AGB and plant height (PH) measurement datasets were collected for rapeseed at the seedling stage (December 28, 2022), bolting stage (February 20, 2023), and early blossoming stage (March 2, 2023). Collection of AGB data: To ensure representative sampling, three plants reflecting the overall growth condition were randomly selected from each plot and placed in sealed plastic bags for transportation to the laboratory. Once separated, the stems and leaves were washed with running water and placed in an oven at 105°C for 1 hour, followed by drying at 80°C for more than 48 hours until a stable weight was reached. The stems and leaves were then weighed using a high-precision balance (accuracy 0.001 g), and the total weight of the samples was calculated. AGB was determined based on the population density and the total weight of the samples. Collection of PH data: Four representative rapeseed plants were selected from each plot, and the distance from the base to the tip of the leaf was measured using a ruler. The average of the four measurements was calculated as the PH for the rapeseed in the plot.

A six-rotor UAV (Matrice 600 Pro, DJI, China) equipped with a Resonon PIKA CX imaging spectrometer (Resonon, USA) was used to capture hyperspectral data on sampling days. The hyperspectral sensor covers a spectral range of 400–1000 nm with 2.2 nm spectral resolution across 150 channels. To ensure data consistency, all flights were conducted from the same takeoff point between 12:00 and 13:30 under clear, windless conditions. Each flight followed a consistent route for all growth stages, flying at an altitude of 50 m (with a 6 m transect width) and a speed of 2 m/s. An 80% overlap was maintained in both forward and side directions. Radiometric calibration was performed before each flight using black and white reference panels. Hyperspectral images were acquired at four key growth stages: bare-soil (October 20), seedling, bolting, and early blossoming, with a spatial resolution of 2.5 cm.

RGB images were also captured on the same days and at the same altitude using a quad-rotor UAV (Dajiang Yu2, DJI, China). These images, along with digital surface models (DSM), were processed using Pix4D software. Ground control points (GCPs) were used to align the RGB photos for spatial consistency across growth stages. The images were processed into high-density point clouds, grids, and textures using structure-from-motion (SfM) algorithms, and mosaicked into digital orthophotos for each growth stage.

The UAV hyperspectral images were preprocessed in three steps: (1) Correction and mosaicking: The data were corrected and mosaicked using Spectronon software (USA) for both radiometric and geographic corrections. (2) Image stitching: ENVI software (Harris Exelis, USA) was used to stitch the images within the specified navigation area, incorporating position data from GCPs to minimize correction error. This resulted in a new hyperspectral image. (3) Extracting rapeseed canopy reflectivity: Rapeseed canopy reflectance curves were extracted from the images. Using ArcGIS software, different maximum area vectors were delineated, and vector data were assigned numbers based on samples. The average spectral reflectance of each region of interest was then extracted using the interactive data language (IDL) in ENVI software. These values were considered the spectral reflectance of the rapeseed canopy in different plots. The images and the average values for each plot were used as the rapeseed canopy spectrum, as illustrated in Figures 2A–F .

Canopy spectral information, obtained through optical sensors, is a crucial parameter for monitoring crop growth. The VI is closely linked to the physiological and biochemical characteristics of crops, making it an essential tool for assessing crop development. This index captures the interaction between spectral bands and enhances the response to specific crop properties. Based on previous research, 15 spectral vegetation indices were selected to estimate the AGB of rapeseed. The band calculation tool in ENVI 5.3 software was used to compute these VIs, as detailed in Table 1 .

The Gray Level Co-occurrence Matrix (GLCM) is one of the most widely used methods for texture extraction, originally proposed by Haralick (1973). GLCMs became popular due to their ability to maintain rotational invariance, capture multi-scale features, and allow for low-complexity calculations (Haralick et al., 1973). In this study, three texture metrics—Data Range (DR), Variation (VAR), and Entropy (ENT)—were extracted from UAV RGB bands. The selected window size effectively captures variations in spatial information among the rapeseed plants within the experimental plot. A window that is too small can increase computational complexity and the volume of calculations, while a window that is too large may result in the loss of detailed texture information (Bai et al., 2021). To address this, an averaging technique that combines the functionality of different window sizes was employed. Through trial and error, texture features were computed using the average values of two window sizes (3 pixels × 3 pixels and 5 pixels × 5 pixels) and four directional orientations (0°, 45°, 90°, and 135°) rotated clockwise along the x-axis. These features have been shown to be effective in quantifying changes in crop canopy structure and estimating AGB (Yue et al., 2018a).

In this study, UAV RGB images of rapeseed were captured to create a base map for the DEM before sowing. Canopy point cloud images were then acquired to construct the DSM at various growth stages of rapeseed. The height model for each growth period was derived by subtracting the DEM from the DSM. Using the statistical toolbox in ArcGIS and the Kriging interpolation algorithm, the average plant height (PH) was extracted from the region of interest for each image (Xu et al., 2022). Plant roughness (PR), a metric that characterizes the irregularities of the canopy surface, was measured using 3D point clouds from UAV RGB images. PR has been shown to have a significant correlation with crop AGB (Herrero-Huerta et al., 2020). The fraction of plant cover (PC) was determined through image classification, where ground objects in the RGB images were categorized as either crops or soil (Maimaitijiang et al., 2019). PC for each image was calculated by dividing the number of cropped pixels by the total number of pixels in the image. Building on previous research, the volume method was applied to estimate crop biomass within a defined spatial range. The plant volume metric (PVM) of rapeseed was calculated as the product of PC and PH, along with the canopy elevation fluctuation rate (CEFR) to describe the relative shape of the canopy, as commonly used in forestry studies (Han et al., 2019). The canopy structure metrics extracted from UAV RGB images—PH, PR, PC, PVM, and CEFR—are defined in Table 2 .

The input layer of the DNN model used in this study consists of four hidden layers with 256, 128, 64, and 32 neurons, respectively (Hu et al., 2024). A ReLU activation function was applied after each hidden layer. To address overfitting, a dropout layer with a 0.2 ratio was added after the first hidden layer. The network was trained using the Adaptive Moment Estimation (ADAM) optimizer, with a maximum of 600 training iterations and a batch size of 256. The initial learning rate was set at 0.001, decreasing by 10% every 100 rounds. For the Random Forest (RF) model, bootstrap sampling was used to create a training dataset, and random decision trees were generated based on the integrated classifier (Niu et al., 2019). The RF model was configured with 80 decision trees (n tree = 80) and a maximum number of variables considered at each split (m try = 4). The final prediction was determined through a majority voting process among the decision trees. Support Vector Regression (SVR) was applied for linear and nonlinear regression tasks (Liu et al., 2023). The training dataset was binary-classified using a kernel function to minimize the distance of all samples from the hyperplane. The sample data were then fitted to generate predictions.

A total of 60 datasets were collected for each period during the 2022-2023 season. Repeats 1 and 2 were selected as the calibration dataset, while plots from Repeat 3 were used as the validation dataset. The statistical results were presented in Table 3 . To construct an AGB estimation model for rapeseed across various growth stages, a ten-fold cross-validation approach was employed. Pearson correlation analysis was performed to examine the relationship between features and AGB. The model’s performance and stability were assessed using the coefficient of determination (R²), prediction root mean square error (RMSE), and relative root mean square error (rRMSE). The study workflow was shown in Figure 3 .

Where y is the observed value (manual measurement), and y ′ is the predicted value (model extracted value), y ¯ is the average value, and n is the sample size.

For the AGB samples, the average value in the calibration dataset was 1785.6 kg/ha, with an overall coefficient of variation of 38.85%, while the validation dataset had a higher coefficient of variation at 42.38% ( Table 3 ). The minimum AGB value observed was 147.5 kg/ha, and the maximum was 5573.39 kg/ha. For PH, the average was 43.74 cm in the calibration dataset and 42.76 cm in the validation dataset. The overall coefficient of variation for PH was 41.75%, while the validation dataset exhibited a larger variation of 53.18%. These results indicate that the validation dataset generally showed larger coefficient of variation values compared to the calibration dataset.

The correlation between VI metrics and AGB across different growth stages was illustrated in Figure 4 . The AGB and VI values for the three different growth stages of rapeseed showed strong significance (p< 0.01). The strongest correlation during the seedling stage was observed with the RVI, which had a correlation coefficient of r = 0.82 (p< 0.01). Significant correlations were also noted between AGB and SAVI (r = 0.75, p<0.01) as well as NDVI (r = 0.72, p< 0.01). The results suggested a linear relationship between VI and AGB at all growth stages, although the strength of the correlation decreased as the crop developed. It was important to note that VI tends to saturate when AGB was high, meaning the accuracy of AGB estimation using a single VI may require validation through a more comprehensive estimation model.

In this study, the correlation between nine TF metrics and rapeseed AGB across different growth stages was evaluated, as shown in Figure 5 . The strongest correlation was observed between the GVAR and AGB across all three growth stages, with the average correlation exceeding 0.5 (p< 0.01). The correlations between GDR and BENT were also close to 0.5 (p< 0.05). The results indicate that DR, VAR and ENT metrics exhibit significant variability in relation to AGB, suggesting they fluctuate more throughout the growth stages in response to AGB.

The correlation between SF metrics and AGB at the seedling, bolting, and Early blossoming stages of rapeseed was shown in Figure 6 . SF metrics, such as PH, PR, PC, and CEFR, showed significant correlations with AGB across all three growth stages (pp< 0.01). The correlation between these metrics and AGB increased progressively through the growth stages. Notably, the PVM showed a rise in correlation with AGB at first, followed by a decrease in the later growth stages. PH displayed the strongest correlation across all stages, with r values of 0.56 (p< 0.05), 0.67 (p< 0.01), and 0.75 (p< 0.01) at the seedling, bolting, and Early blossoming stages, respectively. These results demonstrate that SF metrics have a linear relationship with AGB at each growth stage.