Monitoring impacts of soil bund on spatial variation of teff and finger millet yield with Sentinel-2 and spectroradiometric data in Ethiopia

Abstract

Data from remote sensing devices are essential for monitoring environmental protection practices and estimating crop yields. However, yield estimates in Ethiopia are based on time-consuming surveys. We used Sentinel-2, spectroradiometeric, and ground-truthing data to estimate the grain yield (GY) of two major crops, teff, and finger millet, in Ethiopia's Aba Gerima catchment in 2020 and 2021. At the flowering stage, we performed supervised classification on October Sentinel-2 images and spectral reflectance measurement. We used regression models to identify and predict crop yields, as evaluated by the coefficient of determination (adjusted R²) and root mean square error (RMSE). The enhanced vegetation index (EVI) and normalized-difference vegetation index (NDVI) provided the best fit to the data among the vegetation indices used to predict teff and finger millet GY. Soil bund construction increased the majority of vegetation indices and GY of both crops. We discovered a strong correlation between GY and the satellite EVI and NDVI. However, NDVI and EVI had the greatest influence on teff GY (adjusted R² = 0.83; RMSE = 0.14 ton/ha), while NDVI had the greatest influence on finger millet GY (adjusted R² = 0.85; RMSE = 0.24 ton/ha). Teff GY ranged from 0.64 to 2.16 ton/ha for bunded plots and 0.60 to 1.85 ton/ha for non-bunded plots using Sentinel-2 data. Besides, finger millet GY ranged from 1.92 to 2.57 ton/ha for bunded plots and 1.81 to 2.38 ton/ha for non-bunded plots using spectroradiometric data. Our findings show that Sentinel-2- and spectroradiometeric-based monitoring can help farmers manage teff and finger millet to achieve higher yields, more sustainable food production, and better environmental quality in the area. The study's findings revealed a link between VIs and soil management practices in soil ecological systems. Model extrapolation to other areas will necessitate local validation.

1. Introduction

Crop yields in Ethiopia are estimated based on sampling surveys, expert assessments, and farmer reports. Furthermore, using conventional yield estimation techniques in most African countries, including Ethiopia, is costly, time-consuming, and prone to large errors [1,2]. For large areas, manual crop measurements are prohibitively time-consuming and expensive. Furthermore, national crop production projections are finalized months after a crop harvest, making it difficult to make prompt import and export decisions to assure food security and economic growth. As a result, it is challenging to pinpoint food shortages and guarantee food security in the nation. As a result, there are discrepancies between actual production and projected crop yields from the government. Furthermore, due to insufficient field observations, intercropping techniques, and limited field sizes, it is difficult to accurately measure crop output and its spatial variation [3].

Despite the aforementioned drawbacks, recent research has demonstrated that Sentinel-2 photos can be used to predict yields on specific plots for a variety of crops, such as maize, wheat, and millet [4,5]. For agricultural operations, estimating crop output and planted area is essential. Predicting agricultural yields before harvest is crucial for organizing and deciding on various policy options. Governments also gain from yield forecast data, which supports the improvement and reform of many different areas such as agricultural trade, sustainable use of water and soil resources, and development and reform of food security [6]. Moreover, data on agricultural productivity is crucial for humanitarian organizations that provide food relief. In order to estimate agricultural growing areas and to obtain primary data for yield estimation, remote sensing technology must be used [7,8].

Besides, freely available and high-resolution data have recently demonstrated significant potential for more precise extraction of vegetation properties [9], creating opportunities to estimate the crop yield [[10], [11], [12], [13]], and crop monitoring [14,15]. Agricultural development and environmental impact analysis have all made extensive use of remote sensing data. The satellite data can also provide critical information about crop yield at a low cost [2]. In general, remote sensing offers opportunities for rapid and accurate monitoring and evaluation of crop growth and yield of crops and maximizes the productivity of the cropland [16,17].

Moreover, remote sensors such as Sentinel-2 imagery and spectroradiometer are efficient tools for evaluating crop growth and estimating yields at different tempo-spatial scales [14,18]. The Sentinel-2 satellite collects data that can be used to forecast crop leaf area index (LAI), leaf water content, and chlorophyll content [19]. Many studies reveal that remote sensing-based crop yield estimation provides more reliable and more accurate results than the conventional method [12,20].

Sentinel data has been widely used in agricultural development as well as monitoring and analyzing environmental impact. The Sentinel-2 data allow for frequent long-term crop growth monitoring [21,22], classifying land cover [8,23,24] as well as estimating crop yield [25,26]. Crop monitoring with Sentinel-2 and crop canopy data revealed that wheat grain yield (GY) is strongly related to crop growth [27]. The presence of non-leaf photosynthetically active organs such as spikes and leaf sheaths, as well as changes in leaf erectness, may, however, affect crop spectral values [28]. Furthermore, because crop reflectance varies across spectral bands, it is necessary to account for the reflectance in multiple bands to obtain useful information. Lastly, in Ethiopia, the potential for mapping yield variability using Sentinel-2- and spectroradiometric-based data has not been assessed.

Researchers calculate crop growth and yield parameters using remote sensing-based vegetation indices (VIs), which have become effective tools for mapping crops and estimating crop GY at a low cost [21,29]. However, the accuracy of the crop yield modeling depends upon the accurate selection of VIs [9]. So, it becomes vital to investigate the effect of different VIs for accurate monitoring of crop yield. Bunds increase soil fertility by reducing runoff and soil loss, thereby preserving environmental sustainability [30]. Furthermore, soil bunds can help to restore degraded soil quality and functions while also ensuring long-term production and environmental quality. In the Aba Gerima catchment, soil erosion and depleted soil productivity have been major issues, resulting in low crop yields. Furthermore, replacing traditional crop surveys with more accurate and low-cost methods, such as remote sensing-derived approaches, would be advantageous.

In this case, we wanted to look into using Sentinel-2 satellite and spectroradiometric-derived data to predict crop yield in Ethiopia. We chose the Aba Gerima catchment in Ethiopia, where soil erosion and depletion of fertility have been major causes of lower crop yields. Bund construction has been carried out in the catchment area as part of extension programs [31]. However, due to the cost and time required, their effects on crop productivity have not been studied, which could be solved using a remote sensing approach. Moreover, in Ethiopia, where food production remains at subsistence farming levels, the yield estimation methods could not be benefited from the advancement of high-resolution data (spectroradiometer) and freely available Sentinel-2 data. Besides, crop yield prediction accuracy depends on many different factors such as climate, weather, soil, use of fertilizer, and seed variety [32]. Better performance in yield prediction is also crucial [33].

To decrease the work required during data collection and the disparities between actual and predicted yield in the research area, it is crucial to develop timely and precise crop-yield estimation systems. The current study used various vegetation indices to estimate agricultural yields using red and NIR reflectance [2,34]. Therefore, calculating agricultural yields at a specific crop growth stage using spectral data could be quite important. In addition to the advantages listed above, the vegetations suggested in this study can be utilized by farmers and agricultural policy-makers to monitor crop development and soil management condition in real-time. Thus, the objective of this research was to assess the effects of bund construction on GY of the key growth stage of major crops (teff and finger millet) using Sentinel-2- and spectroradiometric-derived data in the Aba Gerima catchment.

2. Materials and methods

2.1. Description of the research area

The Aba Gerima catchment is located in the Blue Nile Basin, which as altitudes ranging from 472 to 4261 m above sea level (Fig. 1a). Specifically, it is situated in Ethiopia's Amhara Region, between 11°39′0″ N and 11°40′30″ N and between 37°30′0″ E and 37°31′0″ E (Fig. 1b). The site's elevations range from 1914 to 2121 m above sea level (Fig. 1b). The distribution of the sample sites is also depicted in Fig. 1c.

Fig. 1.

Location of study catchment and crop sampling plots: (a) The Blue Nile Basin altitude, (b) Aba Gerima topography, (c) Crop validation plots across land-use type. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

The study area receives a mean annual rainfall ranging from 1076 to 1953 mm, with a mean monthly maximum temperature of 27.0 °C and a mean monthly minimum temperature of 12.6 °C, according to records from 1994 to 2021 at nearby meteorological stations (Fig. 2; S1 table). The rainy season lasts from June to September; otherwise, it is dry [35]. In general, understanding such long-term rainfall and temperature variability and trends within the study area could have significance in potential agricultural production and crop yield prediction. Besides, their influence can be seen in the discrepancy between what we would have predicted and what happened. The catchment's common soil is Acrisols [36]. Teff (Eragrostis tef. Zucc.), finger millet [Eleusine coracana (L.) Gaertn]), and maize (Zea mays L.) are important crops.

Fig. 2.

Long-term (1994–2021) monthly rainfall (RF), maximum temperature (Tp_max), and minimum temperature (Tp_min) in the study area.

3. Methods of data collection

3.1. Experimental setup and crop sampling

We used the digital elevation model (DEM) in ArcGIS v. 10.5.1 to map locations within the Aba Gerima catchment (www.esri.com). We generated three topographic classes (gently sloping (2–5%), medium sloping (5–10%), and strongly sloping (10–15%)) of the watershed using a high-resolution (0.5 × 0.5 m) DEM purchased from the Advanced Land Observing Satellite-2 (ALOS-2) operated by the Japan Aerospace Exploration Agency (JAXA; http://en.alos-pasco.com/).

We created two land management scenarios: 48 plots without soil bunding (WB) and 48 plots with soil bunds (SB). Thus, ninety-six representative crop-sampling plots with a minimum size of 40 m × 40 m (1600 m²) each were identified for this study. The bunds are five years old, with a bottom width of 0.8 m and a height of 0.5 m. All plots were distributed on purpose (Fig. 1c). Each plot was geo-coded using a handheld GPS device (GPSMAP64, Garmin, Olathe, Kansas, USA).

An ox-drawn plow was used to cultivate the soil in each sample plot to a depth of 20 cm. Teff seeds (at a rate of 25 kg/ha) were sown on June 15/2020 and June 22/2021 and finger millet seeds (at a rate of 25 kg/ha) were also sown on April 26/2020 and May January 2021. Throughout the growing season, weeds were manually controlled. To control crop insect pests, no agricultural pesticides were used. At the time of sowing, we applied urea and di-ammonium phosphate fertilizer (at a rate of 100 kg/ha) for each crop.

3.2. Sentinel-2 and spectroradiometric data collection

From the Copernicus Open Access Hub (https://sentinels.copernicus.eu/), we downloaded cloudless Sentinel-2 images covering the Aba Gerima catchment during the flowering stage from 1 to October 31, 2020 and 2021; we used the 2020 data for model development and the 2021 data for model validation. Preprocessing Sentinel-2 data with radiometric and atmospheric correction was used to increase the image quality and model robustness [37,38]. To supervise the satellite images and map cover by forest, shrub/bushland, maize, teff, and finger millet, we used approximately 32 ground-truthing training sites (Fig. 3a). Dataset was used to calibrate (N = 96) and validate the VIs-based model (N = 96). Teff and finger millet plants were submitted to spectral reflectance recording at flowering stages (September October 2020 and 17/2021 for teff and August 20/2020 and 28/2021 for finger millet) using FieldSpec IV spectroradiometer (Analytical Spectral Devices [ASD] Inc., Boulder, CO, USA, 350–2500 nm; Fig. 3b).

Fig. 3.

Crop reflectance and ground data measurement and sampling: (a) Teff and finger millet stands, (b) Teff and finger millet reflectance, and (c) Teff and finger millet ground data collection.

Several VIs have been developed to characterize vegetation canopies. The most common of these indices, which utilize red and NIR wavelengths were used in this study. These indices are well correlated with various vegetation parameters including LAI, biomass, productivity, and photosynthetic activity [[39], [40], [41]]. The use of VIs provides on-time predictions of the crops’ productivity and does not require extensive in-field measurements for the forecasting [42]. Thus, we used the VIs (Table 1) and calculated them using ArcGIS' raster calculator. The workflow for estimating grain yield (GY) of teff and finger millet is depicted in Fig. 4.

Table 1.

Description of the vegetation indices used in the Aba Gerima catchment.

Index	Sentinel-2 bands and equations	References
EVI	$2.5\left[\frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{r}\mathrm{e}\mathrm{d}}{\mathrm{N}\mathrm{I}\mathrm{R}+6\times \mathrm{r}\mathrm{e}\mathrm{d}-7.5\times \mathrm{b}\mathrm{l}\mathrm{u}\mathrm{e}+1}\right]$	[43]
NDVI	$\frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{r}\mathrm{e}\mathrm{d}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{r}\mathrm{e}\mathrm{d}}$	[44]
SAVI	$1.5\left[\frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{r}\mathrm{e}\mathrm{d}}{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{r}\mathrm{e}\mathrm{d}+0.5}\right]$	[45]
GCVI	$\frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}-1$	[46]
GNDVI	$\frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}{\mathrm{R}\mathrm{e}\mathrm{d}+\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}$	[47]

Abbreviations: blue (band 2) represents wavelengths of 460–520 nm, green (band 3) 540 to 580 nm, red (band 4) 630 to 690 nm, and near-infrared (NIR, band 8) 760 to 900 nm. Vegetations indices (VIs): EVI enhanced VI, NDVI normalized-difference VI, SAVI soil-adjusted VI, GCVI green chlorophyll VI, GNDVI green NDVI.

Fig. 4.

Workflow for estimating Sentinel-2 and spectroradiometric-based crop yield.

3.3. Actual grain yield measurement

The main phenological stages of teff are germination (June 25–28), leaf development (July), stem elongation (August–September), flowering September 8–20), and ripening (September 21- October 5). While germination (May 5–10), leaf development (May), stem elongation (June 10 - July 15), flowering September (July 20 - August 5), and ripening (August 15 - September 07) are main growth stages of finger millet. Upon maturity, all teff plants were harvested (Fig. 3c) on November 1st, and all finger millet plants were harvested on November 25th/2020 and 2021 and air-dried for one week. Thousand-seed weight for both crops was measured using a standard weighing balance (gm/mg). A hand-held grain moisture tester was used to determine the moisture content of the seeds (model AG-12, A-Grain, India). GY (kg/ha) was calculated as seen in equation (1).

$$\mathrm{G}\mathrm{Y}\left(\frac{\mathrm{k}\mathrm{g}}{\mathrm{h}\mathrm{a}}\right)=\frac{\mathrm{G}\mathrm{Y}\mathrm{P}\mathrm{P}\times 10000\times \left(100-\mathrm{G}\mathrm{M}\right)}{\mathrm{H}\mathrm{A}\times 0.88}$$ Eq.1

Where, GYPP = grain yield per plot (kg), GM = grain moisture content at harvest (%), and HA = harvested area (m²).

3.4. Data analysis

Linear regression is the second most used algorithm [48]. In SAS v. 9.4, thus, we used the analysis of variance applied to stepwise linear multiple regression equations to identify the VIs that contributed significantly to the prediction of GY. We calculated the coefficient of determination (adjusted R²) and the root mean square error (RMSE) using equations Eq.2, Eq.3), respectively. The best-fit models were also chosen based on high R² and low RMSE between crop yields and the proposed VIs. Using the best-fit regression models, we generated yield maps for teff and finger millet in the study area under ArcGIS platform.

$${\mathrm{R}}^2=\frac{\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{ŷ}}_{\mathrm{i}}-{\mathrm{ӯ}}_{\mathrm{i}}\right)}^2}{\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{y}}_{\mathrm{i}}-{\mathrm{ӯ}}_{\mathrm{i}}\right)}^2}$$ Eq.2

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\frac{1}{\mathrm{N}}\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{ŷ}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}\right)}^2$$ Eq.3

Where, ŷ = predicted value; ӯ = mean measured value; y = measured value; and N = number of observations with i = 1, 2 … N.

4. Results and discussion

4.1. Principal component analysis

The data's suitability for the principal components analysis was checked by the Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy and Bartlett's Test of Sphericity. The KMO value for vegetation indices of teff and finger millet were 0.64 and 0.79, respectively which are greater than the recommended value of 0.5. Bartlett's Test of Sphericity showed an approximate chi-square value equal to 103.52 and 141.69 for teff’ and finger millet's SVIs, respectively with a statistical significance at p < 0.001. Thus, the KMO confirms the sampling adequacy for the data set suitable for PC analysis. Furthermore, Bartlett's test of sphericity, the very small p-value (<0.001), revealed that PCA could be performed efficiently on the dataset.

As shown in Table 2, this study identified two PCs that recorded eigenvalues greater than one and cumulative variance greater than 70%. The two PCs can capture the most spectral variations [49,50]. The total variances for the two PCs were 82.69% and 74.67% for teff and finger millet, respectively. The EVI with loading (0.91) and GNDVI (0.79) for teff and EVI (0.89) and GNDVI (0.91) for finger millet were the most influential contributors (Table 3). This information helps to assess crop yields with minimum costs and monitor changes related to the different land management practices on crop yields.

Table 2.

Eigenvalues and variance explained by principal components of the Teff's Sentinel-2 vegetation indices and finger millet's spectroradiometric vegetation indices.

Crop	Component	Extraction sums of squared loadings			Rotation sums of squared loadings
		Total	Variance (%)	Cumulative (%)	Total	Variance (%)	Cumulative (%)
Teff	1	2.41	48.27	48.27	2.37	47.58	47.58
Teff	2	1.72	34.42	82.69	1.75	35.11	82.69
Finger millet	1	2.65	53.07	53.07	1.92	38.45	38.45
Finger millet	2	1.08	21.59	74.67	1.81	36.22	74.67

Table 3.

Rotated components of loading of the vegetation indices of teff and finger millet.

Finger millet			Teff
Index	Principal component		Principal component
	1	2	1	2
EVI	.90	−.02	.89	.10
NDVI	.60	.69	.88	.12
SAVI	.84	.29	.84	.26
GNDVI	.11	.79	−.20	.91

Extraction method: Principal component analysis; Rotation method: Varimax with Kaiser normalization; EVI, enhanced vegetation index; NDVI, normalized difference VI; SAVI, soil-adjusted VI; GNDVI, green NDVI.

4.2. Calibrating and validating yield prediction models

We analyzed the potential of each VI (Sentinel-2 and spectroradiometric data) and their best combinations using adjusted R² and RMSE to predict teff and finger millet yields. In this study, we regressed each VI with GY and combine those VIs having higher R² (greater than 0.50) and lower RMSE values [[51], [52], [53]]. Thus, tor teff (N = 48 in 2020), the individual contributors were EVI (adjusted R² = 0.68; RMSE = 0.19 ton/ha) and NDVI (adjusted R² = 0.70; RMSE = 0.18 ton/ha) using Sentinel-2 data. NDVI and EVI extracted from Sentinel-2 data better explained teff yield variability.

However, the equation for best combinations for teff GY with the highest adjusted R² (0.81) and close to the lowest RMSE (0.14 ton/ha) using Sentinel-2 data was:

$$\mathrm{G}\mathrm{Y}=5.60\times \mathrm{E}\mathrm{V}\mathrm{I}+3.13\times \mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}+1.05\times \mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}-3.92$$

This indicates that a predictive model combining multiple VIs was more accurate than models based on a single VI in predicting teff GY. Thus, we chose it to validate and predict spatial teff yield variability.

For finger millet GY (N = 48 in 2020), the individual contributors were EVI (adjusted R² = 0.61; RMSE = 0.45 ton/ha) and NDVI (adjusted R² = 0.68; RMSE = 0.41 ton/ha) using Sentinel-2 data. The best combination fit (adjusted R² = 0.88, RMSE = 0.26 ton/ha) was: GY = 5.74 × EVI + 2.50 × NDVI + 1.13 × GNDVI – 3.31.

Using spectroradiometric data, the individual contributors were EVI (adjusted R² = 0.75; RMSE = 0.12 ton/ha) and NDVI (adjusted R² = 0.61; RMSE = 0.14 ton/ha) for teff in 2020. The best combination fit (adjusted R² = 0.76, RMSE = 0.11 ton/ha for teff):

$$\mathrm{G}\mathrm{Y}=2.62\times \mathrm{E}\mathrm{V}\mathrm{I}+0.54\times \mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}+0.54$$

For finger millet GY (N = 48 in 2020), the individual contributors extracted from spectroradiometric data were NDVI (adjusted R² = 0.76; RMSE = 0.36 ton/ha) and SAVI (adjusted R² = 0.76; RMSE = 0.37 ton/ha). However, the best combination fit (adjusted R² = 0.91, RMSE = 0.22 ton/ha for finger millet) was:

$$\mathrm{G}\mathrm{Y}=4.20\times \mathrm{E}\mathrm{V}\mathrm{I}+0.11\times \mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}+1.49\times \mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}+5.69\times \mathrm{G}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}–5.66$$

This shows that a predictive model combining multiple VIs (developed from spectroradiometric data) was more accurate than models based on a single VI in predicting finger millet GY at the catchment scale. We chose it to validate and predict spatial finger millet yield variability in Aba Gerima watershed. Thus, we can conclude that combining canopy development metrics improved crop yield prediction accuracy. This result is in line with reports made [54].

To validate the spectral model, we plotted the observed yield (ton/ha) against the estimated yield (ton/ha) at the catchment scale (Fig. 5). We tested the linear regression models on the independent validation set (S2 table). The scatterplots best showed R² = 0.83 and RMSE 0.14 ton/ha for teff GY using Sentinel-2 data and R² = 0.84 and RMSE = 0.07 ton/ha for finger millet GY using spectroradiometric data. The magnitudes between the calibration and validation dataset of finger millet was consistent, but not for teff vegetation indices datasets. The growth stage, nutrient availability, and water status variables that impact leaf area [55], crop row orientation [56,57], and plant canopies could be the cause of these variations. Additionally, the model's resilience, accuracy, repeatability, and reproducibility were improved by preprocessing Sentinel-2 and spectroradiometric data with radiometric and atmospheric adjustment and exclusion of noisy spectral areas [37,38]. In light of these findings, it appears that the models for estimating crop yields from spectroradiometeric and Sentinel-2 data at broad spatial scales are accurate and may be repeatable in the study area conditions.

Fig. 5.

Scatterplots of the model validation results for GY teff using spectroradiometric (a), Sentinel-2 data (b), and GY finger millet using spectroradiometric (c) and Sentinel-2 data (d). N, number of observations; adjusted R², coefficient of determination; RMSE, root mean square error; GY, grain yield and broken line represents an iso-yield line.

Overall, compared to calibration datasets, the independent validation approach was more accurate. Validation results indicated that developed models accurately predicted the characteristics of independent samples [58]. As opposed to independently calibrated models produced from the independent validation approach, validation derived models in predicting new independent samples had significantly lower accuracy [59]. Therefore, relying on calibrated models before using them on an independent test set to evaluate the model is not entirely practical. To reduce the risk of very low and high GY prediction Thus, validation is necessary before application at very large scales.

The EVI (Fig. 6), NDVI (Fig. 6), and SAVI (Fig. 7) were the greatest predictors of teff and finger millet yields. The amount of green pigments in the most bunded teff and finger millet plants may be the cause of the significant EVI and NDVI contributions. It might also be brought on by the red light absorption and NIR reflection levels of bunded plants [44,60]. EVI is a high-sensitivity VI designed for measuring biomass in dense vegetation as a result [61]. According to earlier research [62], the variance of teff GY was accurately described by NDVI data. When crops reach their peak growth, a strong relationship between VIs and GY is expected [5]. However, models with a high RMSE and lower R² may underestimate or overestimate yield due to pest infestations or diseases, drought, and the effects of temperature and soil nutrient stresses [63], as well as saturation of the VI values [64].

Fig. 6.

Spatial distribution maps of EVI (enhanced vegetation index) for teff (a), finger millet (b), and NDVI (normalized difference vegetation index) values for teff (a), and finger millet (b) in the study area.

Fig. 7.

Spatial distribution maps of SAVI (soil-adjusted vegetation index) values of teff (a) and finger millet (b) at the study area.

EVI and NDVI also explained the majority of the variation in GY; thus, the models performed well in predicting crop yields at the flowering stage. This could be because EVI is based on the red, blue, and near-infrared regions of the spectrum. It could also be because EVI is a high-sensitivity VI optimized for measuring terrestrial vegetation in dense vegetation [61]. As a result, a study found that EVI had a stronger linear association with LAI in field crops than NDVI [43]. Besides, LAI retrieved from Sentinel data could be used as an approach to predict yields of wheat and barley [9]. In previous studies, NDVI was used to monitor vegetation growth and crop yield [65]. However, as biomass density increases, NDVI becomes saturated [66,67]. NDVI data accurately explained the variation of sorghum yield [68,69] (maize GY [70,71] and teff GY [72] in previous researches.

According to Ref. [73], EVI and NDVI at flowering were moderately to strongly correlated (R² ranging from 0.56 to 0.89) with irrigated cotton yield. However, at LAIs ranging from 5 to 6 m² m⁻² [9] and 3 to 6 m² m⁻², NDVI became ineffective [74]. Besides, NDVI poorly indicated vegetation vigour phenology for crops having LAI >3 m² m⁻²) that have accumulated high levels of chlorophyll in their leaves [75]. However [76], found a positive relation between NDVI and SAVI with LAI and sorghum GY. In other studies, EVI and NDVI were found to be the best predictors of crop GY based on reflectance values [18,77]. This could be because they are derived from the red and near-infrared spectral bands, which are heavily influenced by chlorophyll absorption and light reflection [61,78].

4.3. Effect of soil bund on crop yield and spatial crop yield mapping

Most bunded plots have higher GY values due to improved water and nutrient contents, increased LAI, and plant density (Fig. 9). This could be due to the variations in biotic and abiotic factors [63] and their saturation [79,80]. It could also be due to the plants' increased absorption of red light and reflection of NIR light [44]. Besides [81,82], reported comparable results in wheat. The absorption of photosynthetically active radiation by crop canopies is linearly related to VIs [83].

Fig. 9.

Contribution of soil bunds on teff's (a) and finger millet's (b) grain yields (GY) estimated from Sentinel-2 data and teff's (c) and finger millet's (d) GY) predicted from spectroradiometric data using box plots.

In line with the current study's result, a strong association between NDVI and pearl millet yield and green difference vegetation index and millet GY [84,85] was observed. According to research, NDVI asymptotically saturates in high LAI [61,86]. [87] found that NDVI highly correlated (R² = 0.71) with the chlorophyll content of wheat, however, it loses sensitivity after plants accumulate a critical level of saturation level. On other hand, NDVI and LAI showed a strong association (R² = 0.82; RMSE = 0.003) with soil moisture [9,87]) and LAI was the best vegetation index for the retrieval of the soil moisture in wheat and barley croplands [9,88,89].

We used the best-fit models to predict the GY of the two crops across the Aba Gerima catchment (Fig. 8). Teff GY ranged from 0.64 to 2.16 ton/ha for bunded plots and 0.60 to 1.85 ton/ha for non-bunded plots under Sentinel-2 data (Fig. 8a and b; Fig. 9a). Teff GY ranged from 1.01 to 2.38 ton/ha for bunded plots and 1.10 to 1.48 ton/ha for non-bunded plots under spectroradiometric data (Fig. 8e and f; Fig. 9c). Besides, finger millet GY ranged from 0.91 to 2.83 ton/ha for bunded plots and 1.07 to 2.90 ton/ha for non-bunded plots using Sentinel-2 data (Fig. 8c and d; Fig. 9b). Finger millet GY ranged from 1.92 to 2.57 ton/ha for bunded plots and 1.81 to 2.38 ton/ha for non-bunded plots under spectroradiometric data (Fig. 8g and h; Fig. 9d). These differences could be attributed to spatial differences in soil fertility, water availability, crop cultivars, crop management practices, sowing date, and environmental conditions [5,29] and slope class and the presence or absence of a soil bund. Despite the impact of these factors on yield, strong linear relationships were discovered between VIs derived from Sentinel-2 imagery and crop yield [90,91].

Fig. 8.

Spatial yield distribution of Sentinel-2-based bunded teff (a), non-bunded teff (b), bunded finger millet (c), non-bunded finger millet (d), and spatial yield distribution of spectroradiometric data-based bunded teff (e), non-bunded teff (f), bunded finger millet (g), and non-bunded finger millet (h) in Aba Gerima catchment. GY grain yield.

The best-fit models were successful in estimating the yield of finger millet and teff at the flowering stage using VIs derived from Sentinel-2 data. However, caution will be required before applying the models to other regions and seasons because differences in characteristics, such as land management practices [92,93], will affect the accuracy of a yield estimation model [26,94]. As a result, before the models can be used to predict crop yield, they must be redeveloped or (if the basic model form is valid) parameterized for each region.

Our findings confirmed the ability of Sentinel-2 image- and spectroradiometric data-derived VIs developed from data to predict yield in the Aba Gerima catchment. As a result, it will provide critical support for improved agricultural management, thereby increasing food security and farmer income. Agriculture managers will also benefit from the information on yield variability because it will assist them in defining different land management zones with specific yield-limiting factors that should be monitored [95]. Moreover, the spatial information on crop yields before harvest will likely help government agencies responsible for crop market forecasts and food security at various scales.

5. Conclusions

The presence of a soil bund increased grain yields (GY) for teff and finger millet crops. As a result, effective soil bunding can aid in the conservation and restoration of soil quality and functions while also ensuring long-term production and environmental protection. We also confirmed our research hypothesis that crop growth and yield estimates could be provided with acceptable accuracy using the available vegetation indices (VIs). Such modern approach can provide critical guidance in the development of agricultural and environmental policies aimed at improving food security, farmer income, and ensuring ecological integrity.

We discovered that Sentinel-2 satellite and spectroradiometric data obtained during the flowering stage accurately predicted teff and finger millet yields in Ethiopia's Aba Gerima catchment. Hence, spectroradiometer and Sentinel-2 are more accurate approach of crop yield estimation than the conventional ones. The majority of VIs calculated from Sentinel-2 images and spectroradiometric data were strongly influenced by soil bund construction, which increased yield, and slope variation, which reduced yield on steeper slopes. The best predictability was provided by NDVI and EVI.

Our findings show that Sentinel-2 and spectroradiometric data can be used to accurately predict the crop yield of teff and finger millet, as well as its spatial variability at the catchment level. Creating a timely and accurate yield-estimation model could assist decision-makers in allocating resources and improving food security and environmental quality. It can also assist land-users including farmers to decide on what to grow and when to grow. Hence, scaling up the models should be done to collect reliable crop yield data under Aba Gerima catchment conditions. The adopted approach also offers an easy way to upscale the crops’ yields from the plot to the catchment level and map their yield variability under similar catchment conditions. The same approach should be tested to see if it can predict the yields of other crops in different regions of Ethiopia.

Although our models can be applied to other regions, they must be tested to ensure that they and their parameterization are appropriate for those regions before being used to support agricultural management in those regions. As the image quality, date of image and spectral data acquisition could affect the outputs of this research, re-calibrating and re-validating the models with increased sampling intensity and climatic variations should be part of future research work.

Author contribution statement

Gizachew Ayalew: Conceived and designed the experiments; Analyzed and interpreted the data; Wrote the paper.

Gizachew Ayalew, Derege Tsegaye, and Enyew Adgo: Performed the experiments; Wrote the paper.

Atsushi Tsunekawa, Nigussie Haregeweyn, Ayele Almaw Fenta, José Miguel Reichert, Temesgen Mulualem Aragie, and Kefyialew Tilahun: Analyzed and interpreted the data; Contributed reagents, materials, analysis tools or data; Wrote the paper.

Funding statement

Dr. Gizachew Ayalew Tiruneh was supported by Japan International Cooperation [JPMJSA1601].

Data availability statement

Data included in article/supplementary material/referenced in article.

Declaration of interest’s statement

The authors declare no conflict of interest.

Appendix A. Supplementary data

The following are the Supplementary data to this article: