3.1. Method 1
The initial method employed to predict the moisture content of standing or whole-plant corn involved two steps. First, ear moisture content was predicted using NIRS. Then, this prediction was utilized in conjunction with a previously developed model that established a relationship between ear moisture and whole-plant moisture. In our previous study [
5], we successfully demonstrated a quadratic model that accounted for 89% of the variation in ear moisture (EM) when compared to whole-plant moisture (WP):
To assess the effectiveness of this model using our new dataset, we applied Equation (1) to estimate the whole-plant moisture. The validation set consisted of ear moisture data collected in 2022. It is important to note that we purposely excluded the 2021 data for this analysis, as we intended to utilize it for NIRS calibration.
A total of 330 plants were examined, and we achieved a root-mean standard error of prediction (RMSEP) of 3.9 moisture percentage points, along with an R
2P value of 0.88 (
Figure 1). Although it is possible to account for the bias of 2.8 and the slope of 0.72 through a multi-point calibration process, the precision error cannot be corrected.
According to the performance criteria established by [
11], our model for predicting whole-plant moisture from ear moisture would fall into the screening performance category with an RPD value of 2.9. However, it would be on the threshold of the quality control category, as the RPD value (2.9) is close to the cutoff of 3.0.
This result shows promise for determining the optimal harvest timing. However, it assumes that the moisture content of the ear is known. One approach for determining the ear moisture content is to harvest ears manually, weigh them, and dry them in an oven overnight. However, this method requires infrastructure that may not be available on many farms. Another option is to involve a consultant or a forage testing laboratory to carry out the drying process, which would delay the time between sampling and harvest decision making.
3.2. Method 2
Predicting ear moisture using a handheld device like the one studied in this research would be desirable to address this limitation and make the process more convenient. However, it is important to consider the moisture prediction error when utilizing the handheld NIRS device for whole-plant moisture prediction. Therefore, the next phase of this study focused on evaluating the performance of the handheld NIRS device in predicting ear moisture.
Since our previous work did not include the same NIRS instrument, we could not simply use the 2021 and 2022 data to evaluate the performance of predicting ear moisture. Therefore, we used the 2021 data to build a model and the 2022 data to evaluate the model performance. In the 2021 dataset, 610 ears were scanned using three instruments, each conducting three scans per ear. Consequently, there were a total of 5481 scans recorded. For this analysis, the scan data were averaged for each ear, resulting in a calibration dataset of 1827 scans.
In 2021, the corn ear moisture content ranged from 26 to 75 %w.b. with an average moisture content of 49 %w.b. and standard deviation of 9.6 moisture percentage points (
Table 1). In 2022, 330 ears were scanned with a moisture range of 37–80 %w.b., an average of 55% w.b., and a standard deviation of 10 moisture percentage points (
Table 2). Both datasets agree with our previous observations, where the range, average, and standard deviation were 27–81 %w.b., 55 %w.b., and 11 percentage points, respectively [
5]. Consequently, a calibration developed with data from any year would include the moisture range and variability expected between growing seasons.
The best-performing calibration (SEC = 2.8, R
2C = 0.92) in predicting ear moisture content from spectra utilized a second derivative with a smoothing kernel length of 11 and using quadratic interpolation (
Table 2, All column). Additionally, there is good agreement between the standard error of calibration and cross-validation (SECV = 2.8), indicating a robust model that is not overfitted. A SEC of 2.8 is numerically better than the 3.1 observed in our previous study when predicting ear moisture with a handheld NIRS in the wavelength range of 740 to 1050 nm [
5]. However, this numerical difference is unlikely to have practical implications when selecting an instrument or wavelength range for predicting ear moisture.
We also developed single- and multiple-instrument prediction models to determine whether the prediction model was affected by variability introduced by different instruments of the same make and model. A common strategy for mitigating instrument-to-instrument variability involves building a calibration model using multiple instruments. This approach ensures model stability as additional instruments are deployed in the field. Thus far, the results take this approach.
To evaluate the necessity of spectra from multiple instruments on calibration data, we developed a calibration with each individual instrument and with each possible combination of instruments (
Table 1). We subsequently used the 2022 season dataset to evaluate these models (
Table 2). Our analysis revealed that predicting the validation set with a single instrument model resulted in an increased RMSEP (see columns labeled 1→ 2, 1→3, 2→1, 2→3, 3→1, 3→2 in
Table 2), primarily due to the introduction of a more substantial bias. The bias was lower when the same instrument was used across both seasons (columns 1→1, 2→2, 3→3), indicating that hardware can introduce bias in addition to seasonal variations. Instrument three exhibited the highest inter-seasonal bias of 2.3 moisture percentage points, while instrument one had the lowest at 0.33.
Generally, the slope remained stable regardless of the instruments used in the calibration development. However, instrument three showed numerically lower sensitivity when used as the sole calibration instrument.
The RMSEP was lowest when all instruments were used to develop a calibration to predict the spectra from one instrument in the second season and highest for single-instrument calibrations. The combination of two instruments to predict a third yielded results like the “All” instrument calibration when instrument three was part of the calibration, with RMSEPs ranging from 2.7 to 2.8 moisture percentage points for the 2&3 and 3&1 calibration models. However, the 1&2 model resulted in an RMSEP of 4.9, providing further evidence that instrument three introduced variability to the calibration model that instruments 1 and 2 did not.
Based on these results, we conclude that instrument variability significantly influenced our results and that using two instruments might be insufficient for capturing this variability adequately. This result supports previous work on handheld NIRS devices [
12]. Practitioners should consider this instrumental variability in calibration model development for more accurate and reliable predictions. Multi-instrument calibrations or calibration transfer methods should be considered when managing calibrations across multiple instruments.
Typically, plotting predicted ear moisture against actual ear moisture at this stage is customary. However, this study primarily focuses on utilizing these predictions to estimate whole-plant moisture, as this parameter is relevant to the ensiling process and the production of high-quality corn silage. Consequently, our next step involved utilizing the ear moisture predictions obtained from the NIRS as input for Equation (1) to predict the whole-plant moisture content.
3.3. Method 3
When predicting whole-plant moisture by first predicting ear moisture using NIRS and then applying Equation (1), we observed a significant bias, leading to an RMSEP of 3.4 (
Figure 2). Additionally, there was a loss of precision with an R
2P value of 0.82 compared to the oven-dry ear moisture reference method. However, it is worth noting that if we assume a single-point calibration could be conducted each season, achieving a lower SEP of 2.5 would be possible. Despite this improvement, the reduced precision firmly categorizes the model in the screening class.
The final approach we evaluated involved directly predicting the whole-plant corn moisture content using handheld NIRS ear scans. If successful, this approach would eliminate the need for a separate calibration model to predict whole-plant moisture from ear moisture (Equation (1)) and simplify the prediction process. Only one model would be required instead of maintaining two models for ear and whole-plant moisture.
We hypothesized that the spectra obtained from the ear scans would contain information about the physiological state of the corn beyond just grain moisture content. Specifically, we expected the spectra to be sensitive to starch accumulation from sucrose as the ear progresses from the dough to the dent stage of physiological maturity. Both starch and sucrose can influence the spectra captured in the NIRS region.
However, this direct approach of predicting whole-plant moisture using ear scans resulted in a larger bias and a decrease in the precision of whole-plant moisture prediction, with an R
2P value of 0.79 (
Figure 3). Despite this performance degradation, the model classification would still fall within the screening category with an RPD value of 2.2, like our previous approaches.
Previous methods used for on-farm moisture prediction serve as performance benchmarks for our study despite their requirement for a chopped sample and thus being less convenient. The first group of methods saves time by only partially drying the forage but uses a gravimetric approach and thus has the best agreement with the oven-drying but at considerable savings in time [
3,
13,
14]. Ref. [
13] reported that moisture determined with a microwave oven or the Koster method took between 30 and 90 min, depending on the crop species, with corn silage having the longest drying time.
More recently, Ref. [
14] compared gravimetric methods to a handheld NIRS for predicting dry matter (DM) content in ensiled forage. They did not report the wavelength range of the instrument but observed that the manufacturer provided the calibration to have a significant bias. After correcting the bias, the instrument performed better in alfalfa silages than in corn. However, their dataset had a sixteen-point dry matter percentage point range in alfalfa compared to an eight-point range in corn, which may have influenced the results.
Another major difference between our work and [
14] was that we were able to establish and assess our calibration over a broader range of moisture contents, specifically from 37 to 80 %w.b., in contrast to the narrower range of 57 to 68 %w.b. This is because we were sampling standing corn where they were sampling ensiled corn where the moisture range was limited by harvest management. We also did not observe a large bias when predicting our validation set, as they reported. However, our data were collected in the same region between two growing years, using the same spectrometers and laboratory reference methods.