Prediction of Maize Yields from In-Season GreenSeeker Normalized Difference Vegetation Index and Dry Biomass as Influenced by Different Nutrient Combinations

To mitigate low maize productivity, improve on-farm planning and policy implementation, the right fertilizer combinations and yield forecasting should be prioritized. Therefore, this research aimed at assessing the effect of applying different nutrient combinations on maize growth and yield and in-season grain yield prediction from biomass and normalized difference vegetation index (NDVI) readings. The research was done in Embu and Kirinyaga counties, in Central Kenya. Nutrient combinations tested were P+K, N+K, N+P, N+P+K, and N+P+K+Ca+Mg+Zn+B+S. The results showed consistently lowest and highest NDVI reading, dry biomass, and grain yields due to P+K and N+P+K+Ca+Mg+Zn+B+S treatments, respectively. Positive NDVI responses of 56%, 14%, 15%, and 15% were recorded with N, P, K, and combined Ca+Mg+Zn+B+S, respectively. These nutrients, in the same order, recorded 54%, 20%, 8%, and 18% positive responses with biomass. The GreenSeeker NDVI reading with grain yield and aboveground dry biomass with grain yield recorded R ranging from 0.23-0.53 and 0.30-0.61 (in Embu), and 0.31-0.64 and 0.30-0.50 (in Kirinyaga), respectively. When data were pooled, the prediction strength increased, reaching a maximum of 67% and 58% with NDVI and biomass, respectively. Yield prediction was even more robust when the independent variables were combined through multiple linear model at both 85 and 105 days after emergence. From this research, it is evident that the effects of balanced fertilizer application are detectable from NDVI readings—providing a tool for tracking and monitoring nutrient management effects—not just from the nitrogen perspective as commonly studied but from the combined effects of multiple nutrients. Also, grain yield could be accurately predicted early before harvesting by combining NDVI and biomass yields.


Introduction
Maize (Zea mays) is a vital crop in the livelihoods of families in Sub-Saharan Africa (SSA). The crop is a source of food, livestock feed, fuel, and thatching materials, among other uses. As a food, maize is the most consumed crop in the region with annual per capita consumption ranging from 31 to 180 kg per person (Awika, 2011;Abate et al., 2015;Kornher, 2018). As a result, food security is always defined based on the availability of maize in these countries. Across Africa, the production has increased in terms of acreages but yields have remained relatively low, less than 2 t ha -1 , under conventional farmers' practices (Otieno et al., 2020). Breaking this cycle of low maize yield and food insecurity requires investments in breeding high yielding and stress-tolerant crop varieties, accurate weather forecasting, optimal soil and water management and other emerging technologies that optimize resource use. In responding to this need, researchers have come up with various interventions ranging from soil acidity management Fontoura et al., 2019), manure application (Naramabuye et al., 2008;, inorganic fertilizer application Otieno et al., 2020), to soil water management through reduced tillage and mulching (Murungu et al., 2011;Otieno et al., 2020). Most of these strategies and technologies have resulted in increased grain yield. Farmers and policymakers always wait until the dry harvesting stages to estimate the yields before proceed to draft and implement new plans and policies in the region. This method of assessing and measuring yields after harvesting usually comes late, leading to poor food insecurity mitigation planning and budgeting by governments and policy-makers. Thus, researchers are coming up with strategies to help in the early detection of possible constraints and likely expected yields based on in-season crop behaviors-yield forecasting. The importance of yield forecasting has been summarized by Habyarimana et al. (2019): provides data to governmental structures, companies, and farmers, which results in strategic advantages such as the rationalization of policy adjustments, price predictions and stabilization, efficient agricultural trade, and simplification of business operations particularly through planning harvest and delivery of the product, better deployments of machinery and logistics, and better management at the end-user level. The commonly used methods of weather, pest, disease and yield forecasting are crop modeling and remote sensing. These forecasting methods use parameters such as normalized difference vegetation index (NDVI), leaf area index, and fraction of absorbed photosynthetically active radiation (fAPAR) (Diouf et al., 2015;Kross et al., 2015;Ngoune et al., 2020). These technologies have evolved and converted into simpler farm tools and equipment for daily use by farmers. For instance, GreenSeeker NDVI equipment, a cheap hand-held remote sensing tool farmers are currently using to make in-season assessment of daily crop health (Verhulst & Govaerts, 2010;Sultana et al., 2014;Kitić et al., 2019;Ngoune & Mutengwa, 2020). However, farmers in Africa, and Kenya in particular, have not been able to use the GreenSeeker NDVI tool to assess the health of their crops and make rapid yield predictions early in the season for prompt farm budgeting and decision making. Thus the region is left out in the use of the technology. And this has exposed farmers and the entire population to chronic food insecurity that would otherwise be managed to some extent. Several researchers have used GreenSeeker NDVI equipment in fertilizer management and yield forecasting tool-reporting significant positive relationship between NDVI and crop N demand (Xia et al., 2016;Ali et al., 2018), biomass prediction (Xia et al., 2016) and grain yield prediction (Sultana et al., 2014;Fernandez-Ordoñez & Soria-Ruiz, 2017). This shows the usefulness of the tool in nutrient management and yield forecasting. In terms of plant health and nutrient management, however, most research has focused on nitrogen use efficiency only (Teboh et al., 2012;Quebrajo et al., 2015;Vergara-Díaz et al., 2016), leaving other nutrients unaccounted for in balanced nutrient requirements for improved crop production. Again, a few researches have looked at the effects of different nutrient combinations on crop's NDVI at different growth stages and how this translates to yield. This research therefore, aimed at investigating this effect. Again, researchers have shown relationships between crop NDVI and biomass and NDVI and grain yield through linear regression models. However, there are no evaluations done to show the effect of combining NDVI reading with its corresponding biomass on grain prediction in Sub-Saharan Africa. This gap could be explored for possible stronger yield predictions. Due to the above research gaps, this research therefore, aimed at assessing the effect of applying different nutrient combinations on maize growth and yield. It also evaluated the potential of in-season grain yield prediction from biomass and NDVI recording. The combination of different nutrients at plot level is important as it, to some extent, portrays the likely heterogeneity in maize growing conditions and interactions between nutrients between farms that have always complicated the expression in NDVI reading.

Description of the Study Site
The trials were carried out in Kenya Agricultural and Livestock Research Organization (KALRO), Embu research station located in Embu County (Referred as Embu hereafter), and Kirinyaga Technical Institute (KTI) research fields located in Kirinyaga County (Referred as Kirinyaga hereafter). These sites cover agriculturally important zones where farmers predominantly grow maize as a source of food. The sites were located in the Upper Wet Mid Altitude Mega-environment. The sites are characterized by bi-modal rainfall patterns, experiencing wet seasons from March to June (long rain season) and September to December (short rain season). The annual rainfall ranges from 930 mm to 1550 mm. The daily mean temperature is about 18 ºC in Embu and 23 ºC in Kirinyaga. The soils in these sites are predominantly Humic Nitisols with clay-loam texture, deep and good water-holding capacity (Jaetzold & Schmidt, 1983). Other site-specific soil fertility characteristics of the study sites were as reported by Otieno et al. (2020). The research was done during the 2013/2014 short rains and 2014 long rains seasons.

Experimental Design and Treatments
The experiment was laid out in a randomized complete block design with each treatment replicated six times. Each plot measured 8 m × 10 m with a space of 1.5 m and 1 m left between blocks and plots, respectively. Between blocks, a trench of 1 m wide and 1 m deep was dug to reduce the chances of nutrients flowing within the soil profile from one plot to the other. The treatments comprised of different nutrient combinations: P+K, N+K, N+P, N+P+K, and N+P+K+Ca+Mg+Zn+B+S. Nitrogen (N), phosphorus (P), potassium (K), calcium (Ca), magnesium (Mg), zinc (Zn), boron (B), and sulfur (S) nutrients were applied at the rates of 120, 40, 40, 10, 10, 5, 5 and 26.3 kg ha -1 , respectively. The nutrients were supplied from urea, triple superphosphate, muriate of potash, calcium sulfate, magnesium sulfate, zinc sulfate, borax, and sulfate sources, respectively. These rates were chosen to ensure maize growth was no limited by nutrients and to target at least 6 tons of grains per hectare. Nitrogen was applied in three equal splits (at planting, V4, and V10 stages of maize vegetative growth) while the rest of the nutrients were applied at planting. Maize variety, DK 8031, was selected and used for the trials in all sites. This maize variety was selected due to its extensive use in the region and adaptability to the prevailing climatic conditions.

Agronomic Practices
The research was done during the 2013/2014 short rains and 2014 long rains seasons. During the 2013/2014 short rain season, DK 8031 maize variety was planted to deplete nutrients from the plots to reduce huge variability due to already present nutrients. Tilling of plots was done a week to the 2014 long rain season using hand-hoes. After three consecutive rains, maize planting was done at 75 cm by 25 cm spacing using a calibrated planting string. At planting, fertilizers were placed in planting holes then mixed with soil before placing seeds to avoid direct contact with fertilizer. Two maize seeds were planted per socket and thinned to one plant per socket seven days after emergence to maintain a population of about 53,000 plants per hectare. The first and second weeding and topdressing (on plots that received N) were done at V4 and V10 stages of maize growth. Pests and diseases were monitored regularly. At 30 days after emergence, Bulldock (Beta-Cyfluthrin 0.5 g/kg) pesticide was applied at the rate of 6 kg ha -1 to control stalk borers. During pesticide application, all protection measures as outlined by  were observed. After maturity stage, dried cobs were harvested manually.

Data Collection
Maize Normalized Difference Vegetation Index (NDVI): Maize NDVI measurements were taken with GreenSeeker™ Handheld Optical Active Sensor (Trimble Navigation Limited, Sunnyvale, California, USA). The sensor emits brief bursts of red and infrared light and then measures the amount of each type of light that is reflected back from the plant; the measuring process continues as long as the trigger remains engaged (https://agriculture.trimble.com). The NDVI reading (ranging from 0.00 to 0.99) is displayed on the LCD screen of the equipment. The strength of the detected light is used to indicate the crop health; the higher the reading, the healthier the plants could be assumed to be. The NDVI measurements were taken at 40, 65, 85, and 105 days after emergence (DAE) in the central rows of all plots. Three readings were taken within each plot, leaving two maize rows from both edges. These readings were then averaged to give a plot reading.
Biomass production: Aboveground biomass production was assessed at 40, 65, 85, and 105 DAE. Biomass production from each treatment was computed from a sub-plot measuring 4.69 m 2 and a subsample containing chopped leaves and stalks weighing 500 g dried at 65 ºC to a constant dry weight. These weights were then used to compute dry biomass production per hectare using Dobermann and Walters (2005) formula.
Grain yield: Yields were computed from a net plot measuring 3.75 m by 4 m (15 m 2 ) taken from the center of each treatment plot leaving at least 2 m on each side of the net plot to minimize the edge effects. After harvesting, total plants and cob numbers were recorded, and total cob weight was determined in the field using a digital scale accurate to 2 decimal places. All cobs were shelled, mixed thoroughly, and a sub-sample of 1 kg grain (fresh weight) taken for further drying to a constant weight at 12% moisture content (dry weight). These weights were then used to compute grain yield production per hectare.

Statistical Analysis
Collected data were subjected to analysis of variance (ANOVA) using Genstat statistics software, 15th version. Where F tests were significant, means were compared using Fisher's protected least significance difference (L.S.D.) procedure at p ≤ 0.05. The NDVI and biomass averages were then used to assess nutrient responses for individual nutrients and their combinations. Several simple and multiple linear regression models were investigated and compared for each site and pooled data. These regression models were done to establish the relationship between NDVI, biomass, and grain yield. Graphical presentations were done using excel package.

Effect of the Site and Nutrient Combinations on GreenSeeker Normalized Difference Vegetation Index (NDVI)
In Kirinyaga, the NDVI readings were 0.05 and 0.02, significantly higher than those recorded in Embu at 40 and 65 DAE respectively. However, this changed at 85 and 105 DAE, where Embu recorded 0.01 and 0.15 higher readings. The change in NDVI recordings was because during the first eight weeks after planting, Kirinyaga site received higher rainfall than Embu site, after which the latter site received more rainfall than the former site

Effect of Site and Nutrient Combinations on Aboveground Dry Biomass Production
The site significantly influenced dry biomass production (p < 0.001) ( Figure 4) and by fertilizer application (p < 0.001) at 65, 85 and 105 DAE only (Table 2). Biomass production at 40 and 65 DAE was 0.19 and 1.8 t ha -1 , respectively, higher in Kirinyaga than in Embu while at 85 and 105 DAE, the trend changed and biomass was 1.15 and 1.42 t ha -1 , respectively, higher in Embu than in Kirinyaga (Figure 4). This trend is similar to that observed with NDVI readings and is attributed mainly to rain variations between sites.
The biomass increased from 40 DAE and peaked at 85 DAE before decreasing towards 105 DAE. This trend was observed both in Embu and Kirinyaga. This finding agrees with that reported by Otieno (2019) while evaluating the growth and yield response of maize to a wide range of nutrients on ferralsols of western Kenya. Similarly, in Central Brazil, Baldé et al. (2011) reported an increase in maize leaf area which peaked between 80-100 days before declining in size towards 180 days after planting. As cells increase in size and multiply in number, maize plants grow and thus increase in size. Consequently, the leaf area increases in size and number, and more photosynthates are accumulated resulting in high biomass production until the maximum size is attained (Bair, 1942;Kohl et al., 2017).
At 40 DAE, the effect of applying different nutrient combinations yielded non-significant differences in aboveground dry biomass. This could be due to low nutrients demanded by young maize seedlings. Hence the amounts that were supplied by the soil were optimal in keeping the same growth rate.  Figure 5. Maize biomass yield response to N, P, K, and combined secondary and micronutrients (Mg+Ca+S+Zn+B) application. The responses were calculated after pooling the data across Embu and Kirinyaga sites The interaction between the site (S) and nutrient combination (NC) did not result in a significant difference in biomass production (

Effect of the Site and Nutrient Combination on Maize Grain Yield
Site and nutrient combinations significantly affected grain yields (Table 3). The N+P+K+Zn+B+Mg+Ca+S treatment generally had significantly higher grain yield than N+K, N+P, and P+K across all sites and NPK treatment in Embu. The P+K treatment had a lower grain yield than N+P treatment at Kirinyaga. Nutrient combinations N+K, N+P, P+K, and N+P+K, were not significantly different in grain yield at Embu. At Kirinyaga site, no significant differences were recorded among N+P+K+Zn+B+Mg+Ca+S, N+P+K, N+K, and N+P treatments. There was no significant interaction effect observed between site and nutrient combinations. The positive effect of combining primary, secondary, and trace nutrients on grain yields has been confirmed in Ghana by Kugbe et al. (2019)

Predicting Maize Grain Yield from GreenSeeker Normalized Difference Vegetation Index (NDVI) Reading and Aboveground Dry Biomass
The in-season precision of predicting grain yield varied between sites and independent variables considered (Table 4). In all sites, there were significant positive relationships between grain and GreenSeeker NDVI reading and between grain and biomass. The GreenSeeker NDVI readings and aboveground dry biomass produced R 2 ranging from 0.23-0.53 and 0.30-0.61 (in Embu), and 0.31-0.64 and 0.30-0.50 (in Kirinyaga) respectively. The use of NDVI reading in predicting grain yields has been reported by several researchers (Sultana et al., 2014;Fernandez-Ordoñez & Soria-Ruiz, 2017;Maresma et al., 2020). The pooled GreenSeeker NDVI readings and aboveground biomass data recorded significant positive prediction of grain yield ( Figure 6). The GreenSeeker NDVI were significant at 65 (p < 0.0001), 85 DAE (P = 0.0185) and 105 DAE (P < 0.0001) while aboveground dry biomass was significant at 65, 85 and 105 DAE (P < 0.0001). Both GreenSeeker NDVI reading and biomass showed an increasing strength in predicting maize grain as the measurements were taken towards crop maturation; R 2 ranged between 0.0007 (at 40 DAE) and 0.6683 (at 105 DAE) in the case of NDVI and between 0.0077 (40 DAE) and 0.57 (at 105 DAE) in the case of dry biomass ( Figure 6). These R levels are well within the ranges reported by other researchers, 0.32-0.78 (Sultana et al., 2014;Naser et al., 2020). Whether at the individual sites or pooled data, stronger yield predictions were recorded from those variables collected towards the reproductive stages from 85 days and the best at 105 days after emergence. These findings resonate with those reported by Maresma et al. (2020) who concluded that best yield predictions are obtained by scanning maize at or after V10 stage of growth. Fernandez-Ordoñez & Soria-Ruiz (2017) also found strong yield prediction when NDVI was recorded at flowering. During the assessment of the usefulness of spectral reflectance indices as durum wheat yield predictors under contrasting Mediterranean conditions, Royo et al. (2003) concluded that the milky-grain stage is the best depictive stage for recording NDVI as it is more directly related to yield than earlier measurements.
When NDVI was combined with biomass collected at corresponding growth stages, the strength of grain prediction increased tremendously (Table 5) compared to when the relationship was considered at the individual site levels (Table 4 and Figure 6). In-season prediction of grain yield was very strong from 85 DAE (adjusted R = 0.706) to 105 DAE (adjusted R = 0.841). This could be due to the synergy resulting from the individual variables all linking towards grain prediction. Although there is no previous work showing this kind of prediction, Royo at al. (2003) found that combining NDVI with other parameters like reflectance at 550 nm (R550), water index (WI), photochemical reflectance index (PRI), structural independent pigment index (SIPI), and simple ratio (SR) explained a 95.7% of yield variability jointly when all the experiments were analyzed together compared to 17-65.2% when regressions were analyzed separately.