Utilising the Potential of a Robust Three-Band Hyperspectral Vegetation Index for Monitoring Plant Moisture Content in a Summer Maize-Winter Wheat Crop Rotation Farming System

Highlights

What are the main findings?

• Normalised water stress index (NWSI), a new three-band index has high potential for better monitoring of drought in winter wheat-summer maize fields.
• Combining these new indices (NWSI and NDI) with the traditional moisture stress monitoring indices improves monitoring accuracy.

What are the implications of the main findings?

• The new three-band indices provide good options for accurate plant moisture stress monitoring in winter wheat-summer maize rotation systems.
• The NWSI and NDI, combined with traditional moisture stress monitoring indices, lay a scientific basis for precision irrigation.

Abstract

Water is vital for producing summer maize (SM) and winter wheat (WW); therefore, its proper management is crucial for sustainable farming. This study aimed to develop new tri-band spectral vegetation indices that enhance the accuracy of monitoring plant moisture content (PMC) in SM and WW. We conducted irrigation treatments, including W0, W1, W2, W3, and W4, in SM–WW rotations to address this issue. Canopy reflectance was measured with a field spectroradiometer. Tri-band hyperspectral vegetation indices were constructed: Normalised Water Stress Index (NWSI), Normalised Difference Index (NDI), and Exponential Water Stress Index (EWSI), for assessing the PMC of SM and WW. Results indicate that NWSI outperformed other indices. In the maize trials, the correlation reached R = −0.8369, while in wheat, it reached R = −0.9313, surpassing traditional indices. Four mainstream machine learning models (Random Forest, Partial Least Squares Regression, Support Vector Machine, and Artificial Neural Network) were employed for modelling. NWSI-PLSR exhibited the best index-type performance with an R² of 0.7878. When the new indices were combined with traditional indices as input data, the NWSI-Published indices-SVM model achieved superior performance with an R² of 0.8203, outperforming other models. The RF model produced the most consistent performance and achieved the highest average R² across all input types. The NDI-Published indices models also outperformed those of the published indices alone. This indicates that these new indices improve the accuracy of moisture content monitoring in SM and WW fields. It provides a technical basis and support for precision irrigation, holding significant potential for application.

1. Introduction

Agricultural drought remains one of the most widely occurring natural disasters, hindering crop production worldwide [1,2]. Thus, its impact remains a challenge to achieving global food security [3], especially in arid and semiarid regions [4]. Agricultural droughts are usually characterised by crop water stress. Crop water stress occurs when plants lack adequate moisture for normal functioning. The traditional approaches to plant water monitoring are time-consuming, costly, destructive, temporally and spatially restricted, and labour-intensive. An effective resolution to these challenges is the deployment of appropriate remote and proximal sensing equipment, which provides a more rapid, cost-effective, extensive, and non-invasive method for evaluating the physiological conditions of crops [5,6]. Over the years, these technologies have made huge impacts in the monitoring of plant physiological traits such as nutrient status, water content, chlorophyll [7], pigment, and carotenoid contents, and have immensely improved precision agriculture [6,8], thus improving water stress management. With all these improvements, coupled with the considerable potential of hyperspectral remote sensing equipment, it is noteworthy to know that the full potential of this equipment has not been fully utilised. As such, there are still considerable opportunities to explore other potential uses of this technology further to improve its performance [9]. These include dimensionality reduction, data preprocessing, machine learning, and the creation of new indices from narrow-band spectral [10,11,12].

In the past, researchers have experimented with different crops and water availabilities to monitor biophysical and biochemical plant characteristics remotely [13,14,15,16,17,18,19,20]. Since different plants have various regions of the spectral wavelength sensitive to specific biophysical and biochemical physiology, and there are so many narrow bands in hyperspectral data, carefully screening useful bands to monitor a particular feature is highly recommended to achieve the goal of precision agriculture [21]. For instance, Zhang et al. [22] reported that a normalised drought index constructed with wavelengths 1287 nm and 1673 nm was the optimal input data for the XGBoost model for predicting rice leaf water content (LWC). They stressed the importance of coupling multisource data to monitor water content. Zhang & Zhou and Yang et al. [23,24] reported that there is a good correlation between the visible (VIS) and wheat water content. With these findings, it becomes a paramount concern to correctly identify spectral regions sensitive to biophysical and biochemical features of plants. This will remove redundant narrow bands, improve model performance and reduce the required time to handle model training.

As prevailing droughts are predicted due to increasing extreme climate variability, a further understanding of various techniques for remotely acquiring meaningful plant information is crucial for the sustainable production of summer maize and winter wheat, a primary food source and source of income in the North China Plain (NCP) and globally.

Researchers have applied machine learning to both raw canopy reflectance spectral data and vegetation indices calculated from canopy reflectance [22,25,26,27,28] to monitor plant water content effectively. This has improved irrigation scheduling, plant nutrient management, and yield production. All these works reported that plant water content significantly influenced canopy spectral reflectance throughout the growth seasons and at specific growth stages. This necessitates the study of applying spectroradiometers, which will be a handy tool for farm-scale monitoring of plant water stress.

Absorption and reflectance in the VIS are influenced by chlorophyll and other pigments in the leaf [29]. A reduction in moisture content reduces chlorophyll and pigments, thereby affecting reflection [30]. Variations in leaf water content are known to affect the internal leaf structures, canopy structures [31], and leaf area index, thus impacting reflectance in the NIR region, which makes NIR helpful for crop water monitoring [32]. The SWIR region has been reported to be sensitive to plants’ internal moisture. According to Rapaport et al. [33], a reduction in plant moisture influences reflectance in the SWIR spectrum by increasing reflectance. They also noted that SWIR bands are capable of detecting other stress-induced changes in plants. Further, Braga et al. [34] also confirmed the ability of SWIR bands to differentiate between different moisture stress levels of soybean genotypes subjected to varying moisture stress. Since water strongly absorbs light in the SWIR spectral region, a decrease in plants’ internal moisture reduces this ability, thereby increasing reflectance [35]. Other works [14,36] have identified the NIR and SWIR as most appropriate for such functions. Based on these findings, our work aims to incorporate the VIS, NIR, and SWIR spectral bands into VI calculations for monitoring plant water stress. A combination of bands from several regions of the spectrum, spanning wavelengths, is used to calculate vegetation indices for monitoring plant water content, making it a helpful tool for assessing plant water content non-destructively and in a timely manner. This also helps to capitalise on the individual bands and regions’ comparative advantage for monitoring plant moisture content (PMC) in varying environments with different climatic conditions [37].

Furthermore, the time interval from the onset of moisture stress to its detection and remedy significantly impacts plants [38]. Remote sensing has enabled the early detection of these occurrences. Elsayed & Darwish, and Mndela et al. [39,40] have also reported the potential of VIs calculated from remotely sensed data to monitor water content at the leaf and canopy levels, using experiments on different crops.

Additionally, in a changing environment where plants are grown, it has become more evident that several factors affect plant biophysical and biochemical characteristics. As such, several approaches are being devised to improve the monitoring of plant water stress. To tackle the growing environmental impact on canopy reflectance while utilising the maximum potential of hyperspectral data, new band combinations have been reported [41,42,43], which led to an improvement in the prediction accuracy of targeted plant biophysical attributes. Li et al. [42] reported the combination of multispectral and hyperspectral data collected with an unmanned aerial vehicle from summer maize and winter wheat to predict fuel moisture content with the formulation of new three-band combinations concentrated in the VIS and NIR spectra (400–1000 nm) without including the SWIR bands and NIR bands beyond 1000 nm. Our work proposes three band indices for plant moisture monitoring, including a constant term and the SWIR bands in the calculation formula, which are lacking in Li et al. [42]. Zhang et al. [22] also reported the use of a new two-band combination index to monitor leaf moisture content of rice during the reproductive stage. Their work focused on a single crop and a single growth stage and highlighted the importance of new band formations to improve leaf moisture monitoring in rice.

Also, there are reports that vegetation indices are sensitive to specific biophysical and biochemical features of plants at specific growth stages [44] and within different environmental locations [4], thus limiting the global application of these VIs. These reports then highlight the need for site- and growth-stage-specific indices which will account for the environmental effects on spectral reflectance. Tian et al. [43] highlighted the superior performance of three-band indices over two-band indices for monitoring leaf nitrogen content in rice. They suggested the newly constructed index (R705/(R717 + R491)) as the most suitable index. These reports suggest that combinations beyond the traditionally recognised bands for monitoring plant biophysical and biochemical properties can produce better results. Building on these findings to improve the plant moisture monitoring capability of spectral vegetation indices, this work incorporates the visible, near-infrared, and shortwave-infrared bands. Further, it introduces a constant term in the calculation formula, which will help stabilise these indices further.

With all these reports mentioned here for the new band combinations, Tian et al. [43] used a spectroradiometer to predict plant nitrogen. For the others, where plant moisture was the research target [41,42], none have reported using a spectroradiometer to collect canopy reflectance for creating indices and building a model that simultaneously monitors the water content of summer maize and winter wheat during their growth stages. Utilising spectroradiometers for such functions will introduce higher resolutions, add the SWIR bands that are highly sensitive to plant moisture, and alleviate the high technical capabilities required for UAV operation to achieve highly accurate results. These spectroradiometers also provide additional advantages, like laboratory applications [8] and the leaf clip usage [45], which prevents weather condition restrictions that influence both UAV and satellite data collection. These qualities ensure timely data collection as per research requirements and needs with the use of spectroradiometers. These quality additions are somewhat lacking in the UAV data, which positions the spectroradiometers as a suitable choice over the UAVs for our work.

Based on the information mentioned earlier, this work aimed to establish the potential of spectroradiometer-observed canopy reflectance data to construct tri-band spectral vegetation indices that will improve the PMC monitoring within a summer maize and winter wheat crop rotation system. This will provide a handy solution to summer maize and winter wheat farmers struggling with unpredictable climatic occurrences, especially drought, and contribute to the knowledge base of remote crop water monitoring.

2. Materials and Methods

2.1. Experiment Site and Treatments

The study area is in the Farmland Irrigation Research Institute Comprehensive Experimental Station in Qiliying, Chinese Academy of Agricultural Sciences, located in the North China Plain (NCP) at 35°18′11′′N and 113°55′34′′E and 81 m above sea level. The region has a warm, subtropical climate with an average annual rainfall of approximately 573 mm, with 65 to 70% of this rain falling during the summer months from June to September. The experiment site has an average annual solar radiation of 4900 MJ m⁻² yr⁻¹ and an average temperature of 14.5 °C, respectively, with between 189 and 240 frost-free days per year. The site has a weather station for recording daily weather results (temperature (°C), wind speed (ms⁻¹), net radiation, relative humidity (%), rainfall (mm), etc.) and averaging them every thirty minutes. Soil bulk density, porosity, and field capacity are shown in

Table 1. This table shows the soil characteristics of the experiment site.

Soil Depth Range (cm)	Soil Bulk Density (g/cm−3)	Field Capacity (%)	Porosity (%)
0–20	1.28	30	42.5
20–40	1.54	21.9	40.2
40–60	1.56	22.4	40.8

, while the study area is presented in Figure 1.

The experiment consisted of five irrigation treatments with three replications each (15 plots), designed to maintain soil moisture within specified lower and upper limits. To achieve this, irrigation was triggered for treatments when the lower limit soil moisture levels were reached, and the ranges were set as follows: no irrigation (W0), 45–65% (W1), 55–75% (W2), 65–85% (W3), 75–95% (W4) of field capacity. Irrigation was applied through drip irrigation. Drip lines were placed 0.6 m apart with emitter spacing of 0.3 m for the summer maize experiment. The summer maize was planted on 11 June 2024, with an in-row spacing of 0.6 m and a between-plant spacing of 0.22 m. Soil samples were collected every 5–8 days to calculate soil moisture using the oven drying method to ascertain irrigation time. The irrigation treatments started on 6 July 2024. The first irrigation was applied to the W4 treatment and was gradually followed by the lower water limit treatments as their lower soil moisture limits were reached. The total maize irrigation per treatment was 0 mm, 129 mm, 166 mm, 276 mm and 300 mm for W0, W1, W2, W3, and W4, respectively. Irrigation was triggered when soil moisture reached the lower limit for each treatment. It was harvested on 2 October 2024, and the variety of maize planted was WeiKe702. The winter wheat was sown on 18 October 2024 and harvested on 29 May 2025, with the same experimental treatment and design as the summer maize. Winter wheat was sown at 0.2 m line spacing. The plot size was 3.4 m by 2.03 m (6.902 m²). The planted variety was Xinmai32. The irrigation drip lines were placed at 0.4 m with 0.3 m emitter spacing. The experimental field features a weather station that collects and averages weather data every thirty minutes. The soil moisture was monitored using the oven drying method, where soil samples were collected and weighed to determine the wet soil weight. The samples were then dried in an oven at 105 °C for a minimum of 24 h and reweighed to determine the soil dry weight and moisture content. Winter wheat irrigation treatment started during the regreening stage on March 6, and the total irrigation per treatment was 0 mm, 248.66 mm, 300.53 mm, 346.65 mm and 356.36 mm for W0, W1, W2, W3, and W4, respectively. Irrigation amounts per irrigation campaign were measured by flow metres attached to the irrigation lines of each replicate plot.

A compound fertiliser, Royal Fertiliser 25-10-10 (N + P₂O₅ + K₂O ≥ 45.0%, B ≥ 0.3%, Zn ≥ 0.05% Green Intelligent Corn Fertiliser) was applied as per local recommendations at sowing as a basal application, and nitrogen fertiliser was applied at jointing following local recommendations for summer maize. A compound fertiliser (N + P₂O₅ + K₂O, 18-18-6) was applied at sowing, and nitrogen fertiliser was used at the jointing stage, following local recommendations for winter wheat.

2.2. Spectral Data Measurement

With a sampling interval of 2.8 nm at 350 nm to 700 nm, 8 nm at 700 nm to 1500 nm, and 6 nm at 2100 nm, measurements were made across a spectral range of 350–2500 nm, as reported by Beegum et al. [46] using the PSR+ 3500 Spectroradiometer (Spectral Evolution, Lawrence, MA, USA) to measure canopy spectral data. Reflectance was measured at a vertical height of 1.0 m above the summer maize and winter wheat canopy, thus providing a 25° field of view. The spectroradiometer’s final output was a 1 nm increase, giving an output of 2151 narrow bands. Spectral measurements were conducted on clear-sky days between 1000 and 1400 Standard China Time. The instrument was calibrated using a 0.14 m² white calibration reference panel with a 99% reflectance, as described by Zhao et al. [47], before the first measurement, and repeated every 15 min. To capture better reflectance, the equipment averaged twenty-five measurements per scan, and every plot was scanned four times at different points. The four scans were averaged to represent the plot’s spectral reflectance per measurement day. Data were collected for summer maize on 28 July, 5, 13 August, 1, 16 September, and 2 October. A six-foot ladder was used to provide an elevated platform to collect the canopy reflectance of summer maize. The winter wheat season data (canopy spectral reflectance, soil moisture, and plant moisture content) were collected on 17, 25 March, 2, 10, 18, 26 April, and 4, 12, 18 May. Canopy reflectance spectra were collected at a 1 m vertical distance above the canopy.

2.3. Plant Water and Soil Moisture Content Measurement

After spectral data collection, a single maize plant and five representative wheat plants were collected for water content measurement, with one plant per replicate. The winter wheat samples included all tillers from one germinated seed. The fresh maize plants were separated into leaves and stems, and their weights were recorded immediately using an electronic balance with a 0.01 g error. The above-ground wheat plants were placed in a sealable plastic bag and weighed in a similar format to the maize. The summer maize plant weight was calculated by summing the weights of the leaves and stem. After weighing, the samples were taken to the oven for drying. Samples were dried at 105 °C for thirty minutes, followed by 75 °C until constant weight was achieved. Plant moisture content was calculated using Equation (1). During data collection campaigns, soil samples were simultaneously collected to analyse the gravimetric soil moisture content. The workflow of this work is presented in Figure 2.

$$PMC(\%)=\left({\displaystyle \frac{FWP-DWP}{FWP}}\right)\ast 100$$

(1)

where PMC represents the plant moisture content, which in this work refers to the above-ground part of the plant, FWP and DWP denote the fresh weight of plants and the dry weight of plants, respectively.

2.4. Construction of Novel Hyperspectral Indices

The crop data was pulled together, and spectral band selection was conducted. This work used three new band calculation formulas to form new vegetation indices with the target of including wavelength (350–2500 nm), as seen in the following Formulas (2)–(4). To select the best band combinations, initial simulation calculations were performed at 100 nm and subsequently refined at 10 nm to minimise computational time while maintaining data quality and accuracy. The band combination included all possible band combinations at a 10 nm wavelength. For Equation (2), constants were simultaneously added to Equation (2) to create an index that would overcome saturation and other challenges. For the constant addition approach, the range was set to 0–2, and adjustments were made to the values in increments of 0.1 while holding the other factors constant until the optimal values for each band were achieved. After each constant was adjusted, the resulting index values were correlated with PMC, and the R² values were recorded. Final comparisons were conducted to verify which set of constant terms produced the best correlations by observing the R² values. Further, the band combined for each index was saved. The final constants that produced the best coefficient of determination (R²) with PMC were a, b, and c, respectively, in Equation (2) for band1, band2, and band3. These values were then added to the calculation formula in Equation (2). Equation (3) is an extension of the concept applied by Li et al. [42]. Here, we included the SWIR bands, which were not part of their work. A combination of all three bands at 10 nm was calculated, and the resulting outputs were correlated with PMC; the R² values were then ranked. This work selected the best band combinations for each formula for further analysis. The formulas used in this work for calculating the new three-band combinations are Equations (2)–(4).

$$NWSI={\displaystyle \frac{B1-B2-B3}{B1\ast a+B2\ast b+B3\ast c}}$$

(2)

$$EWSI={\displaystyle \frac{B1-B2}{B1+B2-2\ast B3}}$$

(3)

$$NDI={\displaystyle \frac{B1-B3}{B2-B3}}$$

(4)

where B1, B2, and B3 are spectral wavebands whose reflectance values were used to calculate the spectral vegetation index; NWSI is the Normalised Water Stress Index, and NDI is the Normalised Difference Index. EWSI is the Exponential Water Stress Index. The a, b, and c in Equation (2) are simulated constants whose values are 0.5, 0.5, and 1, respectively.

Nine published spectral vegetation indices were calculated to compare the performance of the new index with that of the traditionally used indices at critical growth stages of summer maize and winter wheat. These selected indices are based on the proven effectiveness in their ability to monitor moisture stress in plants. They are designed from remotely sensed remote sensing data to capture various aspects of plant health, growth, land cover, and internal plant moisture. For example, Wang et al. [48] in their work reported the high correlation between the normalised difference vegetation index (NDVI) and plant moisture stress, confirming its use to track moisture stress, plant growth and yield. The NDVI has been referenced as one of the most widely used vegetation index for moisture studies for its effectiveness in monitoring vegetation health, which is directly related to available moisture content as a growth control parameter [49]. Thapa et al. in their work [50] highlighted the importance and applicability of NDVI in winter wheat moisture stress monitoring. Gao in 1996 [51] introduced the normalised difference water index (NDWI) as a complementary index to NDVI with specific advantages in moisture monitoring, and has since been widely used in moisture monitoring studies. Ihuoma et al. [52] reported the high sensitivity of water index (WI), optimised soil-adjusted vegetation index (OSAVI) and WI/NDVI to moisture stress in plants. In other reports, Yang et al. [53] reported the suitability of using red-edge normalised index vegetation index (RE-NDVI) and simple ratio water index (SRWI) for monitoring plant moisture stress. Zhang & Zhou in their work [24] also confirmed the sensitivity of moisture stress index (MSI), WI, and NDWI to different plant moisture classifications (canopy, leaf and equivalent water thickness). The water band index (WBI) has, in studies, been referenced as capable of revealing plant moisture conditions [54]. These findings reported by these works confirm the usefulness of these indices in moisture monitoring. Therefore, this informed our decision to select these indices for comparison to the newly created indices.

Table 2. This table shows the published vegetation indices used in the study for performance comparison with the new indices, with their formula, abbreviations, and references.

Name	Abbreviation	Formula	Reference
Normalised Difference Vegetation Index	NDVI	(NIR-RED)/(NIR + RED) = (R783 − R667)/(R783 + R667)	[55]
Normalised Difference Water Index	NDWI	(NIR-IR)/(NIR + IR) = (R860 − R1240)/(R860 + R1240)	[51]
Red-edge NDVI	RE-NDVI	(NIR-Red)/(NIR + Red) = (R781 − R670)/(R781 + R670)	[56]
Normalised Difference Infrared Index	NDII	(NIR-IR)/(NIR + IR) = (R780 − R1660)/(R780 + R1660)	[57]
Optimised Soil Adjusted Vegetation Index	OSAVI	(1 + 0.16) × (R850-R668)/(R850 + R660 + 0.16)	[58]
Simple Ratio Water Index	SRWI	RED/IR = R668/R1240	[59]
Moisture Stress Index	MSI	SWIR/NIR = R1596/R830	[60,61]
Water Index	WI	RED/NIR = R673/R850	[62]
Water Band Index	WBI	NIR/NIR = R970/R900	[63]

R represents the reflectance value (%) of the wavelength of its subscript.

shows these traditional indices.

2.5. Regression Using Machine Learning Models

Random Forest (RF) and Partial Least Squares Regression (PLSR) are two of the most widely used machine learning models [64] for classification and regression tasks in various scientific fields, including bioinformatics, environmental sciences, and precision agriculture. These models have proven capabilities for applications in crop water status prediction from hyperspectral data, yield prediction, biomass prediction, crop nitrogen prediction, disease infection prediction, and many other uses, as they are especially well-suited to managing complex datasets with multicollinearity and non-linear relationships.

PLSR is a dimensionality reduction method that maximises the correlation with the response variable while projecting predictor variables into a smaller collection of latent variables. PLSR is well-suited to handle data sets with predictors showing multicollinearity by identifying latent components that maximise the covariance between predictors and predicted variables, thereby ensuring prediction and improving prediction accuracy [65]. It has been widely utilised in hyperspectral data processing to predict plant traits. PLSR has exhibited superiority in monitoring leaf moisture of maize [8].

RF is an ensemble learning method that builds multiple trees and aggregates their predictions. RF is noted for handling non-linear relationships, resilience against noise and outliers, and resistance to overfitting, requiring minimal data preprocessing. Due to its ability to combine predictions from numerous trees and capture complex non-linear relationships between response and predictors, RF has become a valuable tool for predicting plant water content and mapping spatial and temporal changes. RF has been widely used in agriculture for regression (e.g., yield prediction, biomass prediction, water content assessment, soil properties, moisture, nutrient estimation, etc.) [64,66].

The support vector machine is a component of machine learning techniques founded on statistical learning theory. Support vector machines have gained attention from regular machine learning users because of their increased accuracy. Support vector machines (SVMs) demonstrated strong learning and generalisation capabilities in classification, regression, and forecasting. SVM has long been a machine learning hotspot due to its superior learning performance [67].

The dataset was randomly divided into two sets, with 70% used for training the models and 30% reserved for validation [68]. The dataset’s descriptive statistics are presented in

Table 3. This table presents model training, testing, and full dataset descriptive statistics.

Dataset	Median	Mean	Maximum	Std. Dev.	Coefficient of Variation	Number of Samples
Full	79.7799	76.1399	90.0515	11.4491	15.037	225
Training	79.8434	76.1666	89.1839	11.8007	15.4933	160
Testing	78.8624	76.0743	90.0515	10.6214	13.9619	65

Note: Std. Dev. is the standard deviation.

. For model tuning and parameterisation, we employed the cross-validation (k = 10) and grid search methods to select the best-fit values based on the combinations that achieved the lowest RMSE. The tuning parameter for PLSR was the number of components set at 1–5 for the individual index model and 1–10 for the combined index model; for RFR, the parameters were the number of estimators (200, 500, 800), maximum depth (None, 3, 5, 8), mtry (1, 2, 3, 4, 5), and minimum samples per leaf (1, 2, 5); for the ANN model, the major parameters were size (0–10), learning rate (0.05, 0.1, 0.2), and decay (0–1); and for the SVM model, we used the RBF Kernel C (0–10), gamma (0.001–1) and epsilon (0.001–0.1). This process is best explained in An et al. [69].

Final model input vegetation indices were selected through four feature selection combinations: recursive feature elimination (RFE), LASSO regression, random forest variable importance, and Pearson’s correlation. Indices were selected based on the frequency of selection by the first three methods. The first set of selected indices was those ranked in the top ten feature selection methods across the three methods. When the selected indices across the three methods were fewer than six indices, those selected by the two feature selection methods were ranked and subjected to Pearson’s correlation. Pearson’s correlation was then applied to decide which ones to maintain in the prediction model in the event of equal frequencies.

The robustness and accuracy of the models were assessed by using the RMSE, MAE, and R², which were calculated as follows:

$$RMSE=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)}^2}$$

(5)

$$MAE={\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right|$$

(6)

$${\mathrm{R}}^2=1-{\displaystyle \frac{\mathrm{S}{\mathrm{S}}_{\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{d}\mathrm{u}\mathrm{a}\mathrm{l}\mathrm{s}}}{\mathrm{S}{\mathrm{S}}_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}}}$$

(7)

where RMSE is the root mean square error, MAE is the mean absolute error, R² is the coefficient of determination, n is the total number of observations in the dataset, i is the index of each observation (runs from 1 to n), ∑ indicates adding up all terms from i = 1 to i = n, yi is the observed value of the dependent variable at the i-th observation, $\widehat{{\mathrm{y}}_{\mathrm{i}}}$ is the predicted value of the dependent variable at the i-th observation, |.| is the absolute value, and SS is the sum of squares.

3. Results

3.1. Canopy Spectral Response to Changes in Plant Moisture Content

To assess the influence of different irrigation treatments on spectral reflectance, we plotted the collected canopy spectral data by growth stage and treatment for the two crops (Figure 3). For winter wheat, from the booting stage (Figure 3a), there were differences in reflectance in the NIR bands. As the growth stages progressed, these differences intensified and were extended to the VIS and NIR bands as well. Interestingly, from the heading (Figure 3b) to maturity (Figure 3g), the higher the irrigation amount, the higher the reflectance in the NIR bands. From the filling stage (Figure 3d), W0 treatment showed the highest reflectance in the VIS and SWIR bands, while having the lowest reflectance in the NIR bands. From the filling stage to the maturity stage, NIR reflectance gradually reduced while the VIS and SWIR reflectance increased as plant moisture was gradually reduced and plant senescence occurred. For summer maize (Figure 3h–j), there was an increase from the tasseling (Figure 3h) to milking (Figure 3i) stage and a reduction at the hard dough stage (Figure 3j) as the plant senescence with reduced plant moisture and physiological maturity. These changes in reflectance highlight the spectral response to plants’ internal moisture. To further highlight the spectral sensitivity to PMC, Pearson Correlation was conducted between PMC and spectral bands. It shows a strong negative correlation in the VIS and SWIR regions, while the NIR bands show positive correlations (Figure S1).

3.2. Performance of New Indices in Monitoring PMC

To test the ability of the indices to detect moisture at critical growth stages of winter wheat and summer maize, growth stage correlations were performed, and significant correlations were found between the indices and PMC. The values of Pearson’s correlations of the highest correlation per index type are presented in

Table 4. This table shows the growth stage Pearson’s correlation (R) values of the index type and PMC.

Crop	Growth Stage	Sampling Dates	Sample Size	Vegetation Index
				NWSI	EWSI	NDI	Published
Winter wheat	Booting	17 & 25 March	30	(12), −0.3768 *	(17), −0.5978 ***	(16), 0.5027 **	WBI, −0.4996 *
	Heading	2 & 10 April	30	(12), −0.6681 ***	(17), −0.1785 ns	(10), −0.1334 ns	OSAVI, 0.6675 ***
	Flowering	18 April	15	(7), −0.8694 ****	(4), −0.7981 ****	(16), 0.8779 ****	WBI, 0.8931 ****
	Filling	26 April	15	(14), −0.8680 ****	(2), −0.8193 ****	(7), 0.7468 ****	NDVI, 0.8559 ****
	Dough	4 May	15	(15), −0.8963 ****	(11), −0.9003 ****	(16), 0.9336 ****	MSI, −0.9236 ****
	Ripening	12 May	15	(3), −0.9539 ****	(2), −0.9476 ****	(7), 0.9477 ****	WBI, −0.9456 ****
	Maturity	18 May	15	(7), −0.9892 ****	(4), −0.9905 ****	(1), 0.9905 ****	SRWI, −0.9917 ****
	All	NA	135	(12), −0.9313 ****	(11), −0.9393 ****	(7), 0.9393 ****	SRWI, −0.9511 ****
Summer maize	Tasseling	28 July, 5 August	30	(5), −0.7151 ****	(17), −0.5907 **	(16), 0.6147 ***	SRWI, 0.5379 *
	Milking	13 August & 1 September	30	(14), −0.4222 *	(14), −0.4982 *	(20), 0.4980 *	NDWI, 0.5984 ***
	Hard dough	16 September & 2 October	30	(14), −0.5521 **	(2), −0.2986 ns	(4), 0.1963 ns	NDWI, 0.4697 *
	All	NA	90	(5), −0.8369 ****	(2), −0.7507 ****	(6), 0.7380 ****	NDVI, 0.8125 ****
All data	All	NA	225	(12), −0.8908 ****	(2), −0.8608 ****	(16), 0.8428 ****	SRWI, −0.8845 ****

NWSI is the normalised water stress index, EWSI is the exponential water stress index, NDI is the normalised drought index, and published represent the traditional indices used in this work. WBI is the water band index, OSAVI is the optimised soil adjusted vegetation index, NDVI is the normalised difference vegetation index, MSI is the moisture stress index, SRWI is the simple ratio water index, and NDWI is the normalised difference water index. Numbers in parentheses indicate the specific index (NWSI14, or EWSI17) found in Table S1. NA denotes not applicable. The significant stars * denotes p < 0.05, ** denotes p < 0.01, *** denotes p < 0.001, **** denotes p < 0.0001, ^ns denotes not significant.

. As presented, NWSI had correlation coefficient values ranging from −0.3768 * from the booting stage to a maximum of −0.9892 **** at maturity, with an overall data correlation coefficient (R) of −0.9313 **** for winter wheat. Growth stage correlation coefficients are −0.3768 *, −0.6681 ***, −0.8694 ****, −0.8680 ***, −0.8963 ***, −0.9539 ****, and −0.9892 ****, respectively, for booting, heading, flowering, filling, dough, ripening, and maturity stages. The EWSI had −0.5978, −0.7981 ****, −0.8193 ****, −0.9003 ****, −0.9476 ****, −0.9905 ****, and −0.9393 ****, respectively, for booting, flowering, filling, dough, ripening, maturity, and overall data for winter wheat. The NDI had growth stage correlation coefficients of 0.5027 **, 0.8779 ***, 0.7468, 0.9336 ****, 0.9477 ****, 0.9905 ****, and −0.9393 ****, respectively, for booting, flowering, filling, dough, ripening, maturity, and overall data for winter wheat. Finally, the published indices had correlation coefficients of −0.4996 *, 0.6675 ****, −0.8931 ****, 0.8559 ****, −0.9236 ****, −0.9456 ****, −0.9917 ****, and −0.9511 ****, respectively, for the growth stages of booting, heading, flowering, filling, dough, ripening, and overall data. These correlations demonstrate the sensitivity of the indices to plant moisture at various growth stages of winter wheat, also highlighting the suitability of the new indices for moisture monitoring. Among the winter wheat growth stages, the NWSI showed the highest correlations at heading, filling, and ripening, surpassing the published indices, which had their highest correlations at flowering and maturity. EWSI and NDI recorded the highest correlations during the booting and dough stages, respectively.

With summer maize, NWSI had Pearson’s correlation coefficients of −0.7151 ****, −0.4222 *, −0.5521 *, and −0.8369 **** for tasseling, milking, dough, and overall data, respectively. These performances were superior to the published indices at the tasseling, dough, and overall maize data, where the published indices recorded −0.5379 *, 0.4697 *, and 0.8125 ***, respectively. Both the EWSI (−0.5907 **) and NDI (0.6147 ***) recorded higher correlations at the tasseling stage than the published indices. Although the correlation strength of summer maize was lower than that of winter wheat, the NWSI outperformed other indices, thus suggesting its better suitability for monitoring summer maize moisture stress. We further conducted an overall correlation analysis combining winter wheat and summer maize data, which yielded correlation coefficients of −0.8908 ****, −0.8608 ****, 0.8428 ****, and 0.8845 ****, respectively, for NWSI, EWSI, NDI, and published indices. This analysis confirmed that the NWSI had the best correlation performance, with a correlation coefficient of R = −0.8908 ****. This indicates that the new index has the potential to monitor the water content of both summer maize and winter wheat. The significant stars * denotes p < 0.05, ** denotes p < 0.01, *** denotes p < 0.001, **** denotes p < 0.0001.

3.3. Plant Moisture Simulation Ability of the New Three-Band Indices

Four regression models (RF, PLRS, SVM, and ANN) were fitted using the combined summer maize and winter wheat data to further assess the general predictive ability of the new indices. To better test the performance of the new indices, all three types of novel indices were initially individually fitted, and model training performance metrics are presented in

Table 5. Model training metrics when each index type was used as a separate input.

Models	NWSI					EWSI					NDI					Published Indices
	R2	R2 CI		RMSE	MAE	R2	R2 CI		RMSE	MAE	R2	R2 CI		RMSE	MAE	R2	R2 CI		RMSE	MAE
		LL	UL				LL	UL				LL	UL				LL	UL
RF	0.8524	0.96	0.99	4.1699	2.9750	0.8154	0.95	0.98	4.9207	3.5184	0.8213	0.95	0.98	4.8422	3.5591	0.8881	0.96	0.99	3.8510	2.7826
PLSR	0.8428	0.74	0.84	5.3586	3.5987	0.7816	0.70	0.86	5.7246	4.2421	0.7926	0.72	0.89	5.0859	3.8943	0.8038	0.73	0.88	5.2322	3.6851
SVM	0.8348	0.92	0.98	4.4180	3.0589	0.7948	0.90	0.97	5.2677	3.4942	0.7822	0.86	0.97	5.3841	3.6217	0.8646	0.95	0.98	4.5686	3.0312
ANN	0.8670	0.92	0.97	3.9600	2.8800	0.8270	0.90	0.97	4.6868	3.4901	0.8267	0.79	0.93	4.4894	3.3951	0.8360	0.96	0.99	4.6375	3.3064

Note: NWSI is the normalised water stress index, EWSI is the exponential water stress index, NDI is the normalised water stress index, R² is the coefficient of determination, RMSE is the root mean square error, MAE is the mean absolute error, RF is the random forest model, PLSR is the partial least squares regression model, SVM is the support vector machine model, ANN is the artificial neural network model, CI is the confidence interval (95%), LL is the lower limit and UL is the upper limit.

. The ANN model consistently yielded the highest R² values (0.8670, 0.8270, and 0.8267) and the lowest RMSE (3.9600, 4.6868, and 4.4894) and MAE (2.8800, 3.4901, and 3.6217) for all three new index types. The RF model had the second-best fitting results for the new indices, being slightly outperformed by the ANN model. With the published indices, the RF model produced the best-fitting results, with R², RMSE, and MAE of 0.8881, 3.8510, and 2.7826, respectively (Table 5).

After model training, PMC simulations were performed. Results show (Figure 4) an R², RMSE, and MAE of 0.7773, 4.9861, and 3.5078 for RF, 0.7878, 4.8609, and 3.4021 for PLSR, 0.7416, 5.3882, and 3.8265 for SVM, and 0.7269, 5.5088, and 3.8162 for ANN, respectively, with the NWSI as input. With NWSI, PLSR had the best simulation metrics, slightly outperforming the RF model. Both the SVM and ANN models also produced good results. The EWSI also demonstrated acceptable performance in simulating PMC, as shown in Figure 5. The R² ranged from 0.6638 to 0.7551, the RMSE from 5.285 to 6.1691, and the MAE from 3.9573 to 4.7792 across all models. The SVM model slightly outperformed the RF model, while the PLSR had the lowest accuracy, with an R² of 0.6638, an RMSE of 6.1691, and an MAE of 4.7792. The EWSI performance was less impressive than that of the NWSI. From Figure 6, the NDI presented an overall lower accuracy than EWSI, with a maximum R² of 0.7299, the lowest RMSE of 5.5715, and the lowest MAE of 4.1064 across all model types. The PLSR reported the lowest accuracy, with an R² of 0.5542, RMSE of 7.0807, and MAE of 5.3034. The ANN, RF, and SVM all had R² greater than 0.72. Traditional indices were also fitted to models for performance comparison, with their results in Figure 7. They produced a competing performance to that of the EWSI. The ANN model received the best R² value (0.7650) and RMSE (5.1265). The PLSR continued to produce the lowest accuracy, with an R² of 0.6521, RMSE of 6.2304, and MAE of 4.5011. The RF model had an R² of 0.7578, while the SVM model had an R² of 0.7042.

In summary, the NWSI models consistently produced the best simulation performances, as all models achieved an R² greater than 0.72. Additionally, the NWSI-PLSR model achieved the highest R² value of 0.7878, while the NWSI-RF model achieved an R² of 0.7773. The EWSI and the published indices also yielded good results with maximum R² of 0.7551 and 0.7650, respectively. Notably, the RF model achieved remarkable performance, with an R² greater than 0.70 across all input types. Apart from the NWSI-PLSR model, the PLSR recorded the lowest accuracy across all other index types.

To further investigate how these new indices enhance the accuracy of moisture monitoring in winter wheat and summer maize, they were combined with traditionally recognised moisture stress monitoring indices and used as input data for modelling. As presented in

Table 6. This table presents the combined new index and traditional index model training performance results.

Models	NWSI-Published Indices					EWSI-Published Indices					NDI-Published Indices
	R2	R2 CI		RMSE	MAE	R2	R2 CI		RMSE	MAE	R2	R2 CI		RMSE	MAE
		LL	UL				LL	UL				LL	UL
RF	0.8822	0.96	0.99	3.8091	2.6434	0.8817	0.97	0.99	3.8928	2.7541	0.8825	0.97	0.99	3.9321	2.8059
PLSR	0.8065	0.75	0.85	5.1428	3.5870	0.8106	0.72	0.88	5.0961	3.6773	0.8349	0.78	0.91	4.7837	3.4400
SVM	0.8573	0.96	0.99	4.4463	2.9927	0.8537	0.96	0.99	4.7592	3.1269	0.8545	0.98	0.99	4.6366	3.0676
ANN	0.8983	0.97	0.99	3.6654	2.6705	0.9020	0.98	0.99	3.5745	2.7405	0.9024	0.98	0.99	3.7406	2.5989

Note: NWSI is the normalised water stress index, EWSI is the exponential water stress index, NDI is the normalised water stress index, R² is the coefficient of determination, RMSE is the root mean square error, MAE is the mean absolute error, RF is the random forest model, PLSR is the partial least squares regression model, SVM is the support vector machine model, ANN is the artificial neural network model, CI is the confidence interval (95%), LL is the lower limit and UL is the upper limit.

, these combinations achieved better model training metrics, with the lowest R² value of 0.8065 (NWSI-PLSR model). Additionally, the ANN model consistently achieved the best fitting R² across all input types, ranging from 0.8983 to 0.9024, while the RF model recorded the second-best results with R² ranging from 0.8817 to 0.8825. The PLSR model recorded the lowest R² values, ranging from 0.8065 to 0.8349. Reportedly, there was an increase in the R² values over the individual index models (Table 5), thus suggesting that the combination improved the model fitting accuracy. There were also reductions in the RMSE and MAE. This then suggests that this combination is a meaningful and valuable step in moisture monitoring in winter wheat and summer maize rotation systems.

After model training, PMC simulations were performed. Like the single-index models, the NWSI-Published indices model again achieved the highest R² value. The NWSI-Published indices-SVM model yielded an R² of 0.8203, RMSE of 4.4849, and MAE of 3.2050 (Figure 4 and

Table 7. This table presents the model testing/validation metrics for the combined novel indices and traditional indices models.

Models	NWSI-Published Indices					EWSI-Published Indices					NDI-Published Indices
	R2	R2 CI		RMSE	MAE	R2	R2 CI		RMSE	MAE	R2	R2 CI		RMSE	MAE
		LL	UL				LL	UL				LL	UL
RF	0.7855	0.67	0.89	4.8949	3.4195	0.7508	0.60	0.87	5.3071	3.7089	0.7819	0.64	0.88	4.9832	3.5951
PLSR	0.7249	0.65	0.87	5.5852	4.2168	0.5950	0.55	0.85	6.7356	4.7861	0.5709	0.40	0.76	7.0638	4.9135
SVM	0.8203	0.70	0.90	4.4849	3.2050	0.7597	0.63	0.88	5.1718	3.7078	0.7894	0.63	0.90	4.8423	3.4395
ANN	0.7039	0.62	0.88	5.7500	4.1072	0.5708	0.49	0.81	7.2146	4.9354	0.7858	0.65	0.87	4.9949	3.5208

Note: NWSI is the normalised water stress index, EWSI is the exponential water stress index, NDI is the normalised water stress index, R² is the coefficient of determination, RMSE is the root mean square error, MAE is the mean absolute error, RF is the random forest model, PLSR is the partial least squares regression model, SVM is the support vector machine model, ANN is the artificial neural network model, CI is the confidence interval (95%), LL is the lower limit and UL is the upper limit.

). There was also an increase in the R² of the RF model, from 0.7773 to 0.7855, while the ANN and PLSR models recorded a decrease from 0.7878 to 0.7249 and from 0.7269 to 0.7039, respectively. This indicates that input data combinations have a mixed effect on various models, as some experienced positive impacts while others were negatively affected. The EWSI-Published index models did not exhibit a significant positive impact, as both the RF and SVM models yielded nearly identical results to the single EWSI models. Contrarily, the PLSR and ANN models again recorded reductions (Figure 5 and Table 7). With the NDI-Published indices models, there was a significant positive impact across all models compared to the NDI input models and the published indices models. The RF R² increased from 0.7229 to 0.7819, the PLSR model R² rose from 0.5542 to 0.5709, the SVM model R² increased from 0.7217 to 0.7894, and the ANN model R² rose from 0.7299 to 0.7858. The reported accuracies are significantly higher than those for the models of these indices when fitted without combination. As shown in Figure 6, the highest R² for the NDI model is 0.7299, while Figure 7 reports the best R² at 0.7650. The combined NDI-Published indices-models achieved R² increases for the RF models from 0.7578 to 0.7819, the SVM model from 0.7042 to 0.7894, and the ANN model from 0.7650 to 0.7858 compared to the published indices models.

Further, the contribution of each variable to the RF model was calculated and plotted in Figure 8. For both the NWSI-Published indices and NDI-Published indices combined models, there was a competitive contribution across the index types. In the case of the NWSI, the top five indices are NWSI indices: NWSI12 (9.5%), NWSI15 (9.3%), NWSI14 (9.3%), NWSI5 (9.2%) and NWSI6 (9.2%), topping the traditional indices with the highest contributor, RE-NDVI, with the percentage of 8.9%. The lowest three contributors are NDWI (4.9%), WBI (6.7%), and SRWI (7.6%). In contrast, the conventional indices’ contribution tops the NDI indices. The top seven contributors are NDWI (12.5%), SRWI (11.9%), NDVI (11.1%), EWSI1 (10.2%), WI (10.1%), RE-NDVI (9.4%) and EWSI2 (9.3%). Notably, the contribution gaps remain very small and do not exhibit absolute dominance of any single index type over the others.

4. Discussion

4.1. Monitoring Summer Maize and Winter Wheat Water Content Using the New Three-Band Vegetation Indices

Water stress monitoring in a summer maize-winter wheat rotation cropping system is a significant challenge that requires a solution, which this work aimed to provide. This work combined spectral bands across the three regions (VIS, NIR, and SWIR) for monitoring the plant moisture content of summer maize and winter wheat. Due to the high sensitivity of summer maize and winter wheat to moisture stress during their reproductive growth stages (booting, heading, tasseling, silking, flowering, filling, and ripening), it is crucial to monitor plant moisture levels during these critical stages. Our findings demonstrate that all the new index types calculated in this work possess a strong PMC monitoring capacity for winter wheat and summer maize, consistent with Li et al. [42] (Table 4). This result is consistent with findings from other works that reported high water sensitivity of wheat during these stages [24]. In our work, the NWSI showed the best performance of all the types of indices used in this work due to the constants added to its calculation formula, which help to stabilise the index, and its strategic band locations across the VIS, NIR, and SWIR. It produces the best correlation performance at critical growth stages of summer maize and winter wheat, and provides a better performance than the traditional water stress monitoring indices (Table 4). Its strategic band locations further enhanced its stability and resistance to external effects that generally affect two spectral wavelength combinations, like NDVI and other two-band combinations that commonly face saturation issues and fail to detect subtle changes in moisture content [70]. The performances exhibited by these novel indices have validated the enormous potential of spectral data that can be leveraged to enhance the monitoring of PMC in summer maize-winter wheat cropping systems, highlighting the potential of the new NWSI for this application. Conclusively, these indices are useful tools for guiding irrigation management in summer maize-winter wheat crop rotation systems.

From a practical perspective, these proposed indices have potential for integrating into irrigation scheduling and decision-making systems. When deployed using ground-based remote sensing platforms (spectroradiometers), they provide spatially explicit indicators of plant moisture status at field scales, enabling stress detection before visible symptoms appear. To integrate into the normal irrigation workflow, based on the highly significant growth stage correlations, threshold index values set by crop and growth stage can effectively inform irrigation timing [71]. They can also be used as input data for machine learning models to predict summer maize and winter wheat moisture content and inform decisions on irrigation scheduling. This ensures proper management of water resources and improves crop yield by preventing moisture stress.

The literature has also documented the different sensitivity of vegetation indices to particular plant biophysical characteristics at different growth stages. These changing sensitivities are related to environmental elements (solar angle, soil background), plant characteristics (canopy structure, leaf angle inclination, leaf area index, leaf internal moisture and cell structures, as well as leaf chlorophyll and pigments), and viewing geometry. According to Prudnikova et al. [72], during the early phases of plant growth, soil background and soil variability affect VIS and NIR canopy spectral reflectance. Additionally, background materials affect spectral reflectance in the early phases of growth [73]. When the plant canopy grows and is properly covered, these effects are diminished. The amount of solar radiation that is intercepted and reflected is directly influenced by the leaf area index, canopy coverage, and leaf inclination angles in relation to the plant canopy structure [74]. The varied sensitivity of vegetation indices to moisture content in plants across different growth phases is mostly due to these underlying elements, as well as internal moisture, structural makeup, and plant health.

4.2. PMC Simulation Ability of the New Indices and Spectral Response to PMC Changes

Due to climate change, there are enormous changes to the environmental conditions in which crops are grown. To mitigate these challenges, improvements have also been made in remote sensing equipment, data processing, and analysis. The novel NWSI, after demonstrating its worth as the best-performing index across the growth stages of summer maize and winter wheat (Table 4), was then further tested, as shown in Figure 4. Again, its performance was appreciable in simulating PMC of the combined crops data with a maximum R² of 0.7878 (Figure 4). This result is supported by Li et al., El-Hendawy et al., and J. Zhang et al. [20,41,42]. It is worth noting that the selected bands for each NWSI covered the spectral range of the VIS, NIR and SWIR (Table S1). This combination provides an added opportunity, as the specific advantages of each spectral region are embedded in these NWSI indices. The VIS bands exhibit high sensitivity to changes in chlorophyll and pigments related to stress. In contrast, the NIR bands respond to changes in plant health, internal leaf structure, canopy coverage, and growth parameters. The SWIR bands then add sensitivity to changes in internal plant moisture. This is another key underlying foundation of the NWSI’s superior performance over other indices, in addition to the added constants in the formula.

EWSI and NDI performed appreciably when fitted with the models (Figure 5 and Figure 6). Li et al. [42], who utilised UAV multispectral and hyperspectral data (400–1000 nm range) to create three band indices, also reported that these indices performed better in transfer learning models for predicting PMC in a winter wheat and summer maize rotation system. In our work, we explored the full range of the spectral (350–2500 nm), making use of the SWIR bands. This then created spectral indices that outperformed some conventionally recognised moisture stress monitoring spectral indices.

As droughts become more prevalent, improving moisture stress monitoring by enhancing the accuracy of monitoring models remains a key focus for managers and farmers. For this reason, we decided to assess the possibility of improving our PMC prediction accuracy. To achieve this feature, each index type (NWSI, EWSI, and NDI) was separately combined with the published indices and used as input data for the models. Here, three sets of simulation models were created: viz, NWSI-Published indices, EWSI-Published indices, and NDI-Published indices (Table 6).

After model training, PMC simulations were performed using the 30% hold-out data. As revealed, the NWSI-Published indices-SVM model had the best PMC simulation accuracy with the highest R² (0.8203) and lowest error metrics (Table 7), consistent with findings reported by Cen et al. [75], who reported that the combination of the three-dimensional drought indices and five vegetation indices improved the monitoring of vegetation water content in sorghum and maize. The NWSI-Published indices-SVM model also yielded an RMSE of 4.4849 and an MAE of 3.2050 (Table 7). There was also an increase in the R² of the RF model, from 0.7773 to 0.7855, while the ANN and PLSR models recorded a decrease from 0.7878 to 0.7249 and from 0.7269 to 0.7039, respectively. This indicates that input data combinations have a mixed effect on various models, as some experienced positive impacts while others were negatively affected. Due to this effect, critical considerations are required when selecting input types for modelling PMC.

The EWSI-Published index models did not exhibit a significant positive impact, as both the RF and SVM models yielded nearly identical results to the single EWSI models. Contrarily, the PLSR and ANN models again recorded reductions (Figure 4 and Table 7). With the NDI-Published indices models, there was a significant positive impact across all models compared to the NDI input models and the published indices models. The RF R² increased from 0.7229 to 0.7819, the PLSR model R² rose from 0.5542 to 0.5709, the SVM model R² increased from 0.7217 to 0.7894, and the ANN model R² rose from 0.7299 to 0.7858. The reported accuracies are significantly higher than those for the models of these indices when fitted without combination. As shown in Figure 5, the highest R² for the NDI model is 0.7299, while Figure 6 reports the best R² at 0.7650. The combined NDI-Published indices-models achieved R² increases for the RF models from 0.7578 to 0.7819, the SVM model from 0.7042 to 0.7894, and the ANN model from 0.7650 to 0.7858 compared to the published indices models. The NDI bands feature spectral bands in the range of 400 nm to 1400 nm, with most selected index bands having a double VIS band between 400 nm and 500 nm. Combining them with traditional indices that include SWIR bands facilitated improved performance. In general, these combinations increased the quality of the input data, which subsequently increased model performance.

With the ongoing need for improvements in moisture stress monitoring to prevent crop failure, yield loss, and food shortages, applying such innovations to multispectral data to improve model outputs and accuracy will be a crucial step. This will help promote wider adoption of this concept and its indices.

These findings have significant implications for monitoring water content in the summer maize-winter wheat cropping system, as the NWSI and NDI indices’ ability to detect moisture levels at key growth stages is a valuable tool for informed irrigation scheduling. Based on our results, where index combinations influenced model accuracy both positively and negatively, it is recommended that band combinations be carefully selected when used as input data for monitoring moisture stress in plants. With ongoing advancements in remote sensing technology, along with improvements in data processing, analysis, and application, continued efforts to enhance the usability of remote sensing data for precise moisture stress monitoring will lead to greater resource efficiency and sustainable production.

Our results show a similar pattern of canopy reflectance across treatments [52,76] of both summer maize and winter wheat. There is a clear pattern of NIR spectral reflectance reduction as the plant water content decreases over time (Figure 3). Interestingly, the negative correlation in the VIS region extended to 730 nm, and a negative correlation was observed from 1335 nm. Similar results have been reported in other works [14]. The highest absolute value of the correlation was in the VIS, a finding supported by [24]. According to the literature, the reduction in NIR (716–1300 nm) canopy reflectance due to a decrease in plant water content, as well as changes in canopy parameters such as coverage, LAI, and plant health, is supported by Hunt & Rock and Penuelas et al. [61,77]. This change is accompanied by an increase in the VIS (400–730 nm) and SWIR (1335–2500 nm) wavebands, as also reported by Cao et al. [78]. The change in spectral reflectance is visible across the VIS, NIR, and SWIR spectra during the growth stages. This coincides with the decrease in PMC, chlorophyll, and canopy coverage, which is also associated with plant senescence.

The SWIR and VIS spectra differentiate between the data collection days. The reflectance continued to increase as the growth progressed during the reproductive stage of summer maize-winter wheat, consistent with findings from Cao et al. and Schmidt et al. [78,79]. As water stress increases, which initiates a chain effect on plants’ physiological response, the water absorption troughs at 970 nm [64], around 1200 nm and 1450 nm, gradually reduce and almost disappear sequentially as the water content, chlorophyll, and other pigments decrease during the late reproductive stage. The band selection process and the waveband correlation analysis revealed that the red-edge region remains critical for plant water monitoring, as it showed the highest correlation to plant water content. These changes indicate moisture stress in plants and are helpful for in-season moisture monitoring of summer maize and winter wheat during their early stages of establishment.

4.3. Remote Sensing (RS) and Machine Learning (ML) for Summer Maize Water Content Monitoring

Combining RS and ML to monitor plant water content has gained considerable attention and recognition. This is due to this combination’s reliability, accuracy, and time-efficient result acquisition. Additionally, spectral processing improves model performance and stability [80]. Several works have combined these two to accurately predict and monitor plant water stress by preprocessing and calculating vegetation indices (VIs) from spectral reflectance [81,82]. Others have reported the use of spectral reflectance values directly [83], and some by combining both variables [84]. Either way, these provide reliable results for monitoring plant water stress and preventing yield loss due to moisture stress. This work applied ML to VIs calculated from hyperspectral data and processed spectra to predict plant water content non-destructively. Four models, random forest (RF), partial least squares regression (PLSR), support vector machine (SVM), and artificial neural networks (ANN) models, were applied to the calculated VIs. Their predictions were appreciable, consistent with the results of Zununjan et al. [10]. The SVM model yielded the best overall results when the NWSI and published indices were combined as input (Table 7). The better performance of the SVM was also reported by Shi et al. [85]. The other models also performed well in predicting moisture content. Although the SVM model achieved the best simulation accuracy, the RF model reportedly has the best performance when considering all model input types, consistently maintaining an R² greater than 0.7200 across all inputs. This then put the RF model in a position of consideration for use, as its adaptability to several input types produces better results, an added asset in this regard. The PLSR performed relatively poorly across all input types, except for NWSI. Our model performance result is consistent with findings reported by Mirzaie et al. [86]. The developed models and input combinations are crucial for monitoring PMC and guiding the initiation of irrigation events. This further validates ML’s vital contribution to precision agriculture. These findings could be applied in regions with similar environmental conditions to the study location.

This study was conducted on two crops: summer maize and winter wheat. Each crop has one season, and single-site data were used in this work. It is worth noting that future work will focus on multi-site, multi-season, and multi-climatic regions, as well as additional crop testing of these indices to further establish their superior performance. Yield prediction and analysis, which were not included in this current work, will also be incorporated. This will evaluate their performance across different locations, seasons, and crops. It will also help solidify their use in precision agriculture.

5. Conclusions

Scarce water resources must be utilised effectively for enhanced crop production. We developed new spectral indices, NWSI, EWSI, and NDI, which can monitor moisture stress in a summer maize-winter wheat crop rotation system. Based on growth stage Pearson’s correlation, these new indices produced the highest correlations at the booting stage (−0.5978 ***), dough stage (0.9336 ****), heading stage (−0.6681 ***), filling stage (−0.8680 ****), and at the ripening stage (−0.9539 ****) for winter wheat. For summer maize, their best correlations were observed at the tasseling (−0.7151 ****) and dough (−0.5521 *) stages (Table 4). These performances were superior to those reported by the published indices. When simulating PMC, the NWSI model also achieves an R² = 0.7878. When combined with the published indices, the R² increased to 0.8203. These reportedly were higher than those reported by the published indices. The combined NDI-Published indices model also outperforms the models of the published indices (Table 7). We conclude that these indices have improved the moisture stress monitoring of winter wheat and summer maize, based on their superior performance compared to conventional moisture stress monitoring indices, thereby laying a foundation for effective water management and sustainable crop production.

The NWSI-PLSR model achieved the highest accuracy for the individual index type models (R² = 0.7878), but this performance was not consistent across other index types. The SVM-NWSI performed best when index types were combined as input data (R² = 0.8203). The RF and SVM models demonstrated the most consistent performance across all index types and combinations. The combination of the EWSI-Published indices models resulted in decreased simulation accuracy, as the RF, PLSR, and ANN models yielded lower R² values (Figure 5a,b,d, and Table 7). This highlights the need for careful consideration when creating, selecting, and combining indices for the input data of machine learning models used in moisture stress monitoring. It can be concluded that the RF and the SVM models achieved the best performance in simulating PMC and are recommended for use with the new indices (NWSI and NDI) and their combinations with the conventional indices.