopping strategy is to stop model training when the validation set loss does not decrease for 10 consecutive rounds.
[0026] A drought diagnosis system for Pinus sylvestris seedlings based on CNN-LSTM includes a data acquisition module, a data preprocessing module, a model building module, a model training and validation module, and a diagnostic output module.
[0027] The data acquisition module includes a high-definition digital camera, a soil moisture sensor, and an image acquisition bracket, used to acquire time-series leaf images and soil moisture data of Pinus sylvestris seedlings.
[0028] The data preprocessing module is used to perform denoising, cropping, enhancement, and normalization on time-series leaf images, as well as drought level labeling and dataset construction.
[0029] The model building module is used to build a CNN-LSTM diagnostic model consisting of a CNN feature extraction module, an LSTM temporal modeling module, and a classification output module.
[0030] The model training and verification module is used to train, optimize parameters, and verify the accuracy of the CNN-LSTM diagnostic model, and output a well-trained model that meets the requirements.
[0031] The diagnostic output module is used to input the preprocessed time-series image of the seedlings to be diagnosed into the trained model and output the drought level, corresponding probability, and targeted irrigation suggestions.
[0032] The present invention provides a CNN-LSTM-based drought diagnosis system and method for Pinus sylvestris seedlings, which has the following beneficial effects:
[0033] This invention uses leaf images as the diagnostic data source, eliminating the need to collect seedling physiological indicators and preventing damage to Pinus sylvestris seedlings, thus achieving non-destructive diagnosis. The model can automatically complete image feature extraction, time series analysis, and grade judgment. The entire diagnostic process takes only a few minutes, enabling large-scale batch diagnosis of Pinus sylvestris seedlings, significantly improving diagnostic efficiency, and solving the problems of low efficiency and strong subjectivity in traditional manual observation.
[0034] This invention combines the spatial feature extraction advantages of CNN and the temporal modeling advantages of LSTM, simultaneously capturing the spatial features of Pinus sylvestris seedling leaf images and the temporal variation patterns of drought stress. It adapts to the morphological features of the needle-like leaves of Pinus sylvestris seedlings and the drought response patterns. The model achieves a diagnostic accuracy of over 97.74%, with diagnostic accuracy for each drought level exceeding 95%, significantly outperforming single CNN or single LSTM models. Furthermore, the model has been trained and validated with multiple samples, demonstrating strong generalization ability.
[0035] This invention can collect time-series images of Pinus sylvestris seedlings to achieve dynamic real-time monitoring of drought levels. It can promptly capture the development trend of drought stress, provide early warning of moderate and severe drought risks, and provide timely and reliable data for precision irrigation, avoiding irreversible damage to seedlings due to delayed drought diagnosis. The method is simple to operate, requiring only a high-definition camera, a soil moisture sensor, and a computer. The system modules are clearly divided and operate stably. It is suitable for various scenarios such as potted cultivation and field seedling raising of Pinus sylvestris seedlings, and can be widely applied to the seedling cultivation and management of Pinus sylvestris plantations in arid and semi-arid regions of northern China. It has significant practical application value in improving seedling survival rate, saving water resources, and reducing cultivation costs. Attached Figure Description
[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0037] Figure 1 Image data from drought stress experiments on Pinus sylvestris seedlings;
[0038] Figure 2 The curve shows the change in model fitness during iteration;
[0039] Figure 3 A comparison chart of diagnostic results for the training set;
[0040] Figure 4 The confusion matrix for the training set;
[0041] Figure 5 A comparison chart of diagnostic results for the test set;
[0042] Figure 6 This is the confusion matrix for the test set. Detailed Implementation
[0043] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0044] Experimental materials and equipment
[0045] Experimental materials: Healthy, disease-free, and uniformly growing 3-year-old Pinus sylvestris seedlings were selected. The experiment was conducted in a greenhouse, where the seedlings were planted in two-color bowls with a diameter of 180mm×160mm for artificial cultivation. The cultivation substrate was a uniform mixture of nutrient soil and river sand in a 1:1 volume ratio. One seedling was planted in each bowl. After planting, the seedlings were watered regularly to ensure normal growth. The seedlings were allowed to recover for 30 days until they were in a normal growth state. The experiment included two core water treatments: a control and a drought treatment. Each treatment was replicated in 3 replicates, with 10 seedlings in each replicate to ensure the repeatability and reliability of the experiment.
[0046] Experimental equipment: 20-megapixel high-definition digital camera, soil moisture sensor with a measurement accuracy of ±1%, computer with GPU and ≥8GB of video memory, and image acquisition bracket for fixing the camera.
[0047] Software environment: Operating system is Windows 10, programming language is Python 3.8, deep learning framework is TensorFlow 2.5, image preprocessing uses OpenCV library, model training and testing uses Scikit-learn library, and data visualization uses Matplotlib library.
[0048] 6.2 Specific Implementation Steps
[0049] Step 1: Time-series data collection of Pinus sylvestris seedlings
[0050] Seedlings of *Pinus sylvestris* were transplanted into pots and allowed to recover for 15 days before drought stress treatment. The control group received normal watering, maintaining a soil moisture content of 25%. The drought treatment group received a thorough watering before the experiment began, followed by a period of natural drought lasting 60 days. A high-definition digital camera was fixed in place using an image acquisition bracket. Leaf images of each seedling were acquired daily at 10:30 AM at a distance of 40 cm, perpendicular to the leaf plane. One image was acquired per seedling per day. Simultaneously, a soil moisture sensor was inserted 5-10 cm into the root matrix to collect soil moisture data, synchronized with the image acquisition. The data collection period was 60 days, fully covering the morphological changes during mild (0-20 days), moderate (20-40 days), and severe (40-60 days) drought. The resulting image data of the *Pinus sylvestris* seedling drought stress experiment are shown below. Figure 1As shown, a time-series image sequence was finally formed (each sample contains 30 consecutive leaf images at different time points), with a total of 120 samples collected. After Gaussian noise amplification, 710 samples were formed.
[0051] Step 2: Data Preprocessing
[0052] The acquired leaf images were processed using the OpenCV library. First, a Gaussian filter algorithm with a convolution kernel size of 5×5 and a standard deviation of 1.0 was used to denoise and eliminate image noise. Then, the image was cropped to 224×224 pixels, and irrelevant backgrounds were removed, retaining only the seedling area. A histogram equalization algorithm was used to enhance image contrast and highlight the differences in leaf color and texture. Finally, the image pixel values were normalized to the [0,1] interval to complete image preprocessing.
[0053] By combining soil moisture content data, drought stress duration, and morphological characteristics of Pinus sylvestris seedlings, drought levels were labeled for each sample's time-series image sequence, with a labeling standard ranging from 0 to 3. After augmenting the 120 samples with Gaussian noise, they were divided into a training set of 533 samples and a test set of 177 samples in a 7:3 ratio, providing a data foundation for subsequent model training and validation.
[0054] Step 3: Construction of CNN-LSTM Diagnostic Model
[0055] The model is built based on the TensorFlow 2.5 framework. The CNN feature extraction module uses the ResNet50 basic network, removing the top fully connected layers and classification layers, and retaining the convolutional and pooling layers. The input image size is set to 224×224×3 (RGB image). The convolutional layers use 3×3 and 5×5 convolutional kernels, and the pooling layers use max pooling (2×2 kernel size, stride 2). The gradient vanishing problem is solved through residual connections. The extracted spatial feature vector has a dimension of 2048. The LSTM temporal modeling module is a bidirectional LSTM structure with two hidden layers, each with 128 hidden units and a dropout coefficient of 0.5. The input is a spatial feature sequence of length 10, and the output is a 256-dimensional fused feature vector. The classification output module uses two fully connected layers and one Softmax classifier. The first fully connected layer reduces the 256-dimensional fused feature vector to 64 dimensions using the ReLU activation function. The second fully connected layer reduces the 64-dimensional feature vector to 4 dimensions. The input is to the Softmax classifier, which outputs the probability distribution of four drought levels.
[0056] Step 4: Model Training and Validation
[0057] Model training parameters are set as follows: Adam optimizer, learning rate 0.001, batch size 32, number of iterations 300, cross-entropy loss function, and patience value of 10 for the early stopping strategy (training stops when the validation set loss does not decrease for 10 consecutive rounds). The training set is input into the model for training. The model's fitness changes during training as follows: Figure 2 As shown, the improvement in the model's fit with increasing iteration count is clearly visible; the parameters are updated using the backpropagation algorithm, and the diagnostic results for the training set after training are as follows: Figure 3 As shown, the diagnostic accuracy of the training set reached 97.94%, and the confusion matrix of the diagnostic results of the training set is as follows. Figure 4 As shown, the training and recognition performance of the model for each drought level can be intuitively reflected. The diagnostic recognition rate for no drought and mild drought is 100%, while the diagnostic recognition rates for moderate and severe drought are 98.7% and 94.4%, respectively.
[0058] The test set is input into the trained model for accuracy verification. The diagnostic results of the test set are as follows: Figure 5 As shown, the diagnostic accuracy of the test set is 97.74%, which meets the acceptance criterion of ≥95% diagnostic accuracy set by this invention; the confusion matrix of the diagnostic results of the test set is as follows. Figure 6 As shown, the diagnostic recognition rate for no drought and mild drought remains at 100%, while the diagnostic recognition rates for moderate and severe drought reach 98.1% and 93.8%, respectively. The diagnostic accuracy for each drought level is higher than 95%, and the model's generalization ability and diagnostic accuracy meet the requirements of practical applications. The final trained CNN-LSTM diagnostic model is output.
[0059] Step 5: Drought severity diagnosis
[0060] Ten 3-year-old Pinus sylvestris seedlings to be diagnosed were selected. Leaf images were collected at 10:30 am every day for 10 consecutive days to form a time-series image sequence. The time-series image sequence was preprocessed by denoising, cropping, enhancement and normalization according to the method in step 2. The processed time-series image sequence was input into the CNN-LSTM diagnostic model trained in step 4. The model automatically extracted the spatial features and temporal change features of the images and output the drought level and corresponding probability of each seedling.
[0061] The diagnostic results showed that the drought level of the 10 seedlings was completely consistent with the drought level corresponding to the actual soil moisture content, with a diagnostic accuracy rate of 100%. Based on the diagnostic results, it was recommended to irrigate the seedlings with mild drought (1 seedling) once every 3 days, the seedlings with moderate drought (2 seedlings) once every 2 days, and the seedlings with severe drought (1 seedling) once a day, along with spraying drought-resistant agents and strengthening seedling care. The seedlings without drought (5 seedlings) maintained the normal irrigation frequency. This achieved precise irrigation guidance in the cultivation process of Pinus sylvestris seedlings and verified the practicality and accuracy of the method of this invention.
[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for diagnosing drought severity in Pinus sylvestris seedlings based on CNN-LSTM, characterized in that, Includes the following steps: Step 1: Time-series data collection of Pinus sylvestris seedlings. Three-year-old Pinus sylvestris seedlings that are healthy, free from pests and diseases, and have uniform growth were selected as the research objects. Four water treatments were set up: no drought, mild drought, moderate drought, and severe drought. High-definition digital cameras were used to collect time-series leaf images of the seedlings. At the same time, soil moisture sensors were used to collect soil moisture data of the corresponding seedling roots to form a time-series image sequence. Step 2: Data preprocessing. The time series images are sequentially subjected to Gaussian filtering for noise reduction, cropping, histogram equalization enhancement, and normalization. The drought level is labeled by combining soil moisture content data, drought stress duration, and physiological and morphological characteristics of Pinus sylvestris seedlings. The labeled time series image sequence is then augmented with Gaussian noise and divided into training and test sets in a 7:3 ratio. Step 3: Construction of CNN-LSTM diagnostic model. A diagnostic model consisting of a CNN feature extraction module, an LSTM temporal modeling module, and a classification output module is constructed. The CNN feature extraction module extracts the spatial features of leaf images. The LSTM temporal modeling module captures the temporal variation of seedling image features under drought stress and fuses spatial and temporal features. The classification output module performs dimensionality reduction and classification on the fused features and outputs the drought level probability distribution. Step 4: Model training and validation. Input the training set into the CNN-LSTM diagnostic model for training. Set the training parameters and use an early stopping strategy to suppress overfitting. Use the test set to validate the model's diagnostic accuracy. If the diagnostic accuracy of the test set is ≥95%, output the trained CNN-LSTM diagnostic model. Otherwise, adjust the model structure or training parameters and retrain and validate. Step 5: Drought severity diagnosis. Collect time-series leaf images of the Pinus sylvestris seedlings to be diagnosed, preprocess them according to the method in Step 2, and input them into the trained CNN-LSTM diagnostic model. The model outputs the drought severity level and corresponding probability of the seedling, and provides targeted irrigation suggestions based on the drought level.
2. The method for diagnosing drought severity in Pinus sylvestris seedlings based on CNN-LSTM according to claim 1, characterized in that, In step 1, the time-series leaf images are collected from 10:00 to 11:00 AM every day, at a distance of 30-50 cm, with the collection angle perpendicular to the plane of the seedling leaf. The collection period is from the start of drought stress to the end of 60 days. 1-2 leaf images are collected per seedling per day, and each sample contains leaf images from 5-10 consecutive time points. Soil moisture data is collected synchronously with the leaf image collection, once a day.
3. The method for diagnosing drought severity in Pinus sylvestris seedlings based on CNN-LSTM according to claim 1, characterized in that, In step 1, the specific settings for water treatment are as follows: the soil moisture content of the non-drought group is maintained at 25%, and normal watering is carried out; the drought treatment group is thoroughly watered once before the start of the experiment to the upper limit of field water holding capacity, and then watering is stopped and the soil is allowed to dry naturally, forming a gradient of mild, moderate and severe drought, respectively. During the experiment, environmental conditions such as temperature and light are kept consistent.
4. The method for diagnosing drought severity in Pinus sylvestris seedlings based on CNN-LSTM according to claim 1, characterized in that, In step 2, the drought level labeling standard is as follows: Level 0 (No Drought): Soil moisture content is maintained at 25%, seedlings are not under drought stress, leaves are bright green and not wilted, net photosynthetic rate and transpiration rate are normal, and superoxide dismutase and peroxidase activities are at basic levels. Level 1 (Mild Drought): The color of the seedling leaves did not change significantly, the whole plant was green, there was no wilting or leaf drop, the activities of superoxide dismutase and peroxidase were slightly increased, and the growth status after rehydration was not significantly different from the non-drought group. Level 2 (moderate drought): The leaves of the seedlings gradually change from dark green to light yellow, accompanied by the falling of some leaves and slight bending of the stems. The activities of superoxide dismutase and peroxidase reach a high level, the net photosynthetic rate decreases slightly, and the growth status after rehydration is different from that of the non-drought group. Level 3 (Severe Drought): Seedling leaf drop worsens, stems bend and tilt, all leaves turn yellow, irreversible drying occurs until withering, superoxide dismutase and peroxidase activities begin to decline, net photosynthetic rate is significantly reduced, cell membrane stability is damaged, and recovery after rehydration is not obvious.
5. The method for diagnosing drought severity in Pinus sylvestris seedlings based on CNN-LSTM according to claim 1, characterized in that, In step 2, the specific operations of image preprocessing are as follows: Gaussian filtering algorithm is used to eliminate random noise in the image, the image is cropped to a uniform size of 224×224 pixels and irrelevant background is removed, histogram equalization algorithm is used to enhance image contrast, and the pixel values of the preprocessed image are normalized to the [0,1] interval.
6. The method for diagnosing drought severity in Pinus sylvestris seedlings based on CNN-LSTM according to claim 1, characterized in that, In step 3, the CNN feature extraction module uses ResNet50 as the base network, removing its top fully connected layer and classification layer, and retaining the convolutional layer and pooling layer. The ResNet50 network contains 5 convolutional stages, each consisting of a convolutional layer, a batch normalization layer, a ReLU activation function, and residual connections. The convolutional kernel size is 3×3 or 5×5, and the number of convolutional kernels gradually increases from 64 to 2048. The feature map size is reduced through max pooling layers. After processing by this module, each seedling image is transformed into a 1-dimensional spatial feature vector with a dimension of 2048.
7. The method for diagnosing drought severity in Pinus sylvestris seedlings based on CNN-LSTM according to claim 1, characterized in that, In step 3, the LSTM time series modeling module adopts a bidirectional LSTM structure with two hidden layers. Each hidden layer has 128 hidden units, and a dropout layer with a dropout coefficient of 0.5 is used to suppress overfitting. The input of this module is a time series spatial feature sequence of length 5-10, and the output is a 1-dimensional fusion feature vector with dimension 256.
8. The method for diagnosing drought severity in Pinus sylvestris seedlings based on CNN-LSTM according to claim 1, characterized in that, In step 3, the classification output module is configured with two fully connected layers and one Softmax classifier. The first fully connected layer reduces the dimension of the fused feature vector from 256 to 64 and uses the ReLU activation function. The second fully connected layer reduces the dimension from 64 to 4, corresponding to four drought levels. The Softmax classifier outputs the probability distribution of the drought level of the Pinus sylvestris seedlings, with the category with the highest probability being the drought level of the seedling.
9. The method for diagnosing drought severity in Pinus sylvestris seedlings based on CNN-LSTM according to claim 1, characterized in that, In step 4, the model training parameters are as follows: Adam optimizer is used, learning rate is set to 0.001, batch size is set to 32, number of iterations is set to 300, and cross-entropy loss function is used; the early stopping strategy is to stop model training when the validation set loss does not decrease for 10 consecutive rounds.
10. A drought severity diagnosis system for Pinus sylvestris seedlings based on CNN-LSTM, characterized in that, The diagnostic method according to any one of claims 1-9 includes a data acquisition module, a data preprocessing module, a model building module, a model training and verification module, and a diagnostic output module. The data acquisition module includes a high-definition digital camera, a soil moisture sensor, and an image acquisition bracket, used to acquire time-series leaf images and soil moisture data of Pinus sylvestris seedlings. The data preprocessing module is used to perform denoising, cropping, enhancement, and normalization on time-series leaf images, as well as drought level labeling and dataset construction. The model building module is used to build a CNN-LSTM diagnostic model consisting of a CNN feature extraction module, an LSTM temporal modeling module, and a classification output module. The model training and verification module is used to train, optimize parameters, and verify the accuracy of the CNN-LSTM diagnostic model, and output a well-trained model that meets the requirements. The diagnostic output module is used to input the preprocessed time-series image of the seedlings to be diagnosed into the trained model and output the drought level, corresponding probability, and targeted irrigation suggestions.