Estimation of Total Nitrogen Content in Forage Maize ( Zea mays L.) Using Spectral Indices: Analysis by Random Forest

: Knowing the total Nitrogen content (Nt) of forage maize ( Zea mays ) is important so that decisions can be made quickly and e ﬃ ciently to adjust the timing and amount of both irrigation and fertilizer. In 2017 and 2018 during three growing cycles in two study plots, leaf samples were collected and the Dumas method was used to estimate Nt. During the same growing seasons and on the same sampling plots, a Parrot Sequoia camera mounted on an unmanned aerial vehicle (UAV) was used to collect high resolution images of forage maize study plots. Thirteen multispectral indices were generated and, from these, a Random Forest (RF) algorithm was used to estimate Nt. RF is a machine-learning technique and is designed to work with extremely large datasets. Overall analysis showed ﬁve of the 13 indices as the most important. One of these ﬁve, the Transformed Chlorophyll Absorption in Reﬂectance Index / Optimized Soil-Adjusted Vegetation Index, was found to be the most important for estimation of Nt in forage maize (R 2 = 0.76). RF handled the complex dataset in a time-e ﬃ cient manner and Nt did not di ﬀ er signiﬁcantly when compared between traditional methods of evaluating Nt at the canopy level and using UAVs and RF to estimate Nt in forage maize. This result is an opportunity to explore many new research options in precision farming and digital agriculture. plant phenology and Nt. The data analysis showed the notable association between the TCARI (obtained from the RE wavelength) and the Nt in forage maize. We observed


Introduction
Nitrogen (N), in both its organic and inorganic forms, is essential to promote leaf development and crop growth. Nitrogen is present in the soil in two ways: the decomposition of plants naturally high in N such as legumes, buckwheat, and others (organic N; amino acids, proteins, and nucleotides), and the human application of fertilizer ( Maize (Zea mays) has one of the highest photosynthetic rates per unit of N [1] and is thus considered a C4 plant. The N content in maize determines the yield and quality of grain and must be available in the soil in adequate quantities throughout the growing season. Vos et al. [2] noticed a 29% increase in the foliar area when compared to a high versus low incorporated amount of N in the soil. A lack of N Soil Adjusted Vegetation Index), TCARI/OSAVI (Transformed Chlorophyll Absorption in Reflectance Index/Optimized Soil-Adjusted Vegetation Index), TCARI/OSAVI RE (Red Edge-based Transformed Chlorophyll Absorption in Reflectance Index/Optimized Soil Adjusted Vegetation Index), and MTCI (MERIS Terrestrial Chlorophyll Index) among others. One common result reported by researchers is that these indices tend to exhibit saturation values in moderate to high plant cover [6,23,24], but discussing this situation is out of the reach of this manuscript.
We used these indices to show internal and external relationships between vegetation cover and its response to sunlight. The use of data mining, deep learning, and other data analysis techniques can help to recognize patterns and trends that cannot be seen without these analyses. Among these techniques are decision trees (DTs), Random Forest (RF), support vector machine, and artificial neural networks, all of which perform quite well with very large databases and in current approaches involving plant nutritional status [25], water-demand [26], biomass [27], and chlorophyll content [28]. When the results of any data analysis failed to find a strong relationship between or among variables, the RF analysis may be a viable option [29]. RF is a nonparametric analysis that uses DTs, with their class values assigned according to the maximum number of forecast votes through the DTs [30], and with more predictive power in comparison to classic regression methods [31,32], in essence to efficiently handling many variables that do not adjust to a normal distribution.
We compared the collection and analysis of Nt in forage maize with the industry-standard method versus the collection of multispectral images and RF analysis. To our knowledge, there are no published studies that used UAV-collected multispectral data to estimate total Nt in combination with the RF algorithm for data analysis; therefore, the use of UAV imagery and RF in this context is innovative and will contribute to future research.

Description of the Study Areas
The study areas are in the north-central region of Mexico known as the Comarca Lagunera ( Figure 1). The cultivated area of forage maize has been reported at approximately 59,298 ha [33]. This semidesert to arid region has a mean annual temperature of 20-22 • C and mean annual precipitation of 200-300 mm.

Description of Sampling Sites
Two large agricultural farms were studied (Granada and El Porvenir; Figure 1). The Granada farm was studied during the summer-autumn forage maize production cycle in 2017, while in 2018, the spring-summer and summer-autumn cycles were evaluated at El Porvenir.
The Granada farm occupies an area of 60.4 ha. The soil composition is clay-loam. The maize hybrid sown was Pioneer P3201 and was selected because it is a hybrid of early to intermediate varieties, it has high nutritional content when used as forage for dairy cattle, high grain health, and high tolerance to root and/or stem lodging. Eight seeds meter −1 will produce a density of 103,000 plants ha −1 . The agricultural management practices included fallowing, harrow passing, skirting, and leveling (slope of 2 cm per 100 m). Chemical fertilizer was applied two times. The first application was composed of 42% ammonium sulfate, 42% monoammonium phosphate, and 16% Ryzogen (a granular microbial fertilizer used to stimulate root growth); this was incorporated into the soil at the time of sowing at a rate of 250 kg ha −1 . The second time was immediately after the first additional water had been applied; the fertilizer consisted of urea (46%) at 400 kg ha −1 . The irrigation schedule included five additional dates. The maize was harvested 117 days after sowing (DAS) when the forage reached 32% of dry matter content. The Granada farm occupies an area of 60.4 ha. The soil composition is clay-loam. The maize hybrid sown was Pioneer P3201 and was selected because it is a hybrid of early to intermediate varieties, it has high nutritional content when used as forage for dairy cattle, high grain health, and high tolerance to root and/or stem lodging. Eight seeds meter −1 will produce a density of 103,000 plants ha −1 . The agricultural management practices included fallowing, harrow passing, skirting, and leveling (slope of 2 cm per 100 m). Chemical fertilizer was applied two times. The first application was composed of 42% ammonium sulfate, 42% monoammonium phosphate, and 16% Ryzogen (a granular microbial fertilizer used to stimulate root growth); this was incorporated into the soil at the time of sowing at a rate of 250 kg ha −1 . The second time was immediately after the first additional water had been applied; the fertilizer consisted of urea (46%) at 400 kg ha −1 . The irrigation schedule included five additional dates. The maize was harvested 117 days after sowing (DAS) when the forage reached 32% of dry matter content.
On the El Porvenir site, 52 ha of forage maize was studied in the spring-summer cycle, and 34 ha in the summer-autumn cycle; this variation has no effect on the number of plant samples. The Syngenta N83N5 maize hybrid was sown in loam-clay soil at the same density as at the Granada farm. This forage maize is an earlier cycle variety with high nutritional value and high stability in various soil types and performs well in different environments. This farm was handled with the same agricultural practices as the Granada farm (fallowing, harrow passing, skirting, and leveling). A sole application of a mixture of chemical fertilizer mixture was performed. It was composed of 300 kg ha −1 of ammonium sulfate and 100 kg ha −1 of monoammonium phosphate. The irrigation schedule included a total of four dates. Maize was harvested when it reached 38% of dry matter-at 118 DAS in spring-summer, and 115 DAS in summer-autumn. On the El Porvenir site, 52 ha of forage maize was studied in the spring-summer cycle, and 34 ha in the summer-autumn cycle; this variation has no effect on the number of plant samples. The Syngenta N83N5 maize hybrid was sown in loam-clay soil at the same density as at the Granada farm. This forage maize is an earlier cycle variety with high nutritional value and high stability in various soil types and performs well in different environments. This farm was handled with the same agricultural practices as the Granada farm (fallowing, harrow passing, skirting, and leveling). A sole application of a mixture of chemical fertilizer mixture was performed. It was composed of 300 kg ha −1 of ammonium sulfate and 100 kg ha −1 of monoammonium phosphate. The irrigation schedule included a total of four dates. Maize was harvested when it reached 38% of dry matter-at 118 DAS in spring-summer, and 115 DAS in summer-autumn.

Field Sampling
For both farms, we collected composite plant samples (CPSs) before and after flowering, and each collection site was georeferenced using a Garmin Etrex 20 geographic positioning system receiver, with a 3.6 m margin of error. Before flowering, samples were taken from three plants (stem and leaves) in the same plot, and after flowering samples were taken from the leaf opposite the ear of corn from ten plants. Plant samples were sent to the laboratory for further analysis.
For the Granada farm, there were six dates for plant collection samples each associated with aerial imaging planning. The first sampling was performed at 33 10 CPSs. For the summer-autumn cycle, sampling was conducted at 53 DAS with 10 CPSs; the second at 77 DAS with 10 CPSs; the third at 92 DAS with 10 CPSs.

Laboratory Analysis of Samples
All CPSs were dried in an oven at 65 • C until they all reached a constant weight. Each CPS was then ground using a 0.5 mm mesh. The Dumas method was used to determine the percentage of Nt in each sample. According to Sweeney and Rexroad [34], the determination of Nt is made by thermal conductivity. This method transforms all N into N gas by calcination; the gases produced are reduced by copper, and subsequently desiccated, while the CO 2 (Carbon Dioxide) is trapped. The N is quantified using a Universal Quantifier. The plant samples were processed in the soil laboratory of the Centro Nacional de Investigación Disciplinaria-Relación Agua Suelo Planta Atmósfera (CENID-RASPA), located in Gomez Palacio, Durango. The sampling process was adjusted to the protocol proposed by the Compendium of International Methods of Analysis [35].

Aerial Imaging
The UAV missions were carried out after maize plants had four fully expanded leaves; this phenotypic stage represents the exponential start of the vegetative development curve. Twelve UAV missions were flown in each of the three production cycles. The UAV was equipped with a Parrot sequoia camera that has the capacity to record up to four wavelength ranges: from 530 to 570 nm (green); 640 to 680 nm (red); 730 to 740 nm (red edge); 770 to 810 nm (near-infrared). The spatial resolution of the multispectral image is 1.2 megapixels. The camera also includes a red, blue, and green (RGB) sensor, with a resolution of 16 megapixels. Regardless of the lighting conditions, the camera has a luminosity sensor in its upper section that allows it to obtain accurate images [36]. The UAV's height above the ground was 191 m; this, along with the camera's field of view, yielded images that represented a ground spatial distance of 18 cm. The images from the UAV missions were captured in EMOTION software (an image processing and flight management platform). Orthomosaic and reflectance values were obtained with Pix4Dmapper software 4.5. The spectral indices obtained from the UAV missions are shown in Table 1. Table 1. Spectral indices are derived from aerial images.

Index Equation Reference
NDRE: sensitive to chlorophyll content when crops are in medium to late stages.
NIR−RE NIR+RE [37] CCCI: based on a two-dimensional approach relating the NDVI [38] used to estimate canopy cover, and the NDRE.
NDRE NDVI [37] CCCI_simpl: estimates chlorophyll content with minimal sensitivity to other factors that can influence in the canopy signal of N.
CI_rededge: indicates chlorophyll content; value shows low sensitivity to the background effects of the soil.
NIR RE − 1 [40] MCARI/OSAVI: very sensitive to variations in chlorophyll content and highly resistant to variations in Leaf Area Index (LAI); takes into account the effects of the soil.

Index Equation Reference
MCARI/OSAVI RE: very sensitive to variations in chlorophyll content and highly resistant to variations in LAI; takes into account the effects of the soil and RE.
TCARI/OSAVI: reduces the effect of soil contribution and enhances sensitivity to chlorophyll content TCARI/OSAVI RE: very sensitive to variations in chlorophyll content and highly resistant to variations in LAI; takes into account the effects of the soil and RE. [42] MTCI: has the advantage of being sensitive to N with relatively lesser effect of water or irrigation level for corn.
GNDVI: estimates photosynthetic activity and is a commonly used vegetation index to determine water and N uptake into the plant canopy.
NIR−G NIR+G [45] CI_green: estimates chlorophyll content in leaves using the ratio of reflectivity in the NIR and green bands.
NIR G − 1 [46] NDVI: determines the green biomass during the early and middle development stages.
NIR−R NIR+R [38] RVI: estimates and monitors green biomass, specifically, in the coverage of high vegetation density.

Data Analysis
The RF decision tree classifier was used to ascertain the link between the Nt data obtained in the laboratory and the value of spectral indices derived from the aerial images. RF can develop multidimensional and multicollinear analyses on large databases [48] and is a suitable data mining technique to estimate Nt in crops [49,50]. The RF implementation was in the Python (v 3.8.0) ecosystem. The outputs of this implementation are a set of decision trees. On this, the contribution of each variable to the model is identified hierarchically. The decision tree is a flowchart, where each of the internal nodes represents an attribute, each branch represents the output of the attribute, and each sheet is a class label. The model developed included independent variables (x = vegetation indices) and a dependent variable (y = Nt). The parameters to quantify the precision and analyze the quality of the input data were declared as follows: n_estimators. Declares the number of trees in the forest. A greater number of trees improves the accuracy of the data (RF learns faster as more decision trees are included) [25]; however, a too-large number of trees can encourage RF's implementation during the training process. In the model run, the computing process limits were set at 30, 50, 100, 500, 1000, 1500, 2000, 2500, 3000, 3500, and 4000 trees. These thresholds have been used to evaluate RF's abilities [12,51]. max depth. Represents the depth of each tree in the forest. If the tree is deeper, it has more divisions and can capture more information from the data. Its value ranges from 1 to 32.
min_samples_split. Some samples are required to separate a node. Its value ranges from 10% to 100% of the samples. min_samples_leaf. A minimum number of samples required in the node. Its value ranges from 10% to 50%. max_features. Number of characteristics to consider when looking for the best node division (default is Auto).
The importance parameter was used to reduce the noise effect of uncorrelated variables. This means that the indices that have the highest correlation in the prediction can be selected and resultin a reduction in computing time. To reduce the number of indices, the model was first trained using all vegetation indices as input variables. Subsequently, based on the importance parameter, all with the lowest weight were eliminated to improve the accuracy of the model and make it more robust.
The optimization of nonparametric regression methods uses a training data-based learning process, which builds a model. The model parameters are adjusted to minimize the estimation error of the extracted variables [52]. A k-fold hold-out cross-validation method with 5 subsets was applied to evaluate the performance of the predictive model. This method has been widely used to validate machine learning models such as RF [12,28,29,[52][53][54][55] and consists of dividing the entire dataset into two subsets; the first one is to train or adjust the model and while the second one is to validate it [56]. Thus, a model is generated only for training data, which is compared with the data that were reserved for validation and was not used during the development of the model. In the present study, we used 70% of the data for adjustment and 30% for validation. These percentages of data division are similar to those reported by other authors [12,29,53,54,57].
To avoid overfitting and to optimize the performance of the RF, it is necessary to use the n_estimators and max_features hyperparameters. Different combinations of these hyperparameters are carried out in the training stage to obtain the best performance. However, this can cause overfitting, causing the model to perform well on the training set but poorly on the test set. Therefore, it is necessary to perform the cross-validation k-fold [12,28,55]. Firstly, the dataset was divided into training and test sets. In k-fold, only the training set was used and was divided into five subsets. These subsets were iterated using one-fifth of the samples to validate the model and the remaining samples (four-fifths) for training. The first iteration of the first subset was chosen to perform the validation and the four remaining iterations for training. In the second iteration, the second subset was used for validation and subsets one, three, four, and five for training, and so on until the five iterations were computed. The above indicates that five independent pieces of training are carried out to validate the model and that the final accuracy will be the average of the five previous accuracies.
The Nt estimated by the RF model and the Nt calculated in the laboratory were intercompared. The determination coefficient (R 2 ) and the mean square error was used.

Sampling Universe
The dataset included data from 162 field-collected points and twelve flight missions. The resulting data included Nt (%) obtained from laboratory processing (NLab) of the field samples, and the RF estimated Nt values from five (N 5 ) and thirteen (N 13 ) spectral indices ( Table 2).

Estimate of Nt in Laboratory
The Nt ranged from 0.81% to 4.09% with an average of 2.51% for the 2 years of the study. In the summer-autumn 2017 cycle (n = 103, mean = 2.85%, median = 2.74%, minimum value = 1.8%, maximum value = 4.04%), spring-summer 2018 cycle (n = 30, mean = 2.53%, median = 2.35%, minimum value = 1.53%, maximum value = 4.09%), and summer-autumn 2018 cycle (n = 29, mean = 1.29, median = 1.24%, minimum value = 0.81%, maximum value = 1.96%; Figure 2). The Nt values for summer-autumn 2018 were lower than those for summer-autumn 2017, possibly due to the use of two different hybrids and two different study sites. The summer-autumn period at Granada and the Pioneer P3201 hybrid was used because, in the past, they have resulted in a greater range of nitrogen concentration.

RF Model Optimization
As shown in Figure 3, r 2 raised and dropped dramatically when the model was tried in 100 trees, and subsequently raised again to 1000 and remained almost sustained. When tried in 3000 trees it raised slightly, but if this value is chosen to optimize the n_estimators parameter it can cause an overfitting and slow down the model. In order to avoid the overfitting, 1000 trees were used as an entrance parameter. The Nt values for summer-autumn 2018 were lower than those for summer-autumn 2017, possibly due to the use of two different hybrids and two different study sites. The summer-autumn period at Granada and the Pioneer P3201 hybrid was used because, in the past, they have resulted in a greater range of nitrogen concentration.

RF Model Optimization
As shown in Figure 3, r 2 raised and dropped dramatically when the model was tried in 100 trees, and subsequently raised again to 1000 and remained almost sustained. When tried in 3000 trees it raised slightly, but if this value is chosen to optimize the n_estimators parameter it can cause an overfitting and slow down the model. In order to avoid the overfitting, 1000 trees were used as an entrance parameter.

Validation
Cross-validation results: k-fold with five subsets was r 2

Selection of the Spectral Indices Based on the Parameter of Importance
Two RF iterations were implemented. In the first, all thirteen indices were included, while in the second all those with the best contribution to the value of importance were selected (Figure 4). In the second iteration, one redundancy effect from minor importance indices was assumed. The TCARI/OSAVI, CI_green, TCARI/OSAVI RE, GNDVI, and NDVI indices were included, which together make up 0.8 of the importance value. Given the decrease in the number of indices, these five improved the value of importance; however, they were not as relevant (0.60, 0.06, 0.05, 0.05, and 0.04, respectively). In this second performance, the TCARI/OSAVI registered the highest importance values in the two runs (0.66).

Selection of the Spectral Indices Based on the Parameter of Importance
Two RF iterations were implemented. In the first, all thirteen indices were included, while in the second all those with the best contribution to the value of importance were selected (Figure 4). In the second iteration, one redundancy effect from minor importance indices was assumed. The TCARI/OSAVI, CI_green, TCARI/OSAVI RE, GNDVI, and NDVI indices were included, which together make up 0.8 of the importance value. Given the decrease in the number of indices, these five improved the value of importance; however, they were not as relevant (0.60, 0.06, 0.05, 0.05, and 0.04, respectively). In this second performance, the TCARI/OSAVI registered the highest importance values in the two runs (0.66).
The scatter plot ( Figure 5) shows the relationship between Nt measured in the laboratory and that estimated using thirteen spectral indices, reaching an R 2 of 0.74 with a mean square error (MSE; indicator of model accuracy) of 0.18 (Figure 5a). On the Figure 5b shows that when using the five most important spectral indices, R 2 increased to 0.77 and MSE decreased to 0.15. second iteration, one redundancy effect from minor importance indices was assumed. The TCARI/OSAVI, CI_green, TCARI/OSAVI RE, GNDVI, and NDVI indices were included, which together make up 0.8 of the importance value. Given the decrease in the number of indices, these five improved the value of importance; however, they were not as relevant (0.60, 0.06, 0.05, 0.05, and 0.04, respectively). In this second performance, the TCARI/OSAVI registered the highest importance values in the two runs (0.66).  The scatter plot ( Figure 5) shows the relationship between Nt measured in the laboratory and that estimated using thirteen spectral indices, reaching an R 2 of 0.74 with a mean square error (MSE; indicator of model accuracy) of 0.18 (Figure 5a). On the Figure 5b shows that when using the five most important spectral indices, R 2 increased to 0.77 and MSE decreased to 0.15.

Discussion
During the growing season, any vegetation indices are useful to characterize the phenotypic stages. It has been stated recently that the spectral indices are of fundamental use as a dataset in digital and precision agriculture systems. In determining the Nt [58], the technologies applied to quantify N must be accurate and robust [59,60]. At the production scale, obtaining a more accurate estimation of Nt involves combining methods and processes, in addition to analyzing large datasets-data mining techniques are an example. These technologies are expected to revolutionize agriculture, enabling decision-making in hours or days instead of weeks or months, envisioning a significant reduction in cost and an increase in crop yield [61].
It is broadly recognized and documented that remote sensing has been widely applied in weed mapping, crop growth monitoring, including the health indices of open ecosystems, irrigation management, and many other applications associated with good production practices in most of the production systems. Their use in precision farming practices is diverse and on the rise. The thematic indices show the plant water and/or nutritional needs, plant stress (biotic and nonbiotic), dynamism in time and scale of the geographic space as it occurs, and in general terms, the exchange process among soil cover and the atmosphere. In this manuscript, the indices obtained from high spatial resolution images were useful to estimate the Nt total in forage maize, a crop which is highly sensitive to the N content, but also highly sensitive to the many sources that cause biotic and the abiotic stress, especially during the flowering stage. The RF analysis showed that the TCARI/OSAVI index is the optimal multispectral vegetation index for Nt estimation in forage maize (i.e., R 2 = 0.77). This result is similar to the previous findings of Inoue et al. [62] and Yao et al. [63], who noticed that a combination in indices of NIR and RE bands provides an efficient approach of N status in a plant.

Discussion
During the growing season, any vegetation indices are useful to characterize the phenotypic stages. It has been stated recently that the spectral indices are of fundamental use as a dataset in digital and precision agriculture systems. In determining the Nt [58], the technologies applied to quantify N must be accurate and robust [59,60]. At the production scale, obtaining a more accurate estimation of Nt involves combining methods and processes, in addition to analyzing large datasets-data mining techniques are an example. These technologies are expected to revolutionize agriculture, enabling decision-making in hours or days instead of weeks or months, envisioning a significant reduction in cost and an increase in crop yield [61].
It is broadly recognized and documented that remote sensing has been widely applied in weed mapping, crop growth monitoring, including the health indices of open ecosystems, irrigation management, and many other applications associated with good production practices in most of the production systems. Their use in precision farming practices is diverse and on the rise. The thematic indices show the plant water and/or nutritional needs, plant stress (biotic and nonbiotic), dynamism in time and scale of the geographic space as it occurs, and in general terms, the exchange process among soil cover and the atmosphere. In this manuscript, the indices obtained from high spatial resolution images were useful to estimate the Nt total in forage maize, a crop which is highly sensitive to the N content, but also highly sensitive to the many sources that cause biotic and the abiotic stress, especially during the flowering stage. The RF analysis showed that the TCARI/OSAVI index is the optimal multispectral vegetation index for Nt estimation in forage maize (i.e., R 2 = 0.77). This result is similar to the previous findings of Inoue et al. [62] and Yao et al. [63], who noticed that a combination in indices of NIR and RE bands provides an efficient approach of N status in a plant.
Our results seem to corroborate what was written by Hunt et al. [64] on wheat, that there is a better relationship between the chlorophyll content when the indices utilize the reflectance in the RE wavelengths in comparison to those that do not use it. While some studies about phenotypic characterization and monitoring of corn cultivation have documented the close relationship between the RE band and chlorophyll content [3,5,8,21,22], in this one, the CCCI and NDRE indices showed a weak relationship with Nt; this is contrary to the results documented by authors who report strong relationships with parameters associated with Nt in the plant [65,66]. Nevertheless, this opposite relationship could be associated with conditions of hydric stress, and plant density, as suggested by Rodríguez et al. [67]; though, our results match with those reported by Fitzgerald [68], that the CCCI_simpl index has limited importance in estimating the status of Nt and chlorophyll content in wheat. Comparable results were observed for MCARI/OSAVI and MCARI/OSAVI RE. This study shows the importance of good practices in forage maize to avoid water and nutritional stress.
Results suggest that indices with the RE bandwidth (CCCI_simpl, MCARI/OSAVI, and MCARI/ OSAVI RE) are not associated with a stress condition, and also are poorly associated in estimating the Nt in forage maize. Furthermore, more studies are necessary to record additional meteorological variables to vegetation indices, such as the temperature of the air, and the relative humidity to explore their role in conjunction with the vegetation indices with RE bandwidth with maize growth and Nt estimation.
The very positive result is related to the TCARI/OSAVI index. We found a moderate to a strong relationship in estimating the Nt in forage maize. This result matches a study by Chen et al. [22], where they found a relationship between the TCARI/OSAVI and the Nt in corn (R 2 = 0.66), while Berni et al. [10] reported that the TCARI/OSAVI index minimizes the effect of bare soil and variation of leaf area in olive trees, providing predictive relationships to estimate chlorophyll concentration using multispectral imaging. One study unveiled TCARI/OSAVI as very sensitive to estimate the chlorophyll content and is the least sensitive to variations in the Leaf Area Index (LAI) [69]. The robust relationship can be explained by the inclusion of the RE region where chlorophyll exhibits the lowest absorption rate. Therefore, the use of this band in indices reduces the saturation effect on the index, and the reflectance remains sensitive to moderate to high chlorophyll values [3,8,20]. The component that divides the TCARI index from the OSAVI index is the one that considers the contribution of bare soil to the reflectance value. The combination of both indices improves the estimation of Nt in forage maize.
Due to the evident contrast in the importance value of TCARI/OSAVI concerning the thirteen evaluated indices, a specific pattern is visualized in estimating Nt in forage maize much better than the resulting from indices associated with the crop topology such as GNDVI, NDVI, and RVI. These results seem also to support that these structural indices have a weak prediction power to estimate Nt in essence because of the tendency to exhibit saturation values in moderate to high ground cover conditions [5,70,71].

Conclusions
Monitoring the growing season of forage maize through field data collection and UAV-based images was a technical challenge because of the high demand for storage and computing resources. In this manuscript, we explored the usefulness of RF analysis to demonstrate whether or not Nt in forage maize plants could be estimated using remotely sensed data. Multispectral image indices were selected to reflect the plant phenology and Nt. The data analysis showed the notable association between the TCARI/OSAVI index (obtained from the RE wavelength) and the Nt in forage maize. We observed