A method and system for analyzing root system traits of optical walnut

Abstract

The application provides a method and system for analyzing the root system characteristics of Prunus mira, belonging to the technical field of intelligent analysis. The method includes the following steps: collecting high-resolution images of the distribution area of Prunus mira by using a UAV, combining GIS positioning to construct an altitude gradient database; applying an image processing algorithm to classify the images, identify the crown coverage area of Prunus mira, and estimate the population density; generating a digital surface model and a digital elevation model to analyze the topographic and geomorphic features; using spectral analysis to identify soil types and land use conditions, and combining thermal infrared data to determine the distribution of soil moisture and surface temperature; collecting Prunus mira root system samples and using three-dimensional imaging technology to quantify the root system characteristics; constructing a Prunus mira resource acquisition capacity prediction model, and using a neural network model including residual attention blocks and spiral fully connected layers to predict the water and nutrient absorption rate of the root system, thereby providing a scientific basis for optimizing irrigation and fertilization strategies and improving the yield and quality of Prunus mira.

Description

Technical Field

[0001] This invention belongs to the field of intelligent analysis technology, specifically relating to a method and system for analyzing the root traits of *Juglans regia*. Background Technology

[0002] As a plant resource with significant economic and ecological value, the study of the root morphology and distribution characteristics of *Juglans regia* is crucial for revealing its growth habits, resource acquisition capabilities, environmental adaptability, and optimizing cultivation and management strategies. However, traditional root research methods often rely on cumbersome field excavation, manual measurement, or inefficient laboratory analysis, which are not only time-consuming and labor-intensive but may also cause irreversible damage to the root system, affecting normal plant growth. Although some non-destructive technologies have been applied to root research, such as underground radar and electrical impedance tomography (EIT), these methods often require complex equipment and are costly. Furthermore, these methods struggle to accurately and comprehensively capture the complex morphology and distribution patterns of the root system in three-dimensional space, failing to meet the demands of modern forestry scientific research and practice for refined and dynamic information on *Juglans regia* root systems.

[0003] Therefore, there is an urgent need for a simple, efficient and low-cost method to analyze the root shape and traits of walnut trees in order to promote the scientific management of walnut tree cultivation. Summary of the Invention

[0004] Based on the above-mentioned technical problems, the present invention provides a method and system for analyzing the root traits of *Juglans regia*, which utilizes UAV remote sensing monitoring, geographic information system positioning, hyperspectral and thermal infrared imaging, image processing and analysis, three-dimensional imaging technology and neural network model to achieve efficient, non-destructive and high-precision observation and analysis of *Juglans regia* root system.

[0005] This invention provides a method for analyzing the root traits of *Juglans regia*, the method comprising:

[0006] Step S1: Using UAV remote sensing monitoring and geographic information system positioning, construct an altitude gradient database of the distribution of *Juglans regia*.

[0007] Step S2: Collect root samples from *Juglans regia*.

[0008] Step S3: Perform three-dimensional imaging on the collected root samples to quantify root morphological characteristics;

[0009] Step S4: Construct a predictive model for the resource acquisition capacity of walnut trees to obtain the absorption rate of water and nutrients by the root system.

[0010] Optionally, the application of UAV remote sensing monitoring and geographic information system positioning to construct an elevation gradient database of the distribution of *Juglans regia* specifically includes:

[0011] High-resolution images of the distribution area of ​​*Juglans regia* were acquired using drones equipped with sensors.

[0012] The acquired images undergo data processing and analysis, specifically including:

[0013] Image processing algorithms are used to classify images and identify the canopy coverage area of ​​walnut trees; population density is calculated by combining image resolution and canopy size; digital surface models and digital elevation models are generated using UAV imagery to analyze topographic features; soil type and land use are identified by analyzing images through spectral analysis, and thermal infrared data is used to help determine soil moisture and surface temperature distribution.

[0014] An altitude gradient database was established based on the results of UAV remote sensing interpretation.

[0015] Optionally, the step of classifying the image using an image processing algorithm to identify the canopy coverage area of ​​the *Juglans regia* tree; and calculating the population density by combining the image resolution and canopy size, specifically includes:

[0016] The acquired images undergo data preprocessing, including radiometric and geometric correction.

[0017] Feature extraction is performed on the image, including extracting spectral features using reflectance information from different bands in the image; extracting texture features using the gray-level co-occurrence matrix; and extracting shape features using edge detection and morphological processing methods.

[0018] Based on the extracted features, the K-means clustering algorithm is used to classify the images, that is, the image pixels are automatically divided into several categories according to the similarity of spectral and texture features, and then the canopy area of ​​the walnut tree is identified by expert knowledge or posterior probability.

[0019] After identifying the bare canopy region of the walnut tree, canopy segmentation and boundary refinement are performed to accurately define the bare canopy region, specifically including:

[0020] The classification results are converted into a binary image using a threshold segmentation method, where the white area represents the crown of a walnut tree and the black area represents other ground features or background.

[0021] Perform dilation, erosion, opening, and closing operations on the image to optimize the canopy mask;

[0022] Calculate the population density of *Juglans regia*.

[0023] Optionally, the step of performing three-dimensional imaging on the collected root samples to quantify root morphological characteristics specifically includes:

[0024] The collected root samples are first cleaned to remove attached soil and other impurities; then they are fixed with a fixative to maintain the root morphology; the fixed root samples are placed in a resin box and filled with silica gel to ensure that the roots are evenly dispersed and without folds in the container; the sealed root samples are placed in a CT scanner and the scanning parameters are set; the CT scanning program is run, and the device will emit X-rays from multiple angles and receive transmission signals to obtain a series of tomographic images.

[0025] ImageJ image processing software was used to remove random noise generated during the scanning process through filtering algorithms; grayscale adjustment was used to improve image contrast and signal-to-noise ratio; and feature matching was used to ensure that all slice images were precisely aligned in space.

[0026] The Amira 3D reconstruction software is used to convert each 2D slice into a thin layer in 3D space to form a 3D volume dataset; a threshold is set according to the grayscale difference between the root system and the background to divide the voxel dataset into root system and non-root system parts; adjacent voxels are identified and connected to form a continuous root system structure; a triangular mesh model of the root system surface is generated.

[0027] Using a threshold segmentation method, individual root systems are automatically identified and separated based on the grayscale difference between the root system and the background; root trait characteristics are then calculated.

[0028] Optionally, the construction of the predictive model for the acquisition capacity of *Hemiberlesia lataniae* resources specifically includes:

[0029] Define a custom layer class for the residual attention block; this layer contains the unit number setting (units), i.e., the output dimension; two fully connected layers with ReLU activation function; an attention mechanism layer; an addition layer for residual connections; and a projection layer.

[0030] First, the input is forward-propagated through two fully connected layers, and the ReLU activation function is applied. Then, an attention mechanism layer is used to calculate the attention weights. Next, it is checked whether the input dimension matches the output dimension. If they do not match, the input dimension is adjusted through a projection layer. Finally, the result processed by the attention mechanism is added to the original input with the adjusted dimension through an addition layer to achieve residual connection.

[0031] Define a custom layer class for a spiral fully connected layer; during initialization, it receives multiple parameters: output dimension units, maximum offset magnitude amax, number of spirals k, number of cycles t, and input dimension input_units; create two weight variables: weight matrix w and bias term b, which are initialized using Glorot uniform and zeros initializers respectively, and set to trainable.

[0032] The forward propagation process of the spiral fully connected layer is as follows: there are k independent fully connected sub-layers, each containing a set of weight matrices and bias terms; for each input sample, it is multiplied by the weight matrices of the k sub-layers respectively, and the corresponding bias terms are added to obtain k spiral outputs; the outputs are stored in a list outputs; the average of the spiral outputs in the outputs list is taken to obtain avg_out; then, a weight vector is generated through a fully connected layer with a linear activation function; the generated weight vector is extended by one dimension to match the dimension of the outputs tensor, and the two are multiplied element-wise using a broadcast mechanism, and finally summed along the spiral axis to obtain the final spiral fully connected layer output;

[0033] Define a function to create the model. The input is a tensor of root trait features. The model consists of a 128-unit fully connected layer with ReLU activation; a 128-unit residual attention block; a 64-unit residual attention block; a spiral fully connected layer with parameters of units=128, amax=3, k=2, t=8, and input_units=64; and finally, a 2-unit fully connected layer with linear activation as the output.

[0034] This invention also provides a root trait analysis system for *Juglans regia*, the system comprising:

[0035] The database construction module is used to construct an elevation gradient database of the distribution of *Juglans regia* by applying UAV remote sensing monitoring and geographic information system positioning.

[0036] Root sample collection module, used to collect root samples of walnut trees;

[0037] The root trait feature extraction module is used to perform three-dimensional imaging of the collected root samples and quantify the root trait features.

[0038] The walnut tree model building module is used to construct a predictive model of the walnut tree's resource acquisition capacity, and to obtain the root system's absorption rate of water and nutrients.

[0039] Optionally, the database construction module specifically includes:

[0040] The image acquisition submodule is used to acquire high-resolution images of the distribution area of ​​*Juglans regia* using sensors mounted on a drone.

[0041] The image processing submodule is used to process and analyze the acquired images, specifically including:

[0042] Image processing algorithms are used to classify images and identify the canopy coverage area of ​​walnut trees; population density is calculated by combining image resolution and canopy size; digital surface models and digital elevation models are generated using UAV imagery to analyze topographic features; soil type and land use are identified by analyzing images through spectral analysis, and thermal infrared data is used to help determine soil moisture and surface temperature distribution.

[0043] The remote sensing interpretation result generation submodule is used to establish an altitude gradient database based on the UAV remote sensing interpretation results.

[0044] Optionally, the step of classifying the image using an image processing algorithm to identify the canopy coverage area of ​​the *Juglans regia* tree; and calculating the population density by combining the image resolution and canopy size, specifically includes:

[0045] The acquired images undergo data preprocessing, including radiometric and geometric correction.

[0046] Feature extraction is performed on the image, including extracting spectral features using reflectance information from different bands in the image; extracting texture features using the gray-level co-occurrence matrix; and extracting shape features using edge detection and morphological processing methods.

[0047] Based on the extracted features, the K-means clustering algorithm is used to classify the images, that is, the image pixels are automatically divided into several categories according to the similarity of spectral and texture features, and then the canopy area of ​​the walnut tree is identified by expert knowledge or posterior probability.

[0048] After identifying the bare canopy region of the walnut tree, canopy segmentation and boundary refinement are performed to accurately define the bare canopy region, specifically including:

[0049] The classification results are converted into a binary image using a threshold segmentation method, where the white area represents the crown of a walnut tree and the black area represents other ground features or background.

[0050] Perform dilation, erosion, opening, and closing operations on the image to optimize the canopy mask;

[0051] Calculate the population density of *Juglans regia*.

[0052] Optionally, the root trait feature extraction module specifically includes:

[0053] The tomographic image generation submodule is used to first clean the collected root samples to remove attached soil and other impurities; then fix them with a fixative to maintain the root morphology; place the fixed root samples in a resin box and fill it with silicone to ensure that the roots are evenly dispersed and fold-free in the container; place the encapsulated root samples into a CT scanner, set the scanning parameters; run the CT scanning program, and the device will emit X-rays from multiple angles and receive transmission signals to acquire a series of tomographic images;

[0054] The tomographic image processing submodule is used to remove random noise generated during the scanning process using ImageJ image processing software through filtering algorithms; improve image contrast and signal-to-noise ratio using grayscale adjustment; and ensure precise spatial alignment of all slice images through feature matching.

[0055] The 3D root system property generation submodule is used to convert each 2D slice into a thin layer in 3D space using Amira 3D reconstruction software to form a 3D volume dataset; it sets a threshold based on the grayscale difference between the root system and the background to divide the voxel dataset into root system and non-root system parts; it identifies and connects adjacent voxels to form a continuous root system structure; and it generates a triangular mesh model of the root system surface.

[0056] The root trait feature acquisition submodule is used to automatically identify and separate individual root systems based on the grayscale difference between the root system and the background using a threshold segmentation method; and to calculate root trait features.

[0057] Optionally, the light walnut model construction module specifically includes:

[0058] The Residual Attention Block Definition submodule is used to define custom layer classes for residual attention blocks. This layer includes the unit number setting (units), i.e., the output dimension; two fully connected layers with ReLU activation function; an attention mechanism layer; an addition layer for residual connections; and a projection layer.

[0059] The residual attention block execution submodule is used to perform forward propagation of the input through two fully connected layers and apply the ReLU activation function; then, the attention mechanism layer is used to calculate the attention weights; then, it checks whether the input dimension matches the output dimension. If they do not match, the input dimension is adjusted through the projection layer; finally, the result processed by the attention mechanism is added to the original input with the adjusted dimension through the addition layer to realize the residual connection.

[0060] The Spiral Fully Connected Layer Definition submodule is used to define custom layer classes for spiral fully connected layers. During initialization, it receives multiple parameters: output dimension units, maximum offset magnitude amax, number of spirals k, number of cycles t, and input dimension input_units. Two weight variables are created: weight matrix w and bias term b, which are initialized using Glorot uniform and zeros initializers respectively, and set to trainable.

[0061] The spiral fully connected layer execution submodule is used for the forward propagation process of the spiral fully connected layer, which consists of k independent fully connected sublayers, each containing a set of weight matrices and bias terms. For each input sample, it is multiplied by the weight matrices of the k sublayers and the corresponding bias terms are added to obtain k spiral outputs. The outputs are stored in a list called outputs. The average of the spiral outputs in the outputs list is taken to obtain avg_out. Then, a weight vector is generated through a fully connected layer with a linear activation function. The generated weight vector is expanded by one dimension to match the dimension of the outputs tensor, and the two are multiplied element-wise using a broadcast mechanism. Finally, the summation is performed along the spiral axis to obtain the final spiral fully connected layer output.

[0062] The Model Function Definition submodule defines the function that creates the model. The input is a tensor of root trait features. The model consists of a 128-unit ReLU activated fully connected layer; a 128-unit residual attention block; a 64-unit residual attention block; a spiral fully connected layer with parameters of units=128, amax=3, k=2, t=8, and input_units=64; and finally, a 2-unit linearly activated fully connected layer as the output.

[0063] Compared with the prior art, the present invention has the following advantages:

[0064] This invention utilizes UAV remote sensing technology to remotely, extensively, and efficiently acquire information on the distribution of *Juglans regia* roots and their environment without on-site excavation, significantly reducing manpower and avoiding direct interference with the root system. Through hyperspectral and thermal infrared imaging, image processing, and analysis, the invention precisely extracts the canopy boundaries of *Juglans regia* trees and quantifies root characteristics. Combined with three-dimensional imaging technology, it accurately reflects the morphology and distribution of the root system in three-dimensional space, improving data accuracy and reliability. By integrating multi-source information such as altitude gradient, soil type, topography, soil moisture, and surface temperature, a comprehensive database of *Juglans regia* root growth environment is constructed, contributing to a deeper understanding of the tree's adaptation mechanisms to its environment. Furthermore, a neural network model is used to predict the root system's absorption rate of water and nutrients, enabling a quantitative assessment of *Juglans regia*'s resource acquisition capacity and providing a scientific basis for optimizing irrigation and fertilization strategies and improving the yield and quality of *Juglans regia*. Attached Figure Description

[0065] Figure 1 This is a flowchart of a method for analyzing the root traits of *Juglans regia* according to the present invention;

[0066] Figure 2 This is a structural diagram of a root trait analysis system for *Juglans regia* according to the present invention. Detailed Implementation

[0067] The present invention will be further described below with reference to specific implementation examples and accompanying drawings, but the present invention is not limited to these embodiments.

[0068] Example 1

[0069] like Figure 1 As shown, this invention discloses a method for analyzing the root traits of *Juglans regia*, the method comprising:

[0070] Step S1: Use UAV remote sensing monitoring and Geographic Information System (GIS) positioning to construct an altitude gradient database of the distribution of walnut trees.

[0071] Step S2: Collect root samples of *Juglans regia*.

[0072] Step S3: Perform three-dimensional imaging on the collected root samples to quantify the root characteristics.

[0073] Step S4: Construct a predictive model for the resource acquisition capacity of walnut trees to obtain the absorption rate of water and nutrients by the root system.

[0074] The following is a detailed discussion of each step:

[0075] Step S1: Use UAV remote sensing monitoring and Geographic Information System (GIS) positioning to construct an altitude gradient database of the distribution of walnut trees.

[0076] Step S1 specifically includes:

[0077] High-resolution imagery of the walnut distribution area was acquired using drones equipped with multispectral, hyperspectral, or thermal infrared sensors, specifically including:

[0078] The drone is equipped with multispectral, hyperspectral, or thermal infrared sensors. For example, a multispectral camera records visible light and near-infrared bands, a hyperspectral camera provides information on dozens or even hundreds of narrow bands, and a thermal infrared sensor measures heat emissions from the ground. The drone flies over the distribution area of ​​the walnut tree according to a preset route. The sensors capture ground image data in real time, covering the entire study area. The image resolution is high enough to clearly identify the walnut tree canopy and other ground features.

[0079] In this embodiment, parameters such as flight altitude, speed, and heading overlap are set according to actual needs and UAV performance to ensure coverage and image quality. The UAV is commanded to perform flight missions through a ground control station or autonomous flight program. Multispectral, hyperspectral, and thermal infrared sensors capture electromagnetic radiation in different bands. Multispectral data reflects the surface reflectance spectral characteristics and can be used for vegetation classification and health assessment. Hyperspectral data provides more refined spectral information, which helps in the accurate identification of soil and vegetation types. Thermal infrared data reveals the surface temperature distribution, which is closely related to environmental factors such as vegetation cover and soil moisture.

[0080] The acquired images undergo data processing and analysis, specifically including:

[0081] Multispectral or hyperspectral images are classified using image processing algorithms (such as NDVI, SVM, deep learning, etc.) to identify the canopy coverage area of ​​*Juglans regia* trees. Combining image resolution and canopy size, the number of *Juglans regia* trees per unit area is calculated to estimate the population density. The steps are as follows:

[0082] First, data preprocessing is performed, including radiometric correction and geometric correction, to eliminate errors and ensure data consistency. If multispectral data is used, it can be fused with high-resolution panchromatic imagery using super-resolution Bayesian method to improve spatial resolution and facilitate more precise identification of tree canopies.

[0083] Secondly, feature extraction is performed from the images. Reflectance information of different bands in multispectral or hyperspectral images is used, such as calculating vegetation indices (e.g., NDVI) or extracting spectral features using spectral angle mapping. Texture features of the images, such as roughness and contrast, are extracted using gray-level co-occurrence matrix. Shape features of the images, such as the area and perimeter of tree canopies, are extracted using edge detection, morphological processing, and other methods.

[0084] Then, based on the extracted features, the K-means clustering algorithm is used to classify the image, that is, the image pixels are automatically divided into several categories according to the similarity of spectral and texture features, and then the canopy area of ​​the walnut tree is identified by expert knowledge or posterior probability; a supervised classification method can also be used, that is, sample data of the canopy of the walnut tree and other ground features are collected and labeled; then, the SVM algorithm is used to classify the hyperspectral image and identify the canopy area of ​​the walnut tree.

[0085] Next, after identifying the bare walnut tree canopy area, canopy segmentation and boundary refinement are performed to accurately define the bare walnut tree canopy area. The operation is as follows:

[0086] The classification results are converted into a binary image using a threshold segmentation method, where white areas represent the crown of a walnut tree and black areas represent other land features or background. The specific steps are as follows:

[0087] The choice of threshold usually depends on histogram analysis of the classification results; for binary classification problems (white walnut tree canopy and non-canopy), the boundary between the distribution of pixel values ​​of the two classes can be found on the histogram, or the optimal threshold can be found automatically using Otsu's method, or a fixed threshold can be set according to the actual situation and experience.

[0088] A threshold is applied to the classification result image, performing a threshold judgment for each pixel. Pixels with classification values ​​greater than or equal to the threshold are marked as 1 (representing a bare walnut tree canopy), and the remaining pixels are marked as 0 (representing a non-canopy). This generates a binary image (mask) where white areas represent bare walnut tree canopies and black areas represent other ground features or the background.

[0089] Morphological operations, while thresholding segmentation can initially separate the canopy region of the walnut tree, may have some problems, such as noise patches, irregular or broken canopy boundaries, etc. Morphological operations use structuring elements (such as squares, disks, etc.) to perform dilation, erosion, opening, and closing operations on the image to further optimize the canopy mask. The specific steps are as follows:

[0090] To eliminate small noise patches, an opening operation is used, which involves erosion followed by dilation. Small noise patches are completely eliminated during the erosion stage, while larger real canopy areas are restored to their original state during the dilation stage because there are enough similar pixels around them. This effectively removes isolated small noise patches while maintaining the integrity of the canopy area.

[0091] To smooth canopy boundaries, a closing operation is used for broken or discontinuous canopy boundaries, which involves first dilating and then eroding. The dilating operation connects adjacent canopy pixels, allowing broken boundaries to be joined. The subsequent erosion operation eliminates excessive edge effects caused by dilution, preserving the overall shape of the canopy region. After the closing operation, the canopy boundaries will be smoother and more coherent.

[0092] An optional operation is skeleton extraction. Skeleton extraction can be performed to further simplify the canopy structure and accurately depict its skeletal morphology. Common methods include thinning or median transformation. These methods gradually remove non-skeleton pixels within the canopy, retaining only the "skeleton" pixels representing the canopy outline and branch structure. Skeleton extraction is helpful for subsequent canopy shape analysis or individual tree identification.

[0093] After threshold segmentation and morphological operations, the preliminary canopy mask was refined, forming a precise walnut canopy region, providing accurate basic data for subsequent population density estimation.

[0094] Finally, the population density of *Juglans regia* was calculated, as follows:

[0095] For canopy area calculation, the pixel area of ​​each individual canopy region is calculated; taking into account image resolution (e.g., meters per pixel), the pixel area is converted into the actual area (square meters).

[0096] Population density is calculated by using the total area of ​​the study area as the denominator and the sum of the canopy areas of all walnut trees as the numerator. The number of walnut trees per unit area is then calculated, which is the population density (trees / hectare or other units). Alternatively, the average canopy area of ​​walnut trees can be obtained through field measurements or by referring to existing data. The population density of walnut trees can be obtained by dividing the total canopy area by the canopy area of ​​a single tree.

[0097] The following steps are taken to generate a digital surface model (DSM) and a digital elevation model (DEM) from UAV imagery for analyzing terrain features:

[0098] The DSM generation process is as follows:

[0099] Stereo matching utilizes the principles of stereo photogrammetry to register adjacent or overlapping UAV image pairs. Correlation algorithms (such as SIFT, SURF, ORB feature matching or dense bundle method) are used to find corresponding points and construct stereo image pairs. Then, dense matching is performed on the stereo image pairs to estimate the disparity between each pixel, that is, the difference in the projected position of the same ground point on different images.

[0100] Three-dimensional reconstruction, based on the disparity information obtained by stereo matching, uses a multi-view geometry algorithm (BundleAdjustment) to perform spatial triangulation, converting the pixel coordinates on the two-dimensional image into three-dimensional spatial coordinates, and constructing a three-dimensional point cloud model of the study area.

[0101] The DSM generation process involves filtering and denoising the 3D point cloud (using the RANSAC algorithm to remove outliers) and smoothing it (using moving least squares). Then, interpolation methods (Triangulated Interpolation (TIN) and Inverse Distance Weighted Interpolation (IDW)) are used to convert the discrete point cloud data into continuous grid data, generating a digital surface model. The DSM contains height information of all objects on the surface (including vegetation, buildings, etc.).

[0102] The DEM generation process is as follows:

[0103] By using land cover classification methods, non-topographic features such as vegetation and buildings are removed from the DSM to obtain the DEM.

[0104] For DEM generation, after removing vegetation and buildings, interpolation methods (Triangulated Interpolation (TIN) and Inverse Distance Weighted Interpolation (IDW)) are used to fill any blank areas to ensure the continuity and integrity of the DEM. The DEM only reflects the undulation of the terrain itself and does not include the height information of surface attachments.

[0105] Topographic and geomorphological feature analysis is performed as follows:

[0106] Slope calculation is performed by using the first-order partial derivative of the DEM to calculate the gradient, obtaining the slope value of each grid point, expressed as a percentage or degree; the slope reflects the rate of change of the terrain surface in the vertical direction.

[0107] Slope aspect calculation is performed by calculating the slope direction (i.e. gradient direction) or the main direction of the local terrain to obtain the slope aspect of each grid point. It is usually expressed in degrees, and the value range is usually from 0° to 360° (0° or 360° for north, 90° for east, 180° for south, and 270° for west).

[0108] Topographic index calculation: Based on the DEM, various topographic indices are calculated, such as topographic roughness (RMSEz), topographic relief (SDI), topographic curvature (mean curvature, total curvature), topographic azimuth (Sky View Factor, SVF), and topographic moisture index (TWI). These indices can reflect geomorphic features from different perspectives, such as the complexity of the terrain, hydrological characteristics, and lighting conditions.

[0109] In this embodiment, RMSEz represents the root mean square error of elevation. In terrain analysis, it is often used to quantify the overall undulation of the terrain surface. It is calculated by taking the square root of the average of the sum of squares of the differences in elevation values ​​between each grid point on the DEM and its neighboring grid points, and then representing the average fluctuation of the terrain. A higher RMSEz value indicates a more complex and rugged terrain. SDI represents the vertical variation of the terrain. Undulation usually refers to the difference in elevation between the highest and lowest points in a certain area, used to intuitively reflect the overall elevation difference of the terrain. Another common method of representing undulation is relative undulation, which is the undulation divided by the average elevation of the area, used to standardize comparisons between areas of different scales. Terrain curvature is divided into two types: mean curvature and total curvature. They are based on the concept of differential geometry and are used to describe the degree of local curvature of the terrain surface. Mean curvature is a comprehensive measure of the unevenness of a local terrain. It considers the average curvature of the terrain in two orthogonal directions, reflecting the rate of change in the vertical direction per unit area of ​​the terrain surface. Total curvature further distinguishes whether the terrain tends to form peaks (positive curvature), valleys (negative curvature), or flat areas (zero curvature), including information on the bending direction of the terrain at a specific location. Curvature parameters help to understand the fine structure and landform of the terrain. SVF, or Sky Visibility Factor, is a dimensionless parameter used to measure the proportion of the sky visible above a specific location. It reflects the influence of terrain on factors such as solar radiation, wind speed distribution, and visibility obstruction. A larger SVF value indicates a more open terrain around the point and higher sky visibility; conversely, a smaller value indicates a more enclosed terrain and a more significant obstruction effect. TWI represents an index of the impact of topography on water retention and drainage capacity. It is based on the principle that topographic slope and aspect affect the convergence or dispersion of water flow. It quantifies the impact of topography on surface water flow and soil moisture distribution by calculating the ratio of topographic slope to cumulative slope along the direction of water flow. A higher TWI value indicates that the topography is more conducive to water retention, and there may be higher soil moisture and potential waterlogging risk. A low TWI value corresponds to a well-drained area with lower soil moisture.

[0110] The following steps are taken to analyze hyperspectral images using spectral analysis to identify soil type and land use, and combine this with thermal infrared data to help determine soil moisture and surface temperature distribution:

[0111] Spectral feature analysis is performed as follows:

[0112] Principal component analysis (PCA) reduces the dimensionality of collected image data and extracts a few principal components that reflect the main trends in data variation. Principal components can often separate the main spectral differences in soil type and land use, simplifying subsequent classification tasks and being used for visualization and feature selection.

[0113] Spectral Angle Mapping (SAM) calculates the spectral angle between each pixel in the study area and a reference spectrum of a known soil type or land use type. This angle measures the similarity between the two spectral curves. SAM helps to quickly identify potential soil or land use categories and can also serve as input features for subsequent classification.

[0114] Spectral classification, based on measured samples (training set) of known soil and land use types, establishes a model between spectral features and category labels using machine learning algorithms (support vector machine, random forest, neural network, etc.). The model is then applied to the entire study area to classify unknown pixels, as follows:

[0115] Feature selection: Based on PCA results, empirical knowledge, or spectral diagnostic characteristics (absorption peaks or reflection valleys of specific soil types in specific bands), the most discriminative bands or principal components are selected as classification features.

[0116] Model training involves using training set data to adjust model parameters and optimize classification performance; evaluation metrics include overall accuracy, user accuracy, producer accuracy, Kappa coefficient, etc.

[0117] The classification application applies the trained model to all pixels in the study area, outputting a classification map of soil type or land use type.

[0118] Thermal infrared data-assisted analysis is performed as follows:

[0119] Soil moisture analysis uses thermal infrared band data (TIR band of Landsat TM / ETM+ / OLI or thermal band of MODIS) combined with vegetation index NDVI and soil emissivity model SSEBop to estimate the distribution of water content on the soil surface. Wet soil and dry soil exhibit different radiation characteristics in the thermal infrared band, which can be used to infer the difference in soil moisture.

[0120] Surface temperature analysis can be performed by directly reading surface temperature data provided by the thermal infrared band or by inversion through a radiative transfer model. Surface temperature is closely related to soil type, land cover (bare land, vegetation, water bodies, etc.) and soil moisture, and can help identify land use types (farmland, forest land, wetlands, etc.) and assess soil moisture status.

[0121] Based on the results of UAV remote sensing interpretation, an elevation gradient database was established, specifically including:

[0122] The steps for spatial data entry and editing are as follows:

[0123] Multispectral, hyperspectral, and thermal infrared image data obtained from UAV remote sensing monitoring, along with corresponding metadata (shooting date, flight altitude, sensor parameters, etc.), are converted into a GIS-compatible format (GeoTIFF). Distribution maps of *Juglans regia* (vector polygons or point layers representing the canopy coverage area or individual tree location), vegetation cover raster data, and digital elevation models (DEMs) obtained from UAV image interpretation are imported into the GIS platform. Field survey records, including GPS coordinates, altitude, and ecological environment descriptions, are compiled and converted into a GIS-supported point data format (Shapefile or GeoJSON), with each point representing a field survey location.

[0124] Create or open a project file in GIS software and load the above types of data into the workspace; for vector data (walnut distribution layer, field survey point layer), perform necessary editing, such as topology checks, attribute field improvements, and error data corrections; for raster data (such as vegetation cover, DEM), check whether its spatial reference system (SRS) is consistent with the vector data, and perform coordinate system conversion if necessary.

[0125] Spatial analysis, the steps are as follows:

[0126] Buffer analysis selects individual trees or covered areas in the distribution layer of *Juglans regia* as the analysis object and creates a buffer zone of a certain distance (such as canopy diameter or ecological impact radius). Through buffer analysis, the spatial impact range around *Juglans regia* is identified, providing a basis for studying its relationship with other ecological elements (such as soil type and topography).

[0127] Interpolation analysis, based on the elevation data of the field survey points, uses Kriging and IDW interpolation methods to generate an elevation surface model of the entire study area; through interpolation analysis, discrete elevation observations are extended to a continuous spatial distribution, so as to facilitate the analysis of the relationship between the distribution of walnut trees and elevation gradient.

[0128] The steps for distribution range and density analysis are as follows:

[0129] By using the overlay analysis function of GIS, the buffer analysis results are combined with the elevation surface model to count the distribution area of ​​*Juglans regia* in different elevation intervals; the number of *Juglans regia* distribution points or covered areas in each elevation gradient is calculated, and the density of *Juglans regia* per unit area is estimated by combining the buffer area.

[0130] Data visualization utilizes the symbolization, color mapping, and layer overlay functions of GIS software to present the analysis results in an intuitive and easy-to-understand map format. The steps are as follows:

[0131] The elevation gradient distribution map is created by using an elevation surface model obtained through interpolation analysis to create a layered and colored topographic map, with the color depth representing elevation. The walnut distribution layer is then overlaid on the topographic map, and the distribution at different elevation gradients is distinguished by the size, color, or pattern of the symbols.

[0132] To create a thematic map of the distribution of *Juglans regia*, the thematic layer is created based on the analysis results of the distribution range and density. Different colors or gradient bands are used to represent the distribution density of *Juglans regia* at different altitude gradients. Map elements such as legends, scale bars, and compasses are added to make the thematic map easy to understand and interpret.

[0133] A suggested layout map of root sampling points was generated by considering factors such as the distribution density of *Juglans regia*, altitude gradient, and topography, as well as the representativeness and feasibility of sampling. The weighted nearest neighbor distance method in GIS was used to design the layout scheme of root sampling points. The recommended sampling point locations were marked on the map to generate a suggested layout map of root sampling points.

[0134] The steps to establish an elevation gradient database are as follows:

[0135] The database structure is designed according to research needs and data types, such as including data tables for walnut distribution points, root sampling points, altitude, soil type, and ecological environment; the processed spatial data and attribute data are imported into the database, and the relationships between the data are established; data backup, updates, and maintenance are performed to ensure the integrity and reliability of the database.

[0136] In this embodiment, based on the elevation gradient distribution map, vegetation cover map, and soil type distribution map obtained from UAV remote sensing and GIS analysis, representative elevation gradient and soil type areas are selected for root sampling to ensure that the sampling data can reflect the differences in root characteristics of *Juglans regia* under different environmental conditions. Understanding the topographic features helps to select the most suitable non-destructive root excavation technology. For example, air shovels may be more suitable in steep and loose soil areas, while trenching or resistance probe methods may be more suitable in flat and compact soil areas. These environmental data (such as elevation, soil type, soil moisture, and surface temperature) will be used together with the root trait data obtained later for correlation analysis to explore the impact of environmental factors on the root system of *Juglans regia*.

[0137] Step S2: Collect root samples of *Juglans regia*.

[0138] Step S2 specifically includes:

[0139] After establishing the walnut elevation gradient database, representative sampling points at different elevation gradients were selected based on the information in the database, and root sampling was carried out by combining non-destructive root excavation technology and micro-root window technology.

[0140] Sample selection specifically includes:

[0141] Select sample points of *Juglans regia* at different altitudes to ensure that the altitude distribution of the sample points is representative and covers the altitude gradient range of the entire study area. In addition to altitude, environmental factors such as slope, aspect, soil type, and sunlight should also be considered. Select sample points with different environmental characteristics to study the impact of environmental factors on the root system of *Juglans regia*. Select individual *Juglans regia* trees of different ages, vigor, and crown sizes to study the impact of individual differences on the root system.

[0142] Air shovels are suitable for soils with relatively loose textures and shallow root systems, and specifically include:

[0143] Near walnut trees at different altitude gradients, select representative locations as root exposure points, considering factors such as the canopy projection range and the possible direction of root extension. Clear debris from the work area to ensure the safety of air shovel operation; if necessary, set up temporary support structures to protect the tree. Connect the air shovel to the high-pressure air compressor, ensuring the air pressure is stable within the recommended range (usually 40-80 psi). Start the high-pressure airflow equipment, hold the air shovel, aim it at the selected soil area, and blow away the soil layer by layer according to the root distribution depth and soil type, controlling the airflow intensity to avoid impact damage to the roots. As the soil is blown away, the roots gradually become visible; observe and record information such as root distribution, branch morphology, and thickness changes, taking photos or videos if necessary. After completing root observation and recording, gently backfill the blown-away soil, restoring the original state as much as possible to minimize the impact on the root growth environment.

[0144] The trenching method is suitable for situations requiring the acquisition of root samples over a large area, specifically including:

[0145] Based on the canopy projection range of the walnut tree and the expected root distribution characteristics, design the location, length, width, and depth of the excavated trench. Typically, the edge of the canopy or one side of the main growth direction is chosen. Using tools such as shovels or small excavators, the trench is excavated along the design line, keeping the trench walls smooth to avoid mechanical damage to the roots. If large roots are encountered, the trench shape is adjusted appropriately to avoid or bypass them. The soil in the trench is cleared, exposing part of the root system. Detailed information on the root distribution, growth direction, branching pattern, and thickness variations within the trench is recorded, along with relevant soil parameters. After observation and recording, the roots are carefully covered back in their original location, the soil is backfilled and compacted, restoring the original soil structure and mulch of the trench area as much as possible to minimize the impact on plant growth.

[0146] The resistance probe method is suitable for detecting the distribution range of root systems without direct digging. Specifically, it includes:

[0147] Within the projection range of the walnut tree canopy, resistance probes are inserted into the soil to a predetermined depth according to a pre-set grid or random points. The probe spacing is determined based on the required detection accuracy and instrument characteristics. The resistance probes are connected to a resistivity measuring instrument, and measurements are taken according to the instrument's operating procedures. Soil resistivity data between each probe are recorded. Based on the correlation between soil resistivity and root distribution, the data is analyzed to generate a two-dimensional or three-dimensional image of root distribution. Resistivity usually decreases with increasing root density, thereby inferring dense and sparse root areas. In dense root areas initially identified by the resistance probe method, local verification can be performed using an air shovel or trenching method to improve the accuracy of the detection results.

[0148] Micro root window technology specifically includes:

[0149] Select suitable locations around the trunk of the walnut tree (usually below the root collar), pre-drill holes, and install transparent microroot tubes (microroot windows), ensuring there are no gaps between the tube opening and the trunk, and that they are well-sealed to prevent moisture ingress. Using a microroot window scanner or manual camera, periodically photograph the inside of the microroot windows (e.g., according to the growing season, rainfall events, or other important time points) to record the dynamic changes in root growth, development, death, and decay within the transparent tubes. Continuously record the root growth data of the walnut tree within the microroot windows at different altitude gradients to compare and analyze the impact of altitude gradients on root growth.

[0150] In this embodiment, by combining the above-mentioned non-destructive root excavation technology and micro-root window technology, representative samples and growth dynamic information of walnut root systems at different altitude gradients can be obtained systematically and accurately without damaging the root structure of walnut. This provides a scientific basis for subsequent root trait research, root ecological function assessment and management decisions.

[0151] Step S3: Perform three-dimensional imaging on the collected root samples to quantify the root characteristics.

[0152] Step S3 specifically includes:

[0153] Using high-resolution X-ray computed tomography (CT) or magnetic resonance imaging (MRI) techniques, combined with root system image analysis software, three-dimensional reconstruction and trait quantification of the root system are performed.

[0154] Sample preparation and scanning, specifically including:

[0155] The steps for processing root samples using CT or MRI are as follows:

[0156] For CT scans, the collected root samples are first cleaned to remove attached soil and other impurities. Then, they are fixed with a non-toxic and harmless fixative (such as formalin or glutaraldehyde) to maintain the root morphology. The fixed root samples are placed in a transparent, non-metallic container with high X-ray penetration (such as a polycarbonate tube or resin box) and filled with a medium with low X-ray absorption (such as silica gel or paraffin) to ensure that the roots are evenly dispersed and without folds within the container. The encapsulated root samples are then placed in a CT scanner. Scanning parameters, such as voltage, current, exposure time, rotation step size, and reconstruction matrix size, are set according to the root size, material properties, and required resolution. The CT scan program is then run, and the device emits X-rays from multiple angles and receives the transmitted signals to acquire a series of tomographic images.

[0157] For MRI, no fixation or cleaning is required. Fresh or refrigerated root samples can be directly placed in a dedicated MR-compatible container, ensuring the sample remains moist to prevent signal attenuation due to drying. On the MRI equipment, select an appropriate imaging sequence (such as T1, T2, or proton density-weighted sequences), and adjust parameters such as magnetic field strength, echo time, repetition time, flip angle, and matrix size according to the root characteristics. Place the container containing the root sample into the MRI scanner and scan according to the preset parameters to generate multiple two-dimensional slice images.

[0158] 3D reconstruction and visualization, specifically including:

[0159] Data acquisition involves two methods: CT scans, which use X-rays to perform tomographic scanning of plant roots and obtain a series of two-dimensional slice images, each slice representing the projection of the root system at a certain depth; and MRI scans living plant roots non-invasively using strong magnetic fields and radio waves, which also produce a series of two-dimensional slice images.

[0160] Data preprocessing uses ImageJ and Avizo image processing software to remove random noise, stripe noise, or other interference factors generated during scanning through filtering algorithms (such as smoothing filtering, median filtering, adaptive filtering, etc.) to improve image quality. Grayscale adjustment, histogram equalization, and edge enhancement are used to improve image contrast and signal-to-noise ratio to better identify the boundary between the root system and the background (such as soil or culture medium). Feature matching and rigid or non-rigid transformation methods are used to ensure that all slice images are spatially precisely aligned to eliminate slight displacement or deformation that may exist during scanning.

[0161] 3D reconstruction relies directly on preprocessed, high-quality, precisely aligned 2D slices. Using Amira and VGstudio Max 3D reconstruction software, each 2D slice is converted into a thin layer (voxel) in 3D space, forming a 3D volume dataset. A threshold is set based on the grayscale difference between the root system and the background to divide the voxel dataset into root and non-root parts. Adjacent voxels are identified and connected to form a continuous root structure, distinguishing different parts such as the taproot, lateral roots, and root tips. Internal voids caused by threshold settings or the scanning process itself are repaired to ensure the integrity of the root model. A triangular mesh model of the root surface is generated, providing a smooth and continuous 3D geometric representation for subsequent visualization.

[0162] Visualization: Using ParaView and VTK (Visualization Toolkit) 3D visualization software, the generated 3D root system model is loaded and rotated, scaled, and translated at any angle for observation; lighting models, material properties, and color mapping are applied to make the root system model realistic or highlight specific structural features.

[0163] Root system image analysis, specifically including:

[0164] Root system division, the steps are as follows:

[0165] Using threshold segmentation, region growing, and convolutional neural network algorithms, individual root systems are automatically identified and separated based on the grayscale difference between the root system and the background. For complex root structures or cases where automatic segmentation is ineffective, interactive tools are used to assist in marking root boundaries for refined segmentation.

[0166] Trait extraction, the steps are as follows:

[0167] Calculate parameters such as total root length, surface area, volume, diameter distribution, branching angle, and number of root tips; estimate the dry or fresh biomass of the root system by combining root density data; statistically analyze the distribution of root length, area, and volume in each vertical layer (e.g., 0-10cm, 10-20cm, etc.) to generate a stratified distribution map; and analyze the vertical and horizontal distribution characteristics of the root system, such as root depth, root spread, and root concentration, by measuring the position coordinates of the root system in three-dimensional space.

[0168] Step S4: Construct a predictive model for the resource acquisition capacity of walnut trees to obtain the absorption rate of water and nutrients by the root system.

[0169] Step S4 specifically includes:

[0170] Data collection and preprocessing specifically include:

[0171] Data source, steps are as follows:

[0172] The elevation gradient, soil type, and topographic features were obtained from UAV remote sensing monitoring, GIS positioning, and analysis results; root characteristics (including total root length, surface area, volume, diameter distribution, branching angle, number of root tips, etc.) were obtained from high-resolution X-ray computed tomography (CT) or magnetic resonance imaging (MRI) technology and root image analysis software; root distribution patterns (such as root depth, root spread, root concentration, etc.) were based on UAV remote sensing and GIS analysis results.

[0173] Data preprocessing, the steps are as follows:

[0174] Missing value handling involves identifying and filling in missing data, which can be done using methods such as mean filling, median filling, and interpolation.

[0175] Data standardization: Standardize or normalize all input features to ensure fair comparison between features of different magnitudes and effective training of neural networks.

[0176] Target variable processing: Transform the root system’s water and nutrient absorption rate data into a form suitable for neural network model processing, such as normalizing it to the [0,1] interval.

[0177] The dataset is divided into the following steps:

[0178] Random partitioning involves randomly dividing the preprocessed dataset into training, validation, and test sets, typically in proportions of 70%, 20%, and 10%.

[0179] To ensure that the data is evenly distributed, each subset has a similar distribution in terms of elevation gradient, soil type, and topography, so as to reduce the risk of model overfitting.

[0180] Constructing a predictive model for the acquisition capacity of walnut resources, specifically including:

[0181] Import the required libraries. TensorFlow is an open-source library for machine learning and deep learning, which includes the Keras high-level API; it uses the Layer module from tf.keras.layers.

[0182] A custom layer class, `ResidualAttentionBlock`, inheriting from `keras.layers.Layer`, is defined. This layer includes the number of units (i.e., the output dimension), two fully connected (Dense) layers with ReLU activation, an attention layer, an adder layer for residual connections, and a projection layer (Dense) to adjust the consistency between the input and output dimensions. It is represented as: `class ResidualAttentionBlock(layers.Layer): def __init__(self, units, **kwargs) sup` er(ResidualAttentionBlock,self).__init__(**kwargs); self.units=units; self.dense1=layers.Dense(units,activation="relu"); self.dense2= layers.Dense(units, activation="relu"); self.attention=layers.Attention(); self.add=layers.Add(); self.projection=layers.Dense(units).

[0183] The `call()` method defines how to process the input data when the layer is called. First, the input is forward-propagated through two Dense layers and the ReLU activation function is applied. Then, the Attention layer is used to calculate the attention weights. Next, it checks whether the input dimension matches the output dimension (units). If they do not match, the input dimension is adjusted through the projection layer. Finally, the result `x` after the attention mechanism is processed is added to the original input `inputs` with the adjusted dimension through the Add layer to achieve a residual connection, represented as `defcall(self,inputs): x = self.dense1(inputs); x = self.dense2(x); x = self.attention([x,x]); if inputs.shape[-1]! = self.units: inputs = self.projection(inputs) return self.add([x,inputs])`.

[0184] A custom layer class, SpiralFC, is defined, also inheriting from keras.layers.Layer. This layer features spiral weight distribution and accepts multiple parameters during initialization: output dimension (units), maximum offset magnitude (amax), number of spirals (k), number of cycles (t), and input dimension (input_units). This layer creates two weight variables: w (weight matrix) and b (bias term), which are initialized using Glorot uniform and zeros initializers respectively, and are set to trainable, denoted as class. SpiralFC(layers.Layer):def__init__(self,units,amax,k,t,input_units=None,**kwargs):super(SpiralFC,self).__init__(**kwargs); self.units=units; self.amax=amax; self.k=k; self.t=t; self.input_unitsin put_units; self.w=self.add_weight(shape=(self.k, input_units, units), initializer="glorot_uniform", trainable=True); self.b=self.add_weight(shape=(self.k, units), initializer="zeros", trainable=True).

[0185] The `call()` method defines the forward propagation process of the spiral fully connected layer. For each spiral (k spirals), it calculates the product of the input and the corresponding weight matrix, adds a bias term, and stores the result in a list `outputs`. The list is then converted into a tensor of shape `[batch_size, k, units]`. Next, the average value (avg_out) of all spiral outputs is calculated, and a weight vector of shape `[batch_size, k]` is obtained by passing it through a fully connected layer (with a linear activation function). The weight vector is expanded by one dimension to match the dimension of the output tensor, and the weights are multiplied element-wise by the output tensor using a broadcast mechanism. Finally, the summation along the first axis (spiral axis) yields the final spiral fully connected layer output, represented as `defcall(self, inputs): outputs = []; for i in range(self.k): x = tf.mat`. mul(inputs,self.w[i])+self.b[i]; outputs.append(x); outputs=tf.stack(outputs, axis=1); avg_out=tf.reduce_mean(outputs, axis=1); Weights=tf.nn.softmax(layers.D ense(self.k,activation="linear")(avg_out)); Weights=weights[:,:,tf.newaxis]; weighted_outputs=tf.reduce_sum(weights×outputs,axis=1); returnweighted_outputs.

[0186] In this embodiment, there are k independent fully connected sublayers (each corresponding to a spiral), each sublayer containing a set of weight matrices self.w[i] and bias terms self.b[i]. For each input sample, matrix multiplication is performed with the weight matrices of the k sublayers respectively, and the corresponding bias terms are added to obtain k spiral outputs. These outputs are stored in a list outputs. The average value of the spiral outputs in the outputs list is taken to obtain avg_out. Then, a weight vector is generated through a fully connected layer with a linear activation function (the weight vector shape is [batch_size,k]), which reflects the importance of each spiral output. The generated weight vector is extended by one dimension to match the dimension of the outputs tensor, and the two are multiplied element-wise using a broadcast mechanism. Finally, the summation is performed along the first axis (spiral axis) to obtain the final spiral fully connected layer output.

[0187] A function `create_model` is defined to construct a neural network model with a specific structure. The input is a tensor of shape `(input_dim,)`. The model contains the following layers: a 128-unit fully connected layer with ReLU activation; a 128-unit ResidualAttentionBlock; a 64-unit ResidualAttentionBlock; a SpiralFC layer with parameters `units = 128`, `amax = 3`, `k = 2`, `t = 8`, and `input_units = 64`; and finally, a 2-unit fully connected layer with linear activation as the output, denoted as `def`. create_model(input_dim):Inputs=keras.Input(shape=(input_dim,)); x=layers.Dense(128, activation="relu")(inputs); x=ResidualAttentionBlock(128)(x); 4) (x);

[0188] By creating a model instance, compiling it, and displaying the model's structure, we can create a model with an input dimension of 10, represented as model = create_model(input_dim = 10); compile the model using mean squared error loss and the Adam optimizer, represented as model.compile(loss = "mse", optimizer = "adam"); ​​and print a summary of the model structure, represented as model.summary().

[0189] Model training and tuning specifically include:

[0190] Initialize weights and biases using random or predefined methods; perform iterative training using the training set data, calculate the loss function value and the predicted value of the target variable after each training round, and update the model parameters according to the optimizer; set an early stopping mechanism to stop training when the validation set loss no longer decreases significantly after a certain number of iterations to prevent overfitting.

[0191] Model performance can be optimized by adjusting hyperparameters such as model architecture (number of layers, number of nodes), optimizer parameters (learning rate), and regularization coefficients through methods such as grid search, random search, or Bayesian optimization; and model convergence speed and final performance can be improved by using learning rate decay or periodic learning rate scheduling strategies.

[0192] Model evaluation and validation specifically include:

[0193] Performance metrics can be selected as: Mean Squared Error (MSE): assesses the average deviation of the model's predictions of root water and nutrient uptake rates; Root Mean Squared Error (RMSE): provides a sense of the average scale of the deviation from actual values; and Coefficient of Determination (R²). 2 Pearson correlation coefficient: Measured by the correlation between the model's predicted values ​​and the actual values. The closer the value is to 1, the better the model fits.

[0194] After each training cycle, the model performance is evaluated using validation set data, and the hyperparameters are adjusted based on the evaluation results. After the model training is completed, it is independently evaluated using test set data to obtain the model's generalization performance on unseen data.

[0195] The trained and validated neural network model is applied to newly acquired data on elevation gradient, soil type, topography, and root characteristics to predict the absorption rate of water and nutrients by the roots of the walnut tree. As new data accumulates, the model is retrained or fine-tuned periodically to maintain its predictive performance. The model is then encapsulated into an API or embedded into relevant applications for easy access and result display in real-world scenarios.

[0196] Example 2

[0197] like Figure 2 As shown, this invention discloses a root trait analysis system for *Juglans regia*, the system comprising:

[0198] Database construction module 10 is used to construct an altitude gradient database of the distribution of *Platycarpus santalinus* by applying UAV remote sensing monitoring and geographic information system positioning.

[0199] The root sample collection module 20 is used to collect root samples of *Juglans regia*.

[0200] The root trait feature extraction module 30 is used to perform three-dimensional imaging of the collected root samples and quantify the root trait features.

[0201] The 40 module for constructing the walnut model is used to build a predictive model of the walnut resource acquisition capacity and obtain the root system's absorption rate of water and nutrients.

[0202] As an optional implementation, the database construction module 10 of the present invention specifically includes:

[0203] The image acquisition submodule is used to acquire high-resolution images of the walnut distribution area using a drone equipped with sensors.

[0204] The image processing submodule is used to process and analyze the acquired images, specifically including:

[0205] Image processing algorithms are used to classify images and identify the canopy coverage area of ​​walnut trees; population density is calculated by combining image resolution and canopy size; digital surface models and digital elevation models are generated using UAV imagery to analyze topographic features; soil type and land use are identified by analyzing images through spectral analysis, and thermal infrared data is used to help determine soil moisture and surface temperature distribution.

[0206] The remote sensing interpretation result generation submodule is used to establish an altitude gradient database based on the UAV remote sensing interpretation results.

[0207] As an optional implementation, the present invention classifies images using image processing algorithms to identify the canopy coverage area of ​​*Juglans regia* trees; and calculates the population density by combining image resolution and canopy size, specifically including:

[0208] The acquired images undergo data preprocessing, including radiometric and geometric correction.

[0209] Feature extraction is performed on the image by using reflectance information from different bands to extract spectral features; using gray-level co-occurrence matrix to extract texture features; and using edge detection and morphological processing methods to extract shape features.

[0210] Based on the extracted features, the K-means clustering algorithm is used to classify the images, that is, to automatically divide the image pixels into several categories based on the similarity of spectral and texture features, and then to identify the canopy area of ​​the walnut tree through expert knowledge or posterior probability.

[0211] After identifying the canopy region of the bare walnut tree, canopy segmentation and boundary refinement are performed to accurately define the canopy region, specifically including:

[0212] The classification results are converted into a binary image using a threshold segmentation method, where the white area represents the crown of a walnut tree and the black area represents other ground features or background.

[0213] Perform dilation, erosion, opening, and closing operations on the image to optimize the canopy mask.

[0214] Calculate the population density of *Juglans regia*.

[0215] As an optional implementation, the root trait feature extraction module 30 of the present invention specifically includes:

[0216] The tomographic image generation submodule is used to first clean the collected root samples to remove attached soil and other impurities; then fix them with a fixative to maintain the root morphology; place the fixed root samples in a resin box and fill it with silicone to ensure that the roots are evenly dispersed and fold-free within the container; place the encapsulated root samples into a CT scanner, set the scanning parameters; run the CT scanning program, and the device will emit X-rays from multiple angles and receive transmission signals to acquire a series of tomographic images.

[0217] The tomographic image processing submodule is used to remove random noise generated during the scanning process using ImageJ image processing software through filtering algorithms; improve image contrast and signal-to-noise ratio through grayscale adjustment; and ensure that all slice images are spatially precisely aligned through feature matching.

[0218] The 3D root system property generation submodule is used to convert each 2D slice into a thin layer in 3D space using Amira 3D reconstruction software to form a 3D volume dataset; it sets a threshold based on the grayscale difference between the root system and the background to divide the voxel dataset into root system and non-root system parts; it identifies and connects adjacent voxels to form a continuous root system structure; and it generates a triangular mesh model of the root system surface.

[0219] The root trait feature acquisition submodule is used to automatically identify and separate individual root systems based on the grayscale difference between the root system and the background using a threshold segmentation method; and to calculate root trait features.

[0220] As an optional implementation, the walnut model construction module 40 of the present invention specifically includes:

[0221] The Residual Attention Block Definition submodule is used to define custom layer classes for residual attention blocks. This layer includes the number of units, i.e., the output dimension; two fully connected layers with ReLU activation function; an attention mechanism layer; an addition layer for residual connections; and a projection layer.

[0222] The residual attention block execution submodule is used to perform forward propagation of the input through two fully connected layers and apply the ReLU activation function; then, the attention mechanism layer is used to calculate the attention weights; then, it checks whether the input dimension matches the output dimension. If they do not match, the input dimension is adjusted through the projection layer; finally, the result processed by the attention mechanism is added to the original input with the adjusted dimension through the addition layer to realize the residual connection.

[0223] The Spiral Fully Connected Layer Definition submodule is used to define custom layer classes for spiral fully connected layers. During initialization, it receives multiple parameters: output dimension units, maximum offset magnitude amax, number of spirals k, number of cycles t, and input dimension input_units. It creates two weight variables: weight matrix w and bias term b, which are initialized using Glorot uniform and zeros initializers, respectively, and are set to trainable.

[0224] The spiral fully connected layer execution submodule is used for the forward propagation process of the spiral fully connected layer, which consists of k independent fully connected sublayers, each containing a set of weight matrices and bias terms. For each input sample, it is multiplied by the weight matrices of the k sublayers and the corresponding bias terms are added to obtain k spiral outputs. The outputs are stored in a list called outputs. The average of the spiral outputs in the outputs list is taken to obtain avg_out. Then, a weight vector is generated through a fully connected layer with a linear activation function. The generated weight vector is extended by one dimension to match the dimension of the outputs tensor, and the two are multiplied element-wise using a broadcast mechanism. Finally, the summation is performed along the spiral axis to obtain the final spiral fully connected layer output.

[0225] The Model Function Definition submodule defines the function that creates the model. The input is a tensor of root trait features. The model consists of a 128-unit ReLU activated fully connected layer; a 128-unit residual attention block; a 64-unit residual attention block; a spiral fully connected layer with parameters of units=128, amax=3, k=2, t=8, and input_units=64; and finally, a 2-unit linearly activated fully connected layer as the output.

[0226] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for analyzing the root traits of *Juglans regia*, characterized in that, The method includes: Step S1: Using UAV remote sensing monitoring and geographic information system positioning, construct an altitude gradient database of the distribution of *Juglans regia*. Step S2: Collect root samples from *Juglans regia*. Step S3: Perform three-dimensional imaging on the collected root samples to quantify root morphological characteristics; Step S4: Construct a predictive model for the resource acquisition capacity of walnut trees to obtain the absorption rate of water and nutrients by the root system; The process of performing three-dimensional imaging on the collected root samples to quantify root morphological characteristics specifically includes: The collected root samples are first cleaned to remove attached soil and other impurities; then they are fixed with a fixative to maintain the root morphology; the fixed root samples are placed in a resin box and filled with silica gel to ensure that the roots are evenly dispersed and without folds in the container; the sealed root samples are placed in a CT scanner and the scanning parameters are set; the CT scanning program is run, and the device will emit X-rays from multiple angles and receive transmission signals to obtain a series of tomographic images. ImageJ image processing software was used to remove random noise generated during the scanning process through filtering algorithms; grayscale adjustment was used to improve image contrast and signal-to-noise ratio; and feature matching was used to ensure that all slice images were precisely aligned in space. The Amira 3D reconstruction software is used to convert each 2D slice into a thin layer in 3D space to form a 3D volume dataset; a threshold is set according to the grayscale difference between the root system and the background to divide the voxel dataset into root system and non-root system parts; adjacent voxels are identified and connected to form a continuous root system structure; a triangular mesh model of the root system surface is generated. Using a threshold segmentation method, individual root systems are automatically identified and separated based on the grayscale difference between the root system and the background; root trait characteristics are then calculated. The construction of the predictive model for the acquisition capability of walnut resources specifically includes: Define a custom layer class for the residual attention block; this layer contains the unit number setting (units), i.e., the output dimension; two fully connected layers with ReLU activation function; an attention mechanism layer; an addition layer for residual connections; and a projection layer. First, the input is forward-propagated through two fully connected layers, and the ReLU activation function is applied. Then, an attention mechanism layer is used to calculate the attention weights. Next, it is checked whether the input dimension matches the output dimension. If they do not match, the input dimension is adjusted through a projection layer. Finally, the result processed by the attention mechanism is added to the original input with the adjusted dimension through an addition layer to achieve residual connection. Define a custom layer class for a spiral fully connected layer; during initialization, it receives multiple parameters: output dimension units, maximum offset magnitude amax, number of spirals k, number of cycles t, and input dimension input_units; create two weight variables: weight matrix w and bias term b, which are initialized using Glorot uniform and zeros initializers respectively, and set to trainable. The forward propagation process of the spiral fully connected layer is as follows: there are k independent fully connected sub-layers, each containing a set of weight matrices and bias terms; for each input sample, it is multiplied by the weight matrices of the k sub-layers respectively, and the corresponding bias terms are added to obtain k spiral outputs; the outputs are stored in a list outputs; the average of the spiral outputs in the outputs list is taken to obtain avg_out; then, a weight vector is generated through a fully connected layer with a linear activation function; the generated weight vector is extended by one dimension to match the dimension of the outputs tensor, and the two are multiplied element-wise using a broadcast mechanism, and finally summed along the spiral axis to obtain the final spiral fully connected layer output; Define a function to create the model. The input is a tensor of root trait features. The model consists of a 128-unit fully connected layer with ReLU activation; a 128-unit residual attention block; a 64-unit residual attention block; a spiral fully connected layer with parameters of units=128, amax=3, k=2, t=8, and input_units=64; and finally, a 2-unit fully connected layer with linear activation as the output.

2. The method for analyzing the root traits of *Juglans regia* according to claim 1, characterized in that, The application of UAV remote sensing monitoring and geographic information system positioning to construct an elevation gradient database of the distribution of *Juglans regia* specifically includes: High-resolution images of the distribution area of ​​*Juglans regia* were acquired using drones equipped with sensors. The acquired images undergo data processing and analysis, specifically including: Image processing algorithms are used to classify images and identify the canopy coverage area of ​​walnut trees; population density is calculated by combining image resolution and canopy size; digital surface models and digital elevation models are generated using UAV imagery to analyze topographic features; soil type and land use are identified by analyzing images through spectral analysis, and thermal infrared data is used to help determine soil moisture and surface temperature distribution. An altitude gradient database was established based on the results of UAV remote sensing interpretation.

3. The method for analyzing the root traits of *Juglans regia* according to claim 2, characterized in that, The process involves classifying images using image processing algorithms to identify the canopy coverage area of ​​*Juglans regia* trees; and calculating the population density by combining image resolution and canopy size. Specifically, this includes: The acquired images undergo data preprocessing, including radiometric and geometric correction. Feature extraction is performed on the image, including extracting spectral features using reflectance information from different bands in the image; extracting texture features using the gray-level co-occurrence matrix; and extracting shape features using edge detection and morphological processing methods. Based on the extracted features, the K-means clustering algorithm is used to classify the images, that is, the image pixels are automatically divided into several categories according to the similarity of spectral and texture features, and then the canopy area of ​​the walnut tree is identified by expert knowledge or posterior probability. After identifying the bare canopy region of the walnut tree, canopy segmentation and boundary refinement are performed to accurately define the bare canopy region, specifically including: The classification results are converted into a binary image using a threshold segmentation method, where the white area represents the crown of a walnut tree and the black area represents other ground features or background. Perform dilation, erosion, opening, and closing operations on the image to optimize the canopy mask; Calculate the population density of *Juglans regia*.

4. A root trait analysis system for *Juglans regia*, characterized in that, The system includes: The database construction module is used to construct an elevation gradient database of the distribution of *Juglans regia* by applying UAV remote sensing monitoring and geographic information system positioning. Root sample collection module, used to collect root samples of walnut trees; The root trait feature extraction module is used to perform three-dimensional imaging of the collected root samples and quantify the root trait features. The walnut model building module is used to build a predictive model of the walnut resource acquisition capacity and obtain the root system's absorption rate of water and nutrients. The root trait feature extraction module specifically includes: The tomographic image generation submodule is used to first clean the collected root samples to remove attached soil and other impurities; then fix them with a fixative to maintain the root morphology; place the fixed root samples in a resin box and fill it with silicone to ensure that the roots are evenly dispersed and fold-free in the container; place the encapsulated root samples into a CT scanner, set the scanning parameters; run the CT scanning program, and the device will emit X-rays from multiple angles and receive transmission signals to acquire a series of tomographic images; The tomographic image processing submodule is used to remove random noise generated during the scanning process using ImageJ image processing software through filtering algorithms; improve image contrast and signal-to-noise ratio using grayscale adjustment; and ensure precise spatial alignment of all slice images through feature matching. The 3D root system property generation submodule is used to convert each 2D slice into a thin layer in 3D space using Amira 3D reconstruction software to form a 3D volume dataset; it sets a threshold based on the grayscale difference between the root system and the background to divide the voxel dataset into root system and non-root system parts; it identifies and connects adjacent voxels to form a continuous root system structure; and it generates a triangular mesh model of the root system surface. The root trait feature acquisition submodule is used to automatically identify and separate individual root systems based on the grayscale difference between the root system and the background using a threshold segmentation method; and to calculate root trait features. The light walnut model construction module specifically includes: The Residual Attention Block Definition submodule is used to define custom layer classes for residual attention blocks. This layer includes the unit number setting (units), i.e., the output dimension; two fully connected layers with ReLU activation function; an attention mechanism layer; an addition layer for residual connections; and a projection layer. The residual attention block execution submodule is used to perform forward propagation of the input through two fully connected layers and apply the ReLU activation function; then, the attention mechanism layer is used to calculate the attention weights; then, it checks whether the input dimension matches the output dimension. If they do not match, the input dimension is adjusted through the projection layer; finally, the result processed by the attention mechanism is added to the original input with the adjusted dimension through the addition layer to realize the residual connection. The Spiral Fully Connected Layer Definition submodule is used to define custom layer classes for spiral fully connected layers. During initialization, it receives multiple parameters: output dimension units, maximum offset magnitude amax, number of spirals k, number of cycles t, and input dimension input_units. Two weight variables are created: weight matrix w and bias term b, which are initialized using Glorot uniform and zeros initializers respectively, and set to trainable. The spiral fully connected layer execution submodule is used for the forward propagation process of the spiral fully connected layer, which consists of k independent fully connected sublayers, each containing a set of weight matrices and bias terms. For each input sample, it is multiplied by the weight matrices of the k sublayers and the corresponding bias terms are added to obtain k spiral outputs. The outputs are stored in a list called outputs. The average of the spiral outputs in the outputs list is taken to obtain avg_out. Then, a weight vector is generated through a fully connected layer with a linear activation function. The generated weight vector is expanded by one dimension to match the dimension of the outputs tensor, and the two are multiplied element-wise using a broadcast mechanism. Finally, the summation is performed along the spiral axis to obtain the final spiral fully connected layer output. The Model Function Definition submodule defines the function that creates the model. The input is a tensor of root trait features. The model consists of a 128-unit ReLU activated fully connected layer; a 128-unit residual attention block; a 64-unit residual attention block; a spiral fully connected layer with parameters of units=128, amax=3, k=2, t=8, and input_units=64; and finally, a 2-unit linearly activated fully connected layer as the output.

5. The root trait analysis system for *Juglans regia* according to claim 4, characterized in that, The database construction module specifically includes: The image acquisition submodule is used to acquire high-resolution images of the distribution area of ​​*Juglans regia* using sensors mounted on a drone. The image processing submodule is used to process and analyze the acquired images, specifically including: Image processing algorithms are used to classify images and identify the canopy coverage area of ​​walnut trees; population density is calculated by combining image resolution and canopy size; digital surface models and digital elevation models are generated using UAV imagery to analyze topographic features; soil type and land use are identified by analyzing images through spectral analysis, and thermal infrared data is used to help determine soil moisture and surface temperature distribution. The remote sensing interpretation result generation submodule is used to establish an altitude gradient database based on the UAV remote sensing interpretation results.

6. The root trait analysis system for *Juglans regia* according to claim 5, characterized in that, The process involves classifying images using image processing algorithms to identify the canopy coverage area of ​​*Juglans regia* trees; and calculating the population density by combining image resolution and canopy size. Specifically, this includes: The acquired images undergo data preprocessing, including radiometric and geometric correction. Feature extraction is performed on the image, including extracting spectral features using reflectance information from different bands in the image; extracting texture features using the gray-level co-occurrence matrix; and extracting shape features using edge detection and morphological processing methods. Based on the extracted features, the K-means clustering algorithm is used to classify the images, that is, the image pixels are automatically divided into several categories according to the similarity of spectral and texture features, and then the canopy area of ​​the walnut tree is identified by expert knowledge or posterior probability. After identifying the bare canopy region of the walnut tree, canopy segmentation and boundary refinement are performed to accurately define the bare canopy region, specifically including: The classification results are converted into a binary image using a threshold segmentation method, where the white area represents the crown of a walnut tree and the black area represents other ground features or background. Perform dilation, erosion, opening, and closing operations on the image to optimize the canopy mask; Calculate the population density of *Juglans regia*.

Images