UAV-Based Multispectral Inversion of Integrated Cotton Growth

Abstract

Cotton growth monitoring can provide macroscopic information for cotton field management. In this paper, we utilised UAV multispectral remote sensing to monitor cotton growth and constructed a Comprehensive Growth Index (CGI) by combining cotton multi-period Leaf Area Index (LAI) and relative chlorophyll content (Soil and plant analyser development, SPAD) through the coefficient of variation method and constructed 28 types of vegetation cover indexes by combining multispectral data. CGI combined with multispectral data to construct 28 vegetation indices and a digital surface model, selected the indexes with higher correlation through correlation analysis, and constructed a single growth model (LAI and SPAD) and CGI by using the support vector machine (SVM), random forest (RF), and gradient boosted decision tree (GBDT) model. In the GBDT model, the comprehensive growth model was better than the single growth index model, with an R² of 0.85 and an RMSE of 0.044, and in the RF model, the comprehensive growth model was better than the single growth index model, with an R² of 0.86 and an RMSE of 0.037. In the comprehensive analysis, the comprehensive growth index can improve the accuracy of the inversion model of the cotton growth and provide a new method for the monitoring of the cotton growth by remote sensing.

1. Introduction

As one of the most important cash crops in the world, cotton occupies an important position in the global agricultural economy. However, traditional cotton growth monitoring methods are limited by the limitations of ground observation, which makes it difficult to monitor growth information comprehensively and accurately. The use of remote sensing technology to achieve the monitoring of the cotton growth process is faster and more accurate than the traditional monitoring methods, which provides a new reference direction for cotton growth monitoring.

The inverse method is a type of mathematical modelling or statistical analysis that derives a system’s internal states or parameters by reversing observed data. In the field of remote sensing, the inverse method is widely used to infer feature attributes or environmental states using remotely sensed image data. The aim of this study is to predict cotton growth using UAV multispectral image data, where the inverse method is able to relate image features to comprehensive growth indicators, providing an indirect measure of variables that are not directly observable. With the advantages of high spatial and temporal resolution, low cost, and flexibility, UAV multispectral technology has gradually become an important tool for agricultural production monitoring and management. By carrying a multispectral camera on the payload platform, UAVs can achieve high-resolution and high-frequency monitoring of crop growth status, providing more accurate data support for agricultural production. It is widely used in the field of agriculture for crop physiological and chemical parameters, pests and diseases, yield prediction, etc. [1,2]. In the field of agriculture, it is widely used in crop physiological parameters, pest and disease, and yield prediction. In terms of model applications, machine learning models have been widely used for tasks such as crop identification, pest and disease monitoring, and yield prediction. Random forest (RF) has shown strong stability and generalisation ability when dealing with high-dimensional feature data and has proved to be suitable for crop identification [3]. Support vector machines (SVMs) show advantages in scenarios with less data, such as pest and disease detection, due to their ability to handle small sample problems [4]. Gradient boosting decision tree (GBDT), on the other hand, excels in complex yield prediction tasks due to its non-linear modelling capabilities [5]. However, the performance of different models varies in different application scenarios, so choosing the right model is crucial to enhancing the effectiveness of agricultural remote sensing monitoring. In monitoring the growth of cotton [6,7], there are now more studies on its leaf area index [8], relative chlorophyll content [9], plant height [10], nitrogen [11], and photosynthetically active radiation [12], and the optimisation of the model has achieved better prediction results. Puchen et al. [13] used UAV multispectral construction of a cotton LAI inversion model; the inversion effect is better, and the results can provide a methodological basis for remote sensing monitoring of cotton growth, but the above studies are mostly from a single growth indicator inversion research, and a single indicator is difficult to comprehensively respond to the real growth of cotton, so in recent years there are a lot of scholars of these single indicators to reduce the dimensionality of construction of a comprehensive growth indicator. In the study of wheat, Pei Haojie et al. [14] combined winter wheat chlorophyll content, leaf area index, plant nitrogen content, plant water content, and biomass according to the mean and other weights to construct a comprehensive growth index. Chunlan Z [15] compared the inverse effect of the equal-weight method and principal component analysis in constructing comprehensive growth indicators and found that the model of comprehensive growth indicators constructed by using principal component analysis was more accurate. In the study of maise, Jinghua Z et al. [16] used the integrated growth indicator combined with the PSO-RF model to invert spring maise growth, while Feng Wenbin et al. [17] constructed the integrated growth indicator on summer maise with a model accuracy of 0.752.

In this study, we combined the UAV multispectral remote sensing technology with the ground survey of LAI and SPAD at three key fertility stages of cotton and constructed a comprehensive cotton growth inversion model based on the coefficient of variation method [18] and assessed the feasibility of the comprehensive cotton growth index by comparing the advantages and disadvantages of the comprehensive growth index with the single growth index in the traditional model and the machine learning model [19], so as to provide a new reference for the cotton growth monitoring method.

2. Materials and Methods

2.1. Overview of the Test Area

2.1.1. Location and Climatic Conditions

The experimental site of this experiment was located in Bincheng District, Binzhou, Shandong Province, China, which has a humid continental climate with four distinct seasons and abundant rainfall under the influence of the natural environment, solar radiation, and monsoon. The average annual temperature is about 12.71 °C, the average annual precipitation is about 564.8 mm, and the average annual sunshine time is about 2632.0 h, which is suitable for cotton growth.

2.1.2. Experimental Design and Layout

Cotton in the experimental site was sown on 25 April 2023 in an area of about 2610 m². Four N gradients (N1–N4) were set for fertiliser in the experimental area prior to sowing, namely 0 kg/hm² (N1, insufficient N fertiliser), 60 kg/hm² (N2, moderate), 120 kg/hm² (N3, higher), and 180 kg/hm² (N4, excessive). These gradients covered the typical range of N fertiliser application from insufficient to excessive and helped to investigate the effects of N fertiliser management on crop growth and remote sensing inversion parameters. Two replicates were set up for each gradient to reduce the interference of soil heterogeneity and fluctuating environmental conditions. Fertiliser in all experimental areas was applied at once as a base fertiliser, without follow-up fertiliser. A 1–2 m-wide separation zone was provided between plots to avoid mutual interference during fertiliser application and operations. The cotton varieties used in the experiment were Zhong100, Lu37, 6269, and Lu338. Each variety was planted in four rows, and three sampling points were evenly divided in each plot during data collection. Field management was in accordance with the local field requirements for unified management (Figure 1).

2.2. UAV Image Acquisition and Pre-Processing

2.2.1. Ground Sampling

Three sampling points were evenly divided in each plot. In this experiment, important reproductive periods of cotton were collected at the boll stage (23 August), fluffing stage (10 September), and diapause stage (27 September). For each sample point, leaf area index (LAI) and chlorophyll content (SPAD) metrics were collected for the above-ground portion of the crop. These data were used to verify the accuracy of the remote sensing inversion results.

2.2.2. Cotton Canopy Image Acquisition

In this experiment, UAV orthophotos of cotton were collected at the boll stage (23 August), fluffing stage (10 September), and diapause stage (27 September) during the important reproductive periods of cotton. The weather was sunny and cloudless on the day of measurement using the DJI Phantom 4 Multispectral Edition, and the flights were conducted at short intervals between 10:00 and 14:00 to ensure sufficient and stable sunlight. During each flight, the standard reflectance panels were fully exposed and unobstructed to the sun, and hand-held shots were taken using the Phantom4-Multi directly overhead for post-radiometric calibration. The shooting angle should be perpendicular to the ground, the overlap rate of the drone’s heading is 80%, the overlap rate of the side direction is 80%, and the wind should generally be no more than level 4, so as to ensure the stability of the flight and the effect of the subsequent image stitching (Figure 2 and

Table 1. UAV-related parameters.

Multi-Spectral Camera Parameter Settings	Numerical Value
effective pixel	2 million
camera interval (between photos)	2 s
ground resolution	1 cm/pixel
flight level	25 m

).

The Phantom4-Multi UAV has RTK, which can achieve centimetre-level positioning with an error of 1 cm + 1 ppm in the horizontal direction (1 ppm means that the error increases by 1 mm for every 1 km of movement of the vehicle) and an error of 1.5 cm + 1 ppm in the vertical direction. Meanwhile, fixed aerial calibrations are added during the whole period of cotton growth and calibrated with the XAG A2 intelligent hand-held GPS device (Pole Fly Ltd., Guangzhou, China) for geographic information calibration during splicing (Figure 3).

2.2.3. Image Pre-Processing

Pix4D 4.5.6 (was used to complete the multispectral remote sensing image stitching operation. Before the images are stitched, radiometric corrections are performed in the Pix4D index calculator using the standard reflector plates captured. Radiometric correction is a key step in the preprocessing of remotely sensed data, and its main goal is to eliminate radiometric errors caused by changes in sensor characteristics, atmospheric conditions, and illumination, thus ensuring the consistency of data collected at different times and locations. This process is particularly important for the accurate interpretation of UAV remote sensing data because UAV platforms usually fly at low altitudes and are susceptible to light conditions and sensor characteristics. The standard reflector (Changhui Guangzhou, China, CH-MXDJ255075) used in this study is adapted to the DJI Phantom4-Multi, which provides a highly stable spectral reflectance reference in the wavelength range of 420–1000 nm and is able to effectively correct for inconsistencies in the sensor data due to variations in the lighting conditions, eliminating the effects of different flight conditions on the data. Before image stitching, radiometric correction was performed in the Pix4D initialisation process with the standard reflectance plate that was shot. After loading the radiometrically calibrated multispectral images into Pix4D and selecting the “Agricultural Multispectral” mode for stitching, it can be seen from the aerotriangulation that the quality of the puzzle is good and there are green lines and no red (bad) dots (Figure 4).

The band synthesis and the calculation of the cotton vegetation index were carried out by ENVI 5.3 software. A total of 28 vegetation indices and DSM were extracted from five bands of red, green, blue, near-infrared (NIR), and red edge based on five multispectral lenses to establish a model remote sensing dataset related to cotton growth. In order to remove the effect of soil noise on vegetation index extraction, OSAVI [20], which is more sensitive to soil and green vegetation, was used. Bare ground was masked by the threshold segmentation method. The geographic coordinates of the sampling points were obtained using the Polaris A2 intelligent handheld positioning point, and this was used to create vector files of the sampling points in ENVI. And use the vector file with the same range to batch extract the DN value of the vegetation index in Arcgis10.8 software, which can ensure that the position of the sampling points taken in remote sensing images in different periods will not be offset (

Table 2. Vegetation index formula for drone multispectral imagery.

Note	Vegetation Index	Formula
1	NDVI [21]	$NDVI=(R_{nir}-R_{red})/\left(R_{nir}+R_{red}\right)$
2	SR [22]	$SR=R_{nir}/R_{red}$
3	EVI [22]	$EVI=2.5\left(R_{nir}-R_{red}\right)/\left(R_{nir}+6R_{red}-7.5R_{blue}+1\right)$
4	DVI [23]	$DVI=R_{nir}-R_{red}$
5	NLI [24]	$NLI=(R_{nir}\ast R_{nir}-R_{red})/\left(R_{nir}\ast R_{nir}+R_{red}\right)$
6	GNDVI [25]	$GNDVI=(R_{green}-R_{red})/\left(R_{green}+R_{red}\right)$
7	RENDVI [26]	$RENDVI=(R_{rededge}-R_{red})/\left(R_{rededge}+R_{red}\right)$
8	MDD [27]	$MDD=(R_{nir}-R_{rededge})/\left(R_{rededge}-R_{red}\right)$
9	SIPI [28]	$SIPI=(R_{nir}-0.5\ast R_{red})/(R_{nir}+0.5\ast R_{red}$)
10	NGI [29]	$NGI=R_{green}/(R_{nir}+R_{green}+R_{rededge}$)
11	OSAVI [30]	$OSAVI=\left(1+0.16\right)(R_{nir}-R_{red})/\left(R_{nir}+R_{red}+0.16\right)$
12	MNLI [24]	$MNLI=1.5\left(R_{nir}\ast R_{nir}-R_{red}\right)/\left(R_{nir}\ast R_{nir}+R_{red}+0.5\right)$
13	WDRVI [27]	$WDRVI=(a{R}_{nir}-R_{red})/\left(a{R}_{nir}+R_{red}\right)$ ($a=0.12)$
14	VI [31]	$VI=R_{Green}/R_{red}$
15	RESR [32]	$RESR=R_{rededge}/R_{red}$
16	SAVI [33]	$SAVI=1.5\left(R_{nir}-R_{red}\right)/\left(R_{nir}+R_{red}+0.45\right)$
17	CCCI [34]	$CCCI=\left(R_{nir}-R_{rededge}\right)/\left(R_{nir}+R_{rededge}\right)$
18	VARI [35]	$VARI=\left(R_{blue}-R_{red}\right)/\left(R_{blue}+R_{blue}-R_{green}\right)$
19	GRVI [36]	$GRVI=(R_{green}-R_{red})/\left(R_{red}+R_{green}\right)$
20	NGBDI [37]	$NGBDI=(R_{green}-R_{blue})/\left(R_{green}+R_{blue}\right)$
21	NGRDI [37]	$NGRDI=(R_{green}-R_{red})/\left(R_{green}+R_{red}\right)$
22	RGRI [35]	$RGRI=R_{red}/R_{green}$
23	MTCI [38]	$MTCI=(R_{nir}-R_{rededge})/\left(R_{rededge}-R_{red}\right)$
24	TSAVI [39]	$TSAVI=(R_{nir}-10.489\ast R_{red}-6.604)/\left(R_{nir}+R_{red}-6.604\right)$
25	ARVI [40]	$ARVI=\left(R_{nir}-2\ast R_{red}-R_{blue}\right)/\left(R_{nir}+2\ast R_{red}+0.5\right)$
26	TVI [41]	$TVI=60\ast (R_{nir}-R_{red})-100\ast \left(R_{red}-R_{green}\right)$
27	NDRE [42]	$NDRE=(R_{nir}-R_{rededge})/\left(R_{nir}+R_{rededge}\right)$
28	NDI [43]	$NDI=R_{rededge}/R_{red}$

).

2.3. Construction of Integrated Cotton Growth Indicators

In order to respond to the comprehensive growth of cotton, combined with the field survey. LAI and SPAD were selected to construct comprehensive growth indicators. Based on the coefficient of variation method, we constructed the cotton comprehensive growth index (CGI) [17]. The main principle is to compare the degree of dispersion of different groups or different variables to determine which growth indicator has a greater degree of variation to determine its proportion in the CGI. By comparing the coefficients of variation of different indicators, the relative importance of each indicator in the data set can be clarified. This helps to highlight more accurately the key indicators that have a greater impact on the evaluation results in a comprehensive multi-indicator evaluation. The formula is as follows:

$$V_i={\displaystyle \frac{{\mathsf{\sigma }}_i}{{\overline{x}}_i}}(\mathrm{i}=1,2,\cdots ,n)$$

(1)

$$W_i={\displaystyle \frac{V_i}{{\sum }_{i=1}^n{V}_i}}$$

(2)

$$CGI=W_i·N_i+W_i·N_i$$

(3)

in which

$V_i$ denotes the coefficient of variation of the ith indicator;

${\mathsf{\sigma }}_i$ denotes the standard deviation of the ith indicator;

${\overline{x}}_i$ denotes the mean of the ith indicator;

$W_i$ denotes the weight of the ith indicator;

$N_i$ denotes the normalisation of the ith indicator.

2.4. Cotton Integrated Growth Prediction Model Construction

Random forest (RF) is an integrated learning method [44] that belongs to an extension of decision trees. It consists of multiple decision trees, each of which is generated based on training on a random subset of input features. When making predictions, each decision tree in the random forest predicts the input data and then determines the final prediction based on voting or averaging.

Support vector machine (SVM) is a supervised learning algorithm [45] that is used to solve classification and regression problems. The main idea is to find an optimal hyperplane in the feature space that separates the samples of different classes while making the interval from the samples to the hyperplane as large as possible in order to achieve the optimal balance between model complexity and learning capability.

The gradient regression tree model is iteratively trained with a series of regression trees, and each tree is constructed to minimise the residuals of the previous tree [46]. Unlike traditional decision trees, the gradient regression tree model uses the technique of gradient boosting. It uses gradient descent to optimise the loss function in each iteration to gradually reduce the residuals. During the construction of the tree, each tree is constructed on the basis of the residuals of the previous tree to make the overall model fit progressively better.

In this study, the coefficient of determination (R²) and root mean square error (RMSE) metrics were selected to evaluate the model accuracy. The coefficient of determination of the model is approximately close to 1. The smaller the root mean square error, the better the prediction effect of the model and the higher the stability.

$$R^2=1-{\displaystyle \frac{{{\displaystyle \sum _i(x_i-y_i)}}^2}{{\displaystyle \sum _i{(x_i-\overline{x})}^2}}}$$

(4)

$$RMSE=\sqrt{{\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n{(x_i-y_i)}^2}}{n}}}$$

(5)

in which

$x_i$ denotes the measured value of cotton growth parameters;

$\overline{x}$ denotes the measured mean value of cotton growth parameters;

$y_i$ denotes the predicted value of the cotton growth model;

n denotes the total number of model samples.

3. Results

3.1. Analysis of Ground-Truthing Data

The LAI and SPAD data measured in the experiment are shown in

Table 3. Statistics of measured data.

	DATA SET	Sample	Maximum	Minimum	Average	Standard Deviation	Coefficient of Variation
SPAD	boll stage	32	64.7	34.2	50.1	5.3	10.7
	fluffing stage	32	64.1	39.6	54.5	4.9	9.1
	diapause stage	32	64.3	34.8	50.1	5.7	11.3
LAI	boll stage	32	4.2	0.1	3.3	1.2	38
	fluffing stage	32	4.3	0.36	2.5	0.82	32
	diapause stage	32	4.4	0.73	2.8	0.69	24

. Outliers were eliminated prior to the calculations to attenuate the effects of measurement errors. Table 3 shows that the average value of SPAD is the largest in the fluffing stage of cotton, which indicates that the chlorophyll of cotton increases with the birth of cotton, and the coefficient of variation is smaller than that of other periods; the distribution of chlorophyll is more homogeneous; it reaches the maximum value in the fluffing stage and then decreases from the flocculation period to the diapause period, and the level of LAI grows continuously and reaches the maximum value in the diapause period, and the differences in the coefficient of variation of the three periods are small, which is suitable for modelling and analysis. The differences in the coefficient of variation in the three periods were small and suitable for modelling analysis.

3.2. Correlation Analysis

In order to select the independent variables with a high correlation with growth and improve the model training accuracy, the extracted vegetation indices and digital elevation model DSM in 29 should be correlated with the measured data in the field. In this study, we chose to use Pearson’s correlation test to carry out the analysis, as shown in Figure 5. By analysing the LAI, SPAD, and the integrated growth indices of the three key fertility periods of cotton, the correlation between LAI and SPAD is weak, which is beneficial to the calculation of the coefficient of variation and can more realistically reflect its degree of dispersion and relative importance in the dataset. It indicates that the indicators constructed by the two can well reflect their contribution to the composite growth index and improve the stability and interpretation of the model.

The correlation between SPAD and GNDVI, NGI, and GRVI was high; all of them were positively correlated, with the absolute values of their correlation coefficients above 0.8, and the results of the significance analysis showed that there was a highly significant relationship between the two (p ≤ 0.01). The correlation between NGI and SPAD is the highest, with an absolute value of correlation of 0.85, which is a strong correlation; the absolute value of correlation between LAI and NDVI, EVI, VARI, NGBDI, and DSM ranges from 0.6 to 0.8, with the absolute value of correlation between NDVI and EVI being greater than 0.75 and is a highly significant relationship. The correlation between CGI and vegetation indices VARI, NLI, MNLI, EVI, CCCI, and DSM is high, in which the absolute value of the correlation of VARI and NLI is greater than 0.8, and the absolute value of the correlation of MNLI, EVI, CCCI, and DSM is between 0.6 and 0.8, and all of them are in a highly significant relationship.

3.3. Quantitative Evaluation of Radiation Correction

Radiometric correction has a significant impact on the consistency of spectral data and the performance of inversion models. In this study, the errors corrected by the radiometric correction come from the changes in atmospheric conditions and light angles at the three experimental sites. By comparing the full-period inversion data of CGI before and after correction, it is found that the data are more consistent after correction, and the performance index of the inversion model is significantly improved. As shown in

Table 4. Radiation correction effect.

CGI	RF		SVM		GBDT
	R2	RMSE	R2	RMSE	R2	RMSE
Before calibration	0.79	0.050	0.66	0.067	0.74	0.065
After calibration	0.86	0.037	0.76	0.061	0.85	0.044

, the R² of the corrected RF model increased by 8.8%, the RMSE decreased by 33.3%, and the SVM had the worst error correction ability. These results demonstrate that radiometric correction is an important step in the preprocessing of multispectral data.

3.4. Modelling and Analysis

RF, SVM, and GBDT methods were used to construct the inversion models of LAI, SPAD, and CGI for the whole plantation period of cotton, and the indices with an absolute value of correlation higher than 0.6 were selected as the input variables of the model to be used in the construction of the inversion model of cotton growth. The modelling results and validation results are shown in

Table 5. Inversion results for the model test set.

Model Variable	RF		SVM		GBDT
	R2	RMSE	R2	RMSE	R2	RMSE
SPAD	0.71	0.110	0.86	0.070	0.85	0.088
LAI	0.77	0.090	0.86	0.070	0.67	0.120
CGI	0.86	0.037	0.76	0.061	0.85	0.044

and Figure 6.

Among the inversion models, the coefficients of determination of the CGI inversion model for the test set under the RF and GBDT-based modelling approaches were higher than those of the LAI inversion model and the SPAD inversion model, and their root mean square errors were lower than those of the LAI inversion model and the SPAD inversion model, which indicated that the integrated growth indicators could effectively predict the sample data and the prediction accuracies were higher compared with those of the LAI inversion model and the SPAD inversion model. Among the CGI inversion models in terms of model principles, the RF model performs the best with a test set coefficient of determination R2 of 0.86 and a root mean square error of 0.037; the GBDT is the next best with an R2 of 0.85 and an RMSE of 0.044; and the SVM model has the lowest performance, with an R² of 0.76 and an RMSE of 0.061. The reason for the better results of the RF model is that it is used as an integrated learning method that consists of multiple decision trees. Compared to SVM and GBDT models, RF is able to capture the complex nonlinear relationship between the input variable (CGI) and the target variable (cotton growth) well. The reason for the poor results of the SVM model is that its kernel function has limited performance in handling high-dimensional nonlinear data. In this study, CGI integrates multiple features (LAI and SPAD), which may have significant interactions with each other. The stochastic segmentation and integration mechanism of RF enables it to mine and exploit such multidimensional variable interactions, which significantly improves the predictive power of the model. Therefore, RF was finally chosen to construct the CGI inversion model to detect the cotton growth.

In order to verify that the growth index has a good inversion effect in each plot, the ground growth data of different treatments in each period were respectively used as the validation set to verify the inversion effect of the CGI model. As shown in Figure 7, among the three models, RF has the best inversion effect and high stability of the model; for plots with higher nitrogen fertiliser content, the difference in cotton growth is smaller, which has a bad influence on the inversion effect, while RF has higher stability in this regard.

CGI inversion of cotton growth at different fertility periods; the results are shown in Figure 8. The research area of cotton growth in the bolling period can be seen with obvious differences; the growth of better is mainly concentrated in the middle of the plot, nitrogen-deficient plots, and the plot edge of the growth of poor, in the spitting period of the cotton tends to stabilise the growth of cotton in the diapause of cotton basically stops growing, and the majority of the leaves withered. Through the image, one can visualise the growth of cotton at different stages.

4. Discussion

In this study, the Comprehensive Growth Indicator (CGI) characterises the overall crop growth by integrating several key growth characteristics (LAI and SPAD). In order to improve the scientific validity and applicability of CGI, it is necessary to undergo radiometric correction when collecting remote sensing data. In this study, radiometric correction significantly improves the stability of spectral data and the prediction ability of inversion models, but the effect of radiometric correction may vary under different experimental designs and scenarios, especially under different platforms or lighting conditions; the effect of using standard reflector plates and the adaptability of correction algorithms may change. In addition, differences between equipment and algorithms can significantly affect the quality of the corrected data [47], making it difficult to generalise and quantify directly.

In this study, the coefficient of variation method was used to pre-weight the features. This method assigns weights to features based on their degree of dispersion. And the theoretical basis is that features with a higher degree of dispersion are usually more capable of distinguishing between growth states, and the variability of LAI in cotton is significantly higher than that of SPAD values, so it is assigned a higher weight, which is consistent with the actual observation of growth dynamics. This study demonstrated the successful application of CGI in cotton growth monitoring, but it is limited by the experimental location, which is relatively single and small; in different regions, of the climate environment and other factors, the experimental results have certain geographical limitations. Subsequently, to improve the model, it can be used based on UAV image investigation of cotton orthophotos in different regions to further explore the inversion accuracy of the integrated growth indicators. When applying CGI to different crops, variable selection and weight allocation may need to be adjusted according to crop characteristics. For example, for maise, key growth parameters such as plant height and leaf thickness may be included, whereas rice may be more dependent on the dynamics of vegetation indices. Subsequent comparative experiments can be conducted by introducing more crop types and ecological conditions to verify its broad applicability.

Regarding the selection of model variables, Cheng Dayu et al. [48] compared two growth monitoring models based on vegetation index and two models based on vegetation index and texture features and selected the top three indicators (NDVI, BNDVI, GNDVI, R_Mean, G_Mean, B_Mean) with the highest absolute values of the correlation coefficients of the vegetation indices and texture features and then combined them and found that the CGMIs constructed by the three modelling methods with the addition of texture features were all more accurate. The models all have improved accuracy. In this paper, only the UAV-acquired multispectral orthophoto and digital elevation models are discussed, and the texture features of the images are not yet involved, and the effect of adding texture features on the accuracy of the models needs to be further investigated.

5. Conclusions

In this study, a comprehensive growth indicator (CGI) was constructed by the coefficient of variation method using UAV multispectral images combined with ground-based measured cotton growth indicators (LAI and SPAD). The results showed that the prediction performance of the inversion model of CGI was better than that of a single growth indicator (LAI and SPAD) under both RF and GBDT models, and the RF model performed the best in terms of prediction accuracy and stability, which demonstrated its efficient modelling ability for multidimensional feature data. In addition, the introduction of CGI not only improved the prediction accuracy but also enhanced the stability of the model, which fully demonstrated the advantages of comprehensive growth indicators in field management.

The CGI indicators constructed in this study effectively make up for the limitations of the prediction model of a single growth indicator, which can reflect the growth status of cotton more comprehensively and provide a more scientific and accurate means of growth monitoring for field management. This method has high application value and can provide data support for agricultural producers to optimise field management, precise fertiliser application, and pest prevention and control. Taking the experimental site in this paper as an example, appropriate fertiliser can be applied to the weak places during the bolling period, so as to improve the yield and quality of cotton.

In future research, the experimental scope should be further expanded to include cotton planting data under different geographical, climatic, and soil conditions to improve the generalisation ability and robustness of the model. In addition, an integrated growth indicator construction method that combines texture features, time series analysis, and more growth parameters (e.g., canopy temperature and biomass) is expected to further improve model accuracy and stability. In practical agricultural applications, CGI can be integrated into UAV-based agricultural management systems to develop integrated platforms that combine time-series analysis and automated decision-making tools to enable continuous monitoring of crops during different growing periods and generation of management recommendations. For example, CGI can be used to monitor nitrogen fertiliser use efficiency and disease expansion areas to provide field managers with precise regulatory solutions.

In addition, advancing the implementation of CGI in actual agriculture requires interfacing with existing precision agriculture platforms (e.g., farm management software or UAV control systems), designing user-friendly interfaces and automated data processing processes, and lowering the threshold of use for agricultural practitioners. At the same time, the feasibility of its large-scale dissemination should be explored in conjunction with policy support and economic analysis, so as to provide technological safeguards for agricultural modernisation and sustainable development.