Unlocking the Secrets of Corn: Physiological Responses and Rapid Forecasting in Varied Drought Stress Environments

Abstract

Understanding the intricate relationship between drought stress and corn yield is crucial for ensuring food security and sustainable agriculture in the face of climate change. This study investigates the subtle effects of drought stress on corn physiological, morphological, and spectral characteristics at different growth stages, in order to construct a new drought index to characterize drought characteristics, so as to provide valuable insights for maize recovery mechanism and yield prediction. Specific conclusions are as follows. Firstly, the impact of drought stress on corn growth and development shows a gradient effect, with the most significant effects observed during the elongation stage and tasseling stage. Notably, Soil and Plant Analyzer Development (SPAD) and Leaf Area Index (LAI) are significantly affected during the silking stage, while plant height and stem width remain relatively unaffected. Secondly, spectral feature analysis reveals that, from the elongation stage to the silking stage, canopy reflectance exhibits peak–valley variations. Drought severity correlates positively with reflectance in the visible and shortwave infrared bands and negatively with reflectance in the near-infrared band. Canopy spectra during the silking stage are more affected by moderate and severe drought stress. Thirdly, LAI shows a significant positive correlation with yield, indicating its reliability in explaining yield variations. Finally, the yield-related drought index (YI) constructed based on Convolutional Neural Network (CNN), Random Forest (RF) and Multiple Linear Regression (MLR) methods has a good effect on revealing drought characteristics (R = 0.9332, p < 0.001). This study underscores the importance of understanding corn responses to drought stress at various growth stages for effective yield prediction and agricultural management strategies.

1. Introduction

Drought stress poses a significant challenge to agricultural productivity and hinders socio-economic development. With severe climate change and frequent human activities, global surface temperatures are increasing year by year, leading to a noticeable increase in the frequency of extreme drought events [1,2]. China, being a major agricultural country, accounts for 31.1% of the world’s agricultural output, and water scarcity severely reduces corn yield, which is an important economic crop and agricultural pillar [3,4,5]. Therefore, identifying the changes in physiological characteristics of vegetation and accurate estimation of corn yield under drought conditions is crucial for providing information for agricultural practices, ensuring food security, and preventing economic losses.

The impact of water availability on plants is a complex process. In the complex water exchange between soil, plants, and the atmosphere, roots absorb soil moisture, transport it to the leaves through capillaries, and distribute it to various organs through physiological responses, such as photosynthesis, osmosis, and regulation [6,7,8]. Drought stress has multiple effects on crops, especially by altering their physiological and ecological responses. Photosynthesis is essential for organ formation and biomass accumulation, but it decreases under drought conditions primarily due to reduced water use efficiency and stomatal conductance. Osmotic regulation helps maintain cell expansion under water stress through the accumulation of solutes [9,10,11]. At the same time, internal plant hormones, especially abscisic acid, increase, triggering stomatal closure to limit water loss [12]. The accumulation of reactive oxygen species exacerbates cell damage under drought conditions, while the presence of antioxidant enzymes mitigates the side effects of drought [13]. Additionally, drought has significant effects on plants, including wilting, curling, chlorophyll degradation, changes in root systems, and growth inhibition of plant height and stem thickness. Among these, internal biochemical parameters, such as leaf area index (LAI) and chlorophyll, as well as external features, such as plant height and stem width, are commonly used drought monitoring indicators due to their ease of measurement and ability to reflect vegetation’s physiological status [14,15,16].

In the field of agriculture, drought, as a key factor affecting crop yield, has historically attracted much attention and stimulated a wide range of research interests. In exploring the impact of drought on agricultural output, we have to mention its complexity and multi-dimensional characteristics. Drought not only affects soil moisture content, but also indirectly affects crop growth and development by affecting temperature, light, air humidity and other environmental factors. Traditional drought assessment methods mainly rely on the field observation of grass-roots units, but this method often ignores the comprehensive impact of these comprehensive environmental factors on crop growth, which is not only time-consuming and laborious, but also easy to lead to inaccurate assessment results due to human factors [17,18]. To overcome this limitation, researchers have begun to explore more accurate and efficient soil moisture measurement methods, such as neutron technology and dielectric methods, which can effectively improve the accuracy and efficiency of drought monitoring by penetrating into the soil layer.

With the advent of satellite technology and the rapid development of remote sensing, the field of drought assessment underwent a revolutionary transformation. Modern drought assessment methods are primarily based on two core concepts: one is the continuation of early soil moisture monitoring approaches, utilizing satellite-based thermal inertia methods to monitor soil moisture content as an indicator of crop drought conditions [19]. However, this method solely considers the impact of soil moisture on crops, neglecting other potential drought influencing factors. The second approach involves constructing diversified drought indices, which have evolved from early water balance models to today’s complex constructions incorporating multi-model coupling, linear and joint distribution functions, and data mining techniques [20,21]. These drought indices not only improve the spatial and temporal resolution of drought monitoring but also rapidly capture crop responses to drought at the appearance level, providing robust data support for agricultural drought monitoring.

In terms of yield prediction and index construction, various models are often employed. Among them, the Random Forest algorithm, as an ensemble learning technique, leverages the advantages of classification and decision trees to demonstrate high precision, the ability to discern complex feature relationships, efficient execution, and resistance to overfitting. It utilizes Bootstrap sampling to construct training and out-of-bag test sets, recursively splits to form multiple decision trees, and then aggregates predictions through voting and scoring to obtain optimal results [22,23,24]. Additionally, Multiple Linear Regression (MLR) is a statistical method used to study the linear relationship between a dependent variable and multiple independent variables. It aids in understanding the combined effects of independent variables on the dependent variable and predicting possible values of the dependent variable given the values of the independent variables, making it suitable for constructing comprehensive drought indices and other applications [25]. Furthermore, Convolutional Neural Networks (CNNs) possess advantages such as automatic feature learning from data, reduced human intervention, and hierarchical abstract feature representation capabilities [26]. They can extract abstract information from data through a multi-layered network structure, which helps in understanding the complexity of the data. However, considering the variability of input data and regional environments, the applicability of different models varies, necessitating an assessment of their suitability during the process of index construction.

Despite significant advancements in remote sensing technology for drought monitoring, it still faces certain limitations. For instance, while data sources are vast, their resolution is often low, and most methods can only detect the occurrence and cessation of drought events, making it difficult to accurately assess the specific impacts and severity of drought on crop yields. The universality promotion of drones provides hope for precision crop monitoring and provides high-resolution data and richer information for dynamic on-site monitoring [14,27]. Therefore, the challenge lies in how to quickly and accurately predict future yields based on real-time crop growth conditions.

In response to this challenge, this study aims to develop a new drought index that reflects crop yields. We employ high-precision resolution drones for aerial monitoring, combined with detailed field experiments, to ensure the accuracy and reliability of the index. The detailed model can be seen in Figure 1. By controlling water stress on maize plants at different growth stages and drought severity levels, we select typical physiological, morphological, and spectral characteristics for drought impact analysis and utilize multispectral, visible, and thermal infrared imagery to screen for the optimal drought index. Building upon this foundation, we integrate commonly used data fusion methods to propose a novel drought index, aiming to pave the way for future satellite-based applications. The significance lies in enhancing public understanding of the dynamic changes in physiological and morphological characteristics of maize plants during growth under varying levels of drought stress, enabling rapid, accurate, and reliable large-scale yield predictions. This aids local irrigation districts and farmers in effectively managing the impact of drought on maize production, safeguarding food security.

2. Study Area, Data and Methods

2.1. Subsection

The experimental field is located in Fuping County, Weinan City, Shaanxi Province (109°17′41″–109°17′44″E, 34°41′18″–34°41′21″N). The area belongs to the Liuqu system of of the Donglei irrigation district phase II Project. The irrigation district located in the east of the Guanzhong plain of Shaanxi Province. It is situated in the northern part of the loess hilly area west of the Yellow River and north of the Wei River, with an altitude between 385 m and 635 m. It borders the Yellow River to the east, Fuping County to the west, the “Shaanxi Province Jiaokou pumping Wei irrigation area” and “Weinan City Luohui canal irrigation area” in the south, and Qiao mountain in the north, where the terrain is higher in the northwest and lower in the southeast (Figure 2). The area has a temperate continental monsoon climate, with cold and dry winters and hot and rainy summers. Rainfall mainly occurs in the summer and the average annual rainfall is between 519 mm and 552 mm. However, the average annual evaporation is relatively high, ranging from 1700 mm to 2000 mm, making it a typical semi-arid region.

2.2. Experiment Design

2.2.1. Plot Design

An area with a wide field of vision and flat terrain was selected as the experimental field. The field is mainly composed of loess soil, with sufficient water sources and easy irrigation. According to the field conditions and experimental needs, the experimental field was divided into two equally sized parts, each measuring 10 m × 16 m (Figure 3). Because we carried out different drought treatment on plots at different growth stages, the corresponding treatment status of specific plots are shown in

Table 1. Treatment period of drought stress conditions.

	Drought Stress State
		Mild	Moderate	Severe	Extreme
Drought treatment period	Elongation	P12	P3	P4	P7
	Tasseling	P2	P9	P10	P8
	Silking	P5	P1	P6	P11

. A closable rain shelter was built over each plot to exclude the impact of precipitation on the drought stress plots. Each part was evenly divided into six 4 m × 4 m plots with a 2 m gap between them to prevent water leakage and to isolate the subplots for smooth implementation of the water control experiment. To accurately control the drought stress, 17 sets of soil moisture sensors were installed in the experimental field, including 11 RainRoot devices (The device was purchased from Beijing Rainroot Scientific Co., Ltd., Beijing, China) and 6 KEBAI devices (The device was purchased from KEBAI Sciences Co., Ltd., Jinan, Shandong, China). Among them, the observation time of the soil moisture sensors was set at 30 min.

The local dominant variety, Yufeng 620, was selected as the corn planting variety for the study. According to the conventional planting standards, the plant spacing in the corn field was set at 70 cm, and the seeding depth was controlled between 3 and 5 cm. The seeding date was set for 27 June, the most suitable time for the local climate conditions. During the growth process of corn, common management measures used by local farmers were adopted, including, but not limited to, timely irrigation, fertilization, weed control, and pest control, to ensure that corn could grow smoothly and achieve good yield and quality.

2.2.2. Water Control Design

According to the growth and development rules of corn, the entire growth period was divided into five stages: emergence, elongation stage, tasseling stage, silking stage and maturity stage. To ensure a high emergence rate and strong growth and reproductive ability of seedlings to cope with drought stress, normal water treatment was applied after seeding until emergence. The drought stress level was divided according to the Agricultural Meteorological Observation Standard for Corn-QX/T361-2016, and specific information is shown in

Table 2. Corn water control scheme.

Drought	Elongation	Tasseling	Silking
Normal	>70	>75	70
Mild	60–70	65–75	60–70
Moderate	50–60	55–65	50–60
Severe	45–50	50–55	45–50
Extreme	≤45	≤50	≤45

. At the same time, soil moisture probes were buried at depths of 10 cm, 20 cm, and 40 cm in each plot to measure the soil moisture at different depths. The drought stress level of each plot was controlled based on the field water holding capacity (29.5%).

2.3. Observation Parameters

2.3.1. Drone-Based Spatial Measurement

For the experiment area, fixed flight points were selected and a drone (DJI Phantom 4 Pro, equipment purchased from DJI Innovation Co., Ltd., Beijing, China. Specific pictures can be seen in the Figure 4a) with visible light lens (CMOS, equipment purchased from DJI Innovation Co., Ltd., Beijing, China) (Figure 4c), multispectral lens (RedEdge-M, equipment purchased from Mapu Technology (Fuzhou) Co., Ltd., Fuzhou, China) (Figure 4b), and thermal infrared lens (VUE PRO, equipment purchased from Beijing Ruilian Toda Technology Co., Ltd., Beijing, China) was used for aerial observation during the measurement period (Figure 4). Due to different resolutions and other parameters of each lens, individual settings were required. Considering the inconsistency of resolution, the visible light lens used mechanical shutter photography with a flight height of 30 m, with a mainline overlap of 70% and a lateral overlap of 80%. The multispectral lens was set to a flight height of 50 m to obtain data, including blue, green, red, near-infrared, and red edge bands. The thermal infrared lens was set to a flight height of 50 m to capture the canopy temperature of the corn plant. For our experiment, we flew the drone once on a cloudless day during elongation (25 July), tasseling (17 August) and silking period (27 August), respectively.

2.3.2. Ground Data Measurement

Plant height: Throughout the entire corn growth period, representative plants were selected, and soil-to-top-leaf-tip length was measured using a tape measure. To ensure accuracy and reliability, measurements were repeated eight times in each test plot, and the average measurement was taken as the final result.

Stem diameter: While measuring plant height, stem diameter was recorded using a tape measure. Measurements were repeated eight times in each test plot, and the average measurement was taken as the final result.

Leaf Area Index (LAI): Portable plant canopy analyzers (LAI-2200C, purchased from Beijing Ligotai Technology Co., Ltd., Beijing, China) were used to measure LAI in each plot at different time points, with eight repeated observations.

Relative Chlorophyll Content: Soil and Plant Analyzer Development (SPAD) values were measured in each test plot using a chlorophyll content analyzer (SPAD-502Plus, purchased from Konica Minolta (China) Investment Ltd., Beijing, China), with eight repeated observations in each plot.

Plant canopy spectral measurement: A portable spectrometer (Field Spec4 Std-Res) from the American company Analytical Spectral Devices (ASD, Boulder, CO, USA) was used to measure the spectral curve of corn plant canopies, with a bandwidth range of 350–2500 nm. To account for sunlight interference, measurements were taken under clear conditions at noon (10:00–2:00). Prior to observing each plot, whiteboard calibration was performed, and the sensor probe was held vertically downward to measure the corn canopy spectrum. To ensure accuracy, fixed points within each plot were selected for measurement, with eight repeated observations at each point. Spectra affected by noise between 1350 and 1410 nm, 1800 and 1940 nm, and 2450 and 2500 nm were excluded.

Yield data measurement: In each plot, a sample square (1 m × 1 m) was selected, and all corn plants within the square were counted for total ear number and ear grain number. One hundred grains were randomly selected and placed in an aluminum box for drying and water removal in a constant temperature drying oven. After being removed from the oven at 75 °C after 12 h, the dry weight was measured using a balance and recorded. The grains were then put back into the constant temperature drying oven for another hour before recording the second weight. The difference between the two weights had to be less than 0.2%, with the specific calculation formula as follows:

$$Y=E\times G\times W\times 10$$

(1)

where Y represents the yield (kg/km²), E represents the total number of ears, G represents the number of grains per ear, and W represents the weight of 100 grains (g).

2.4. Yield-Related Drought Index (YI)

2.4.1. Index Calculation

In this study, six Vegetation index (VI) models and five Color Index (CI) models were carefully selected to lay a solid foundation for the subsequent construction of yield prediction models.

(1)

Color index

CI is mainly based on previous researchers’ index models and is calculated by normalizing the r, g, and b bands in visible light images [28,29,30,31]. The calculation method is as follows:

$$C{I}_1=2\times g-b-r$$

(2)

$$C{I}_2={\displaystyle \frac{(g^2-r^2)}{(g^2+r^2)}}$$

(3)

$$C{I}_3={\displaystyle \frac{(g^2-r\times b)}{(g^2+r\times b)}}$$

(4)

$$C{I}_4={\displaystyle \frac{(r-g)}{(r+g-b)}}$$

(5)

$$C{I}_5=3\times g-2.4\times r-b$$

(6)

$$C{I}_6={\displaystyle \frac{(2\times g-b-r)}{(2\times g+b+r)}}$$

(7)

where r, g and b represent the red band, green band and blue band.

(2)

Vegetation Index

Considering the numerous vegetation indices used for drought monitoring, we selected five representative index models based on previous work and constructed, to determine six indices that better reflect corn yield for subsequent research [32,33,34,35,36]. The bands (R₁, R₂, R₃, R₄, and R₅) used in the construction process represent any one band in multispectral data; these band numbers are only for differentiation purposes and hold no actual significance, resulting in a total of 125 combinations. The specific model construction process is as follows:

$$V{I}_1=(R_1-R_2)$$

(8)

$$V{I}_2={\displaystyle \frac{(R_1-R_2)}{(R_1+R_2)}}$$

(9)

$$V{I}_3={\displaystyle \frac{\left(R_1\right)}{\left(R_2\right)}}$$

(10)

$$V{I}_4={\displaystyle \frac{1.5\times (R_1-R_2)}{(R_1+R_2+0.5)}}$$

(11)

$$V{I}_5={\displaystyle \frac{1.16\times (R_1-R_2)}{(R_1+R_2+0.16)}}$$

(12)

2.4.2. Data Fusion Method

To comprehensively evaluate the impact of drought on corn, this study breaks through the limitations of traditional drought indices by innovatively developing a Yield-related Drought Index (YI) as a new benchmark. Based on extensive data screening and analysis, ten color and vegetation indices highly correlated with corn yield were meticulously selected as key indicators. Utilizing advanced techniques such as Random Forest (RF), Multiple Linear Regression (MLR), and Convolutional Neural Networks (CNN), these indicators were carefully integrated, aiming to create a “golden key” for accurately measuring the effect of drought on corn yield.

$$YI=\left\{\begin{array}{c}RF(x_1,\cdots ,x_n),RMSE=RMS{E}_{\min }\\ CNN(x_1,\cdots ,x_n),RMSE=RMS{E}_{\min }\\ MLR(x_1,\cdots ,x_n),RMSE=RMS{E}_{\min }\end{array}\right.$$

(13)

Random Forest Algorithm (RF)

The Random Forest algorithm (RF) integrates classification and decision trees, offering precise predictions, a strong ability to identify complex feature relationships, efficient execution, and resistance to overfitting [37,38]. It employs Bootstrap sampling to construct training and out-of-bag (OOB) test sets (7:3), recursively splits the training set to form multiple decision trees, and then aggregates predictions through voting and scoring to obtain optimal results. Based on extensive research by numerous scholars, we have developed a new metric using the Random Forest method. The specific formula is as follows:

$$VI(X_j)={\displaystyle \frac{1}{n}}{\displaystyle \sum _t({err}^{\prime }OO{B}_t^j-errOO{B}_j^t})$$

(14)

where $X_j$ is the j variable, n represents the number of decision trees. t represents the decision tree, and ${err}^{\prime }OO{B}_t^j$ is the OOB error when $X_j$ is arranged in OOB while other variables remain unchanged. $errOO{B}_t^j$ is the OOB error of decision tree t.

Convolutional Neural Networks (CNNs)

Based on the constructed Vis and CIs, six indices with the highest correlation to yield were selected from each section for subsequent CNN modeling to construct a new drought index of the corn yield. The principle of the CNN model built for this paper is outlined below, comprising six main parts: input layer, convolutional layer, batch normalization layer, activation function layer, fully connected layer, and output layer (Figure 5) [39,40,41]. The input layer primarily inputs vegetation indices and color index data, while the convolutional layer, consisting of 3 × 3 convolutions and 64 convolutional kernels, serves to extract features by capturing spatial structural characteristics through local filtering operations. The batch normalization layer is introduced to expedite the training process, reduce gradient vanishing issues, and enhance model stability and reliability. The activation function layer employs the ReLU activation function to introduce non-linearity, enhancing the model’s expressive capability and adaptation to complex data patterns. The fully connected layer primarily connects the preceding and succeeding layers, transforming outputs into a form manageable by the next layer. The regression layer performs regression tasks, comparing model outputs with actual target values and calculating losses to update model parameters, iterating to the simulated drought index. The specific model training parameters are shown in

Table 3. Model training parameter table.

Parameter Name	Model/Value/Function
Gradient Threshold	1
Maximum iterations	500
Optimization function	sgdm
Execution Environment	Auto
Initial Learn Rate	0.03
Validation Frequency	10
Mini Batch Size	128
Learn Rate Schedule	piecewise
Learn Rate Drop Factor	0.9
Learn Rate Drop Period	10

.

Multiple Linear Regression (MLR)

Multiple linear regression (MLR) is a statistical method used to study the linear relationship between a dependent variable and two or more independent variables. The selected advantage index is taken as input data, the yield data is taken as Y, the relationship between them is constructed by using multiple linear regression method (MLR), and it is used for subsequent backward prediction:

$$y=a_0+a_1{x}_1+a_2{x}_2+\cdots +a_n{x}_n$$

(15)

where $a_0$ represents the constant term, and $a_1\cdots a_n$ represents the regression coefficient corresponding to each variable ($x_1\cdots x_n$).

3. Results

3.1. Changes in Physiological Parameters of Corn

The variations in physiological parameters of corn plants at different growth stages under varying drought stress conditions were explored. These stages included the elongation stage (Stage 1), tasseling stage (Stage 2), and silking stage (Stage 3). The red sections in the picture represent different drought treatment periods. The parameters mainly include SPAD (Figure 6a–c), LAI (Figure 6d–f), plant height (Figure 6g–i) and diameter width (Figure 6j–l). Figure 6a–c depict the impact of different drought conditions on plant SPAD values across these growth stages. During the elongation stage, a critical early growth phase, the SPAD value increased from 50.55 to 56.13. Compared to the control group, corn plants subjected to drought stress during this stage showed lower SPAD values, with crops under extreme drought conditions exhibiting significant growth inhibition, starting with an initial SPAD value of only 43.58, and inadequate recovery later, highlighting the importance of sufficient water supply early in growth. Notably, plant SPAD values were higher under mild drought conditions compared to moderate and severe conditions, indicating that mild drought can promote SPAD increase in the early stages of plant growth (Figure 6a).

At the tasseling stage, SPAD values peaked under normal conditions (62.275). While corn exhibited relatively minor effects under mild drought, there was a noticeable delay in reaching peak SPAD values under moderate and severe drought. Plants suffered most severely under extreme drought conditions, with SPAD values ranging from 47 to 51.5, significantly lower than other conditions (Figure 6b). By the silking stage, leaf senescence and chlorophyll content gradually decreased, leading to a decline in corn SPAD values (Figure 6c). Although the impact of different drought intensities was less pronounced, plots under extreme drought conditions observed the lowest SPAD values, followed by severe, moderate, and mild drought conditions. These findings underscore the significant effects of varying drought levels on corn crops, with the elongation stage being most sensitive to drought stress, followed by the tasseling and silking stages.

As for LAI, Figure 6d–f illustrates corn plant LAI responses to different drought stress conditions across growth stages. The results show that, at the beginning of the elongation stage, LAI for corn plots under normal conditions was 1.65, peaking at 3.081 during mid-tasseling stage, and gradually declining to 2.79 in the final growth stage. Significant variations in LAI were observed among plots subjected to different drought intensities. Corn crops exhibited the lowest LAI under extreme drought conditions, at only 0.6, followed by plots under severe, moderate, and mild drought conditions (Figure 6d). Figure 6e demonstrates that corn maintained normal LAI during the elongation stage, with an increase in LAI peaking in the second stage. As drought stress varied during the tasseling stage, LAI ceased to increase and began to decline, notably showing significant reductions under severe and extreme drought, while plots under mild and moderate drought eventually adapted and exhibited recovery. By the silking stage, plants across all water stress conditions displayed similar declining trends in LAI, comparable to normal plots. However, plants under extreme drought conditions exhibited the lowest LAI values, indicating significant impact during this growth stage (Figure 6f).

Under normal conditions, initial plant height at the elongation stage was 92.83 cm, gradually increasing to a peak of 252.42 cm by the end of the tasseling stage. However, plant height remained unchanged during the silking stage as plants transitioned from growth to reproductive development. Figure 6g–i demonstrates significant correlations between corn height and different drought levels during early growth stages, with extreme drought notably affecting plant height more than other drought levels. Despite drought stress during the elongation stage, significant height increases were observed when normal conditions were reinstated during tasseling, indicating corn’s potential for recovery (Figure 6g). During the tasseling stage, plants under mild drought conditions even surpassed normal plots in height (Figure 6h). However, under extreme and severe drought conditions, plant height ceased to increase, plateauing at 215 cm and 234 cm, respectively, due to the effects of drought stress. Similar effects of drought stress were observed during the silking stage, where the impact was minimal, as plants had largely reached maximum height and ceased further growth naturally (Figure 6i).

Figure 6j illustrates plant stem width under normal growth conditions, starting at 2.15 cm and peaking at 2.76 cm during the tasseling stage. During early growth stages under different drought levels, no significant impact on stem width was observed. However, as corn entered the tasseling stage, stem width began to decrease under various drought stresses. At this stage, extreme, severe, and moderate drought had nearly identical effects on stem width, with extreme drought conditions showing greater susceptibility. Differences were notably significant between mild drought and other drought stress conditions (Figure 6k). When the tasseling stage ended and normal conditions resumed, corn stem width did not increase further. Figure 6i further highlights the impact of drought stress during the silking stage, where plots exposed to extreme, severe, and moderate drought experienced reduced stem width. However, mild drought had no significant impact on corn stem width during plant growth.

3.2. Physiological Parameters Impact by Drought Stress

Research on the impact of different drought intensities on physiological parameters of plants at different growth stages indicates that drought has the greatest impact on plant growth during the elongation stage and tasseling stage while, under extreme conditions, the SPAD and LAI during the silking stage also suffer significant effects. As shown in Figure 7, the average SPAD loss rate during extreme drought is highest at 16.74% in the tasseling stage, followed by the elongation stage at 14.07%, and 10.11% during the silking stage (Figure 7a–c). The sequence of SPAD loss under mild, moderate, and severe drought is the same, indicating a significant impact of extreme drought during the tasseling stage.

LAI loss under extreme drought conditions (Figure 7e) is highest during the second stage of plant growth (40.48%), with a nearly identical response during silking (36.7%) (Figure 7f), while the impact of extreme drought on LAI during the elongation stage is relatively smaller (9.41%) (Figure 7d). The magnitude of SPAD and LAI loss at different growth stages under different drought stresses follows the order of extreme drought > severe drought > moderate drought > mild drought. This study further reveals the sensitivity of plant height to drought conditions during the elongation stage. Due to extreme drought, plant height decreased by 36.27%, with significant average losses also caused by severe, moderate, and mild drought, at 32.78%, 29%, and 23.62%, respectively (Figure 7g). The impact of drought stress during the silking stage on plant height is not significant, as plants cease further growth (Figure 7i), but under extreme drought conditions, the greatest height loss occurs during the tasseling stage, at 17.76% (Figure 7h). The effect of different drought intensities on plant stem width is also examined. The results indicate that the tasseling stage is the period when plants are most susceptible to drought, with plant stem width loss of 17.47% under extreme drought, followed by severe drought (16.56%), moderate drought (15.16%), and mild drought (11.11%) (Figure 7k). During the elongation stage, the average stem width loss under moderate and mild drought conditions is greater than under extreme and severe drought conditions, possibly due to the influence of certain external factors (Figure 7j). During the silking stage, extreme drought, severe drought, and moderate drought have significant impacts on stem width, with average width losses of 12.89%, 11.09%, and 9.03%, respectively (Figure 7l).

3.3. Canopy Spectral Characteristics

A spectrometer was used to track canopy reflectance of corn from elongation stage to silking stage under different levels of drought stress, as shown in Figure 8a–c. The spectral distribution under different drought stress conditions is basically consistent, showing peak–valley variations. In the visible and shortwave infrared bands, drought severity is positively correlated with reflectance while, in the near-infrared band, drought severity is negatively correlated with reflectance. It is worth noting that a small reflectance peak is observed at 550 nm in the visible light band, attributed to the presence of chlorophyll, with a decrease in chlorophyll content leading to weakened green visual effects under stress. In the near-infrared band, leaf cell structure mainly influences spectral reflectance, with weak water absorption bands near 970 nm and 1200 nm. Meanwhile, in the shortwave infrared band, the spectral characteristics of plants are mainly dominated by total leaf water content, with absorption troughs at 1450 nm and 1950 nm and reflectance peaks at 1650 nm and 2200 nm. The reflectance levels in different bands follow the sequence: normal > mild drought > moderate drought > severe drought > extreme drought. During the elongation stage, corn canopy reflectance is lowest while, under mild drought stress, canopy reflectance is higher, likely due to corn plants not fully expanding their leaves, resulting in relatively less surface area to reflect light, and some degree of drought stress can increase plant reflectance. Reflectance peaks during the tasseling stage, attributed to increased chlorophyll content, improved photosynthetic efficiency, and denser vegetation canopy formation. In the near-infrared region, although the normal spectral curve is slightly higher than the drought stress curve, extreme drought stress leads to a sharp decrease in corn canopy reflectance, indicating a significant impact of severe drought during the tasseling stage. Additionally, during the silking stage, drought stress of moderate severity or higher induces significant changes in corn canopy spectra, suggesting a greater impact of moisture on silking stage corn canopy spectra.

3.4. The Relationship Between Parameters and Yield

To investigate the extent to which physiological parameter can represent yield, a correlation analysis between crop yield and SPAD (Figure 9a), LAI (Figure 9b), plant height (Figure 9c), and stem width (Figure 9d) was conducted. The results indicate that LAI has the highest explanatory power for yield and shows a significant positive correlation (R² = 0.6061, p < 0.05) (Figure 9b). SPAD is also a good indicator for reflecting yield to some extent, expressing yield levels (R² = 0.3981, p < 0.05). This is because crop yield is closely related to photosynthetic capacity (leaf area of green leaves); while leaf chlorophyll is one of the key factors in photosynthesis, SPAD only reflects one aspect of photosynthetic capacity. As for morphological traits, the relationships between plant height and stem width with yield are not close, with relatively low correlations, R² = 0.1320 and R² = 0.1366, respectively. This may be because these indicators mainly reflect the growth stage’s vigor, while yield is more related to the reproductive development stage, resulting in lower correlation.

3.5. Construction of Drought Index

To better select indicators that can represent yield, five representative index models were individually incorporated into different bands for mixed selection. Figure 10a–e shows the correlation heatmaps between vegetation indices (V1–V5) constructed from different bands and corn yield. Overall, each of the five index models has its advantages and disadvantages, with a tendency towards selection of certain bands to a certain extent. Figure 10a indicates that, in the V1 vegetation index construction model, the combination of green and red bands has the highest correlation with yield (R = 0.48), followed by the combination of red and red edge bands, which also has a high correlation with yield (R = 0.46). In the V2 model (Figure 10b), the combination of green and red edge bands has the poorest representativeness (R = 0.20), while the index formed by the combination of blue and green bands can better represent yield (R = 0.62); a similar pattern is observed in other models. For the V3 model, the combination of green and red edge bands has the poorest representativeness (R = 0.23), while the index constructed from the combination of green and blue bands has the strongest representativeness (R = 0.65) (Figure 10c). Additionally, in the V4 model, the combination of green and red bands has the most representativeness (R = 0.48), while the combinations of blue and green bands, and infrared and green bands have the poorest representativeness (R = 0.15) (Figure 10d). Furthermore, Figure 10e shows that the representativeness of indices formed by the blue band and other bands are in the following order: Green < Red < NIR < Red Edge, with correlation coefficients of 0.25, 0.31, 0.37, and 0.53, respectively.

3.6. Accuracy Assessment of Indicators

To evaluate the accuracy of the prediction results, MLR and RF methods were simultaneously employed to simulated drought index for comparative analysis. Concerning input data, it was primarily categorized into four types: using only multispectral data, using only visible light data, integrating visible light and multispectral data, and combining visible light, multispectral, and thermal infrared data. Figure 11a–c represent the results of CNN, RF and MLR, respectively. The results indicate a close correlation between the prediction accuracy of the polynomial regression method and the amount of input data used (Figure 11b). The accuracy of predicting corn yield was lowest when using only visible light data (R = 0.4848, p < 0.1). However, as the amount of data input increased, the prediction accuracy improved. The highest accuracy was achieved when combining visible light, multispectral, and thermal infrared data (R = 0.8505, p < 0.05).

In terms of the Random Forest regression method, the accuracy of predictions was lowest when using only visible light, with an R value of only 0.5080 (p < 0.1). The highest prediction accuracy was achieved when integrating visible light, multispectral, and thermal infrared data (R = 0.6402, p < 0.05), although it was relatively lower compared to other methods (Figure 11c).

From the perspective of CNN methods, the highest prediction accuracy was obtained when integrating visible light and multispectral imagery, showing the highest significance (R = 0.9332, p < 0.001), while using only visible light resulted in the lowest accuracy (R = 0.5617) (Figure 11a).

Overall, in predicting corn yield, the CNN method based on visible light and multispectral data demonstrated the highest accuracy, while the Random Forest method showed relatively poorer accuracy. Through this series of meticulously designed experiments and comparisons, we successfully developed a novel and efficient drought index (YI) that is intimately related to yield. The birth of this index marks a significant breakthrough in the field of drought monitoring and assessment, providing a powerful scientific basis for early warning of agricultural drought disasters, precise prevention and control, and post-disaster assessment. It holds profound implications for safeguarding national food security and promoting sustainable agricultural development.

3.7. Spatial Distribution of Yield

To achieve rapid estimation of corn yield, five bands in multispectral and three bands in visible light, as the basis, were utilized, cross-selecting the optimal six VIs and six CIs, combined with CNN methods to construct the drought index and predict corn yield. Since output is an important validation index that can be obtained, the YI index was used to invert output in order to express the construction result of the index. The results, as shown in Figure 12a, combined with the analysis of drought stress, indicate that the yield of the top four reference plots is the highest, with an average yield of 15.2686 kg per plot. During different growth stages, corn experiences the most significant impact on yield during the second stage due to water deficiency, with an average yield reduction of 24.99%. Extreme drought stress causes the largest reduction in yield, reaching 32.43%; corn is affected to a lesser extent during the silking stage, with an average reduction of 19.77%, and the impact on yield during the elongation stage is minimal, with an average reduction of 16.46%, whereas mild drought has the smallest impact on yield, at only 2.69%. At the same time, the index was normalized and graded planning was carried out according to the field situation (Figure 12b,c). The YI obtained is shown in Figure 12b. YI signifies that, the higher the value, the better the growth state of the plant. At the same time, according to the principle of plot division, we graded YI and obtained different drought degree thresholds (Normal: 0.91–1.0, Slight drought: 0.77–0.91, Middle drought: 0.63–0.77, Severe drought: 0.50–0.63, Extreme drought: 0–0.5). According to the results, the drought conditions of the samples planned by our experiment are consistent. Therefore, agricultural workers should consider local conditions, corn irrigation frequency, and water volume to formulate targeted irrigation and emergency measures in order to mitigate the adverse effects of drought on corn yield.

4. Discussion

This study, based on field-controlled experimental methods, revealed the differential impacts of drought stress on corn growth and development, emphasizing key stages of vulnerability and resilience. Additionally, it elucidated the spectral characteristics of corn canopy under drought stress, providing a foundation for remote sensing-based monitoring and management strategies. In our research, we found that mild droughts can promote rapid growth of corn. A similar phenomenon was also found in the field experiment of Liu et al. [14]. In terms of physiological parameters, with the change in growth period, the internal physiological parameters of maize will also change. In addition, drought resulted in a decrease in LAI and SPAD content. This is because drought can affect the photosynthesis of leaves and even lead to changes in the internal structure of cells [42,43]. For the construction of yield index, the results show that the newly constructed index (YI) has high accuracy and can be used for yield prediction and drought identification. Compared with the research on yield prediction applied to data products derived from satellites, the yield prediction based on drone data is more accurate [26,27,44]. This is because drones have higher resolution and can characterize richer details. This better matches the actual measured yield data. This highlights the superiority of YI, and the raw data of this indicator can be matched with the data of the satellite for further application to the satellite. This could enable large-scale yield forecasting and drought prevention. In conclusion, this research contributes to enhancing understanding of adaptation strategies of corn to drought stress and provides practical significance for improving resilience and productivity of crops in drought-prone areas.

However, there are several limitations in the experimental process and execution. In our actual measurement process, although we strictly follow the measurement operation specifications, considering the problem of manual measurement indicators there will be some human errors. At the same time, the yield measurement process also produces some errors, which may be the reason for the low correlation results in Figure 8.

In terms of data measurement, four representative indicators (SPAD, LAI, plant height, and stem width) were utilized to monitor changes in corn characteristics under different periods and levels of drought stress. While these indicators involve internal physiological features, clarity regarding different enzyme activities and physiological biochemical processes within plants remains elusive. Future research should focus on comprehensive monitoring and analysis of corn responses to drought processes from internal physiological and external phenotypic perspectives [45].

Regarding spectral measurement, ground-based spectrometers were employed to measure canopy spectral characteristics of corn at various growth stages, leading to differences in measurement time among different plots and variations in solar irradiance at different times, resulting in certain errors. It would be preferable to utilize drones equipped with hyperspectral lenses for measuring crop canopy to minimize errors caused by reflectance. Additionally, there are differences in resolution among visible light, multispectral, and thermal infrared lenses carried by drones, and they cannot operate simultaneously. Errors also exist in the image matching and processing. Future research should employ cameras with consistent resolutions and integrate the three types of cameras to perfectly match time points and reduce errors.

In the realm of model, there are two key shortcomings. The primary issue lies in the inherent limitations of the model itself. In our study, we utilized multispectral and visible light data to construct a drought index for yield prediction, where the CNN model demonstrated its exceptional feature extraction capabilities, particularly in processing complex spectral information. It rapidly identified and correlated spectral features with yield variations, exhibiting stronger representativeness compared to traditional models, such as multiple linear regression or principal component analysis. For instance, Chhabra et al. [26] used principal component analysis to predict corn yield with an R² value of 0.7418, while Joshi et al. [27] showed that deep learning models outperformed stepwise multiple linear regression in predicting corn yields in the Midwestern Corn Belt of the United States, aligning with our validation results. Furthermore, the application of multi-source data in predicting yields of other crops further confirmed the high accuracy of deep learning models, validating the accuracy of our research [46]. However, the CNN model is not without flaws; its high dependence on the quality and scale of the dataset, along with the substantial computational resource consumption during model training, pose significant challenges.

On the other hand, the deficiencies in the model training process cannot be overlooked. Specifically, our study only used plot data within the study area as training samples, treating each plot as an individual data point, resulting in a limited sample size and severely insufficient data input for the CNN model. This directly weakened the model’s simulation capabilities, potentially hindering its full performance. The scarcity of samples limited the model’s ability to learn features and patterns from the data, thereby affecting the accuracy and reliability of drought monitoring and yield prediction. To address this challenge, we plan to expand the sample scope in future research by incorporating data points outside the study area during the experimental process to increase the number of plot samples. We hope to enhance the training effectiveness and prediction performance of the model by increasing the number of observation maps and adopting more refined data collection strategies to obtain a richer and more representative dataset.

On the scale issue, given the broad applicability and versatility of the data sources utilized in our newly developed drought index, and the fact that most satellites are equipped with sensors capable of capturing these necessary spectral bands, we intend to actively explore and implement the application of this drought index to a wide range of satellite datasets in our future research endeavors. This initiative aims to fully leverage the advantages of satellite technology, such as its extensive spatial coverage and rapid data updates, to achieve large-scale, high-precision mapping of drought conditions. This will provide robust support for drought monitoring, early warning, and impact assessment [47,48].

Moreover, when conducting yield predictions at the satellite scale, the CNN model demonstrates broad application prospects, capable of efficiently processing high-resolution satellite imagery and rapidly extracting key features related to crop yields. However, issues such as high data acquisition costs and limited model generalization capabilities still need to be considered. To address these shortcomings, we plan to continuously optimize the model structure and algorithms, as well as expand the scale and diversity of the dataset, to improve the accuracy and reliability of the model. This will enable it to better adapt to the actual needs of different regions, crops, and drought conditions, providing more robust support for agricultural drought monitoring and yield prediction [49]. Through ongoing efforts, we firmly believe that we can provide more precise and reliable decision-making tools for agricultural production, promoting sustainable agricultural development. This would enable better monitoring of crop physiological status, facilitate local authorities’ understanding of the impact of drought stress on corn yield, help devise effective coping measures to mitigate agricultural losses due to drought, and maintain local food security and agricultural stability.

5. Conclusions

By controlling different levels of drought stress at different growth stages, the physiological, phenotypic, and spectral characteristics of corn in response to drought were explored, and multiple indicators and fusion methods were combined to predict corn yield. The specific conclusions obtained are as follows: (1) In terms of physiological and phenotypic parameters, different levels of drought stress exhibit a gradient effect on corn growth and development, with the greatest impact during the elongation stage and tasseling stages, and a more significant effect on SPAD and LAI during the silking stage, but with minimal impact on plant height and stem width. (2) Regarding spectral characteristics, under different levels of drought stress, corn canopy reflectance from the elongation stage to the silking stage exhibits consistent spectral distribution characteristics, showing peak-valley variations. Drought severity is positively correlated with reflectance in the visible light and shortwave infrared bands, while negatively correlated in the near-infrared band. Reflectance is lowest during the elongation stage and peaks during the silking stage. Corn canopy spectra during the silking stage are greatly affected by moderate to severe drought stress. (3) In terms of the relationship between parameters and yield, LAI shows a significant positive correlation with yield (R² = 0.6061, p < 0.05), providing the highest reliability in explaining yield. (4) The new yield-related drought index (YI) constructed based on the multi-spectral index and visible light index had the highest accuracy in expressing drought (R = 0.9332, p < 0.001), and was more spatially consistent with experimental control results.