Sentinel-2 NDVI as an indicator of apple orchard yield

1 Introduction

Among temperate zone fruit crops, apples are the most widely cultivated crop worldwide. According to FAOSTAT, over 80 million tons of apples are produced annually worldwide (Aghirbov et al., 2023). Apple yields typically fluctuate significantly from year to year due to biological cycles, climatic conditions, and varietal characteristics (Egorov et al., 2001; Radzhabov et al., 2021). Therefore, independent, accurate assessment and forecasting of apple yields is an economically significant task, significantly impacting the fruit market in both individual regions and countries.

Traditional statistical methods for tracking crop yields at the regional and national levels—based on collecting yield data from farmers—are time-consuming and often contain significant errors. In many countries, statistical yield assessments are also lacking, significantly affecting market prices and fruit availability. In recent years, methods for assessing crop yields using satellite data have been developed (Atzberger, 2013). The greatest progress has been achieved in predicting the yield of annual crops (Corgne et al., 2016; Cuaran and Leon, 2021). However, transferring this methods to perennial crops is largely impossible, due to the fact that fruit plants develop in one place for many years with a constant nonlinear increase in aboveground phytomass and are often cultivated in combination with annual crops (Morenets et al., 2018). Existing studies on satellite-based assessment of fruit trees status and their phenological development are usually local in nature (Savin et al., 2020; Carmona et al., 2024), or are aimed at assessing the condition of individual trees within the framework of precision farming technologies (Usha and Singh, 2013; Pujar et al., 2017). Research aimed at developing methods for forecasting fruit crop yields using satellite data is virtually nonexistent. This is due to the insufficient understanding of the relationship between the satellite imagery of fruit orchards and their actual yields.

The aim of our research was to analyze the relationship between the actual yield of apple orchards and the NDVI vegetation index calculated based on Sentinel-2 data, using an apple orchard in northern Azerbaijan as an example, as a physical basis for constructing yield forecasting models.

2 Research site and methods

The research site is a mixed-age apple orchard in northern Azerbaijan. The geographic coordinates of the center of the orchard are 41°31’10.98”N and 48°35’26.71”E. The site is characterized by hot summers and cool winters. Air temperatures do not drop below -2-4 °C, and in summer often exceed +35 °C. Precipitation is low (250–300 mm), necessitating irrigated apple cultivation. The orchard is located on a plain sloping toward the Caspian Sea, with an elevation difference of approximately 40 meters. Typical meadow-brown soils of varying textures predominate within the orchard (Torpag Farm, 1985; GUGK, 1991). The growing season was defined from March to November. In general bud development and bud burst phase takes place at end of March - early April, inflorescence emergence and pink bud occurs in the first half of April, flowering typically lasts 8 to 16 days at the end of April, depending on the specific cultivar and weather patterns. The ripening groups in research site are structured as: summer varieties (Gala), ripen and are harvested in July and August; autumn varieties (Golden Parsi, Arian): ripen around September; and winter varieties (Golden Delicious, Granny Smith, Fuji): harvested from October through November. However, the exact timing of phenophase onset varies from year to year due to meteorological conditions.

The study area is approximately 303 hectares. It consists of 53 plots with apple orchards of varying ages. Of these, 30 plots, planted between 2007 and 2017 with available apple yield data for different years, were analyzed. The plots range in area from 3 to 10 hectares (Figure 1). Cultivation is carried out under intensive artificial irrigation. The following apple varieties are grown on the plots: Arian (2 plots), Gala (8 plots), Golden (1 plot), Golden Parsi (4 plots), Granny Smith (2 plots), and Fuji (4 plots). Nine more of the studied plots cultivate two-three of the above-mentioned varieties simultaneously. The oldest plantings in the study area were established in 2007, and the youngest in 2017 (Figure 2).

For each plot, apple yield was determined in the field using the direct metering method. For each plot all apples were harvested and weighed. Then the apples weight in tons per plot was divided by the plot acreage in hectares. Such field yield observations are available for every plot for each year from 2018-2024. The total number of plots under analysis is presented in Table 1. NDVI values (Rouse et al., 1973; Rembold et al., 2013), calculated from Sentinel-2 satellite data for the same time period, were used as a satellite-based yield predictor. The spatial resolution of the data used for NDVI calculation (bands B4, and B8) is 10 meters on the ground. A total of 81 cloud-free and cloud-shadow-free Sentinel-2 images (Level-2A product, which provides orthorectified surface reflectance images after atmospheric correction) were selected and downloaded using Sentinel 2 Image Downloader plugin in QGIS for the research site from 2017 to 2024, covering all growing seasons. Presence of clouds and its shadows were detected visually. NDVI was calculated using the generally accepted formula: NDVI = (NIR - Red)/(NIR + Red), using Sentinel-2 bands B4 and B8 (https://sentiwiki.copernicus.eu/web/s2-processing) as the red (Red) and infrared (NIR) bands.

Table 1. Data sets used for the analysis (number of plots).

The seasonal absolute maximum and area-weighted average NDVI values were calculated based on analysis of all images for each growing season for each plot. A number of images for average per plot NDVI value calculation varied from 8 to 14. Absolute maximum of plot NDVI was calculated using pixel level NDVI values of all pixels within the plot during the season. Area-weighted NDVI values was calculated using two step approach. Firstly, the area-weighted NDVI was calculated for each plot for each image under analysis. After that the average plot value was calculated using area-weighted NDVI values for all images per growing season. Examples of maximum and average NDVI curves are presented as Figure 3. The relationship between crop yield and NDVI normally is considered nonlinear (Rai et al, 2022). However, yield modeling using satellite data typically faces the challenge of limiting the length of the time series that can be used to construct the model. For annual crops, this is due to the introduction of new varieties (Samborski et al., 2015), while for perennials, it is due to the annual increase in aboveground plant biomass (Zhu et al., 2020). Therefore, 10–20-year crop yield data and satellite predictors are most often used (Ali et al., 2022), which is insufficient for constructing high-quality nonlinear models. Therefore, in our study, simple regression modeling was used to estimate the relationship between NDVI and crop yield. In our case such models cannot be used to accurately predict crop yield due to limited data sets, but are quite suitable for roughly assessing the presence of relationships. To test the models, a cross-validation procedure was conducted. The sample was randomly divided into 3 identical blocks (33,3% each), one of which was used for testing, and the rest for training. To ensure the stability of the quality assessment parameters, this division was repeated 7 times. The caret package in R (https://cran.r-project.org/web/packages/caret/vignettes/caret.html) was used for cross-validation. The main parameters for assessing the quality of the models based on the cross-validation results were the following: the coefficient of determination (R²cv), the root mean squared prediction error (RMSE_cv), and the ratio of the interquartile range to the root mean squared prediction error (RPIQ).

NDVI values were calculated and spatial modeling was performed using the ILWIS v.3.3 software. Statistical data processing was performed using the Statistica v.6.0 software package.

3 Results and discussion

The following results were obtained from the studies conducted. If all yield measurements for all years, varieties, and plots were combined into a single sample set (210 observations), no reliable regression model was constructed for either the average seasonal NDVI or the seasonal maximum value per plot. The best model was characterized by R_2cv = 0,14, RMSE_cv = 183, and RPIQ = 0,3. This result is expected, as the sample set included apple tree stands of different ages and varieties, which have different phenological development patterns and different plant statuses.

Table 2 presents the modeling results for data grouped by apple variety.

Table 2. Quality indicators of regression models between yield and NDVI by apple variety.

Judging by the data in Table 2, dividing the sample set by apple variety did not result in the construction of good models, indicating the lack of a reliable relationship between the yield of individual varieties and NDVI. This is most likely due to the fact that the samples included stands of the same variety (or a mixture of varieties) but of different ages. However, it should be noted that slightly more reliable models can be constructed using the average seasonal NDVI value compared to using the seasonal NDVI maximum. This is most likely due to the fact that the value of the seasonal NDVI maximum can be influenced by spatial heterogeneity in the condition of the stands, associated with soil fertility variability across the sites (Gasimzade et al., 2025).

Table 3 presents the modeling results for grouping data by apple stand age.

Table 3. Quality indicators of regression models between yield and NDVI by apple stand age.

Grouping the data by planting age revealed that, overall, the correlation exists for middle-aged plantings (13–16 years old) and is not observed for either older or younger plantings. This is fully explained by the existence of a peak in apple fruiting (Sedov, 2020). It should also be noted that the R² values are likely underestimated due to the fact that the sample for each establishment year includes plantings of different varieties.

Table 4 presents the modeling results for each of the analyzed plots.

Table 4. Quality indicators of regression models between yield and NDVI for individual apple tree planting plots.

A good correlation between NDVI and yield was established in more than half of 30 plots, with four of these plots being considered very good (R²>0.8). However, no clear correlation was found between the varieties of trees. Even some plots with mixed varieties demonstrated a good correlation. This is likely due to factors other than cultivar composition exerting a stronger influence, despite the phenological and physiological diversity of the varieties and differences in the spectral reflectance of their leaves (Savin et al., 2023).

The correlation with tree age is also less pronounced. However, no correlation was found for the youngest and oldest trees. This is because annual fluctuations in tree yield in older trees are less determined by the “greenness” of their crowns, which becomes more constant with age. In young stands, the soil background and its spatial heterogeneity greatly influence their image on satellite data (Gasimzade et al., 2025). The relationship between crown condition and tree yield is also less pronounced than in middle-aged stands (Sdvizhkov et al., 2011).

In many cases, a closer relationship between yield in a plot and the seasonal average NDVI value is observed, rather than the seasonal maximum of the index. This is explained by the fact that the seasonal maximum of the index can only indicate the status of individual trees within a plot, while the average value is an indicator of the status of all trees in a plot.

The obtained results showed that the relationship between yield and NDVI in individual plots was generally higher than for the entire plot, grouped by both age and variety. This is entirely logical, since when analyzing a single plot, only the condition of the trees changes each season, while the influence of other factors affecting yield (soil conditions, irrigation and cultivation practices, age, and varietal differences) is minimized.

The aim of our research was to analyze the relationship between the actual yield of apple orchards and the NDVI vegetation index calculated based on satellite Sentinel-2 data. We have found that unlike similar approaches successfully used to predict yield for annual crops (Atzberger, 2013), in the case of perennial apple tree plantations, results of comparable quality cannot be obtained over large areas. This is due both to the varying ages of the plantations and, presumably, to the more contrasting phenology and physiology of the varieties. For annual crops, yield predictive models based on NDVI are more robust and their quality is higher over large areas than in individual small plots (Rembold et al., 2013). Our results showed that, in the case of perennial fruit orchards, more stable relationships were established for individual sections of the orchard than for a larger tract comprising orchards of different ages and varieties. Therefore, when developing satellite-based assessment models for forecasting apple orchard yields based on NDVI, it is more promising to construct separate models for each orchard section using average seasonal index values rather than seasonal peaks, as is often done for annual crops (Savin, 2015; Rodimtsev et al., 2022).

Attempts to forecast apple yields over large areas (country or region level) using satellite data are very rare. Much more research is focused on developing yield estimation methods based on data obtained from drones or proximal sensors (Häni et al., 2020; Chen et al., 2022). This is because the use of satellite data, which does not allow for the analysis of individual fruits or trees, does not lead to reliable yield forecasts or estimates (Apolo-Apolo et al., 2020). The results of our research once again confirmed this, showing that the most reliable results can be obtained by analyzing long-term satellite data series for individual plots with middle-aged stands. Previous results obtained for other perennial crops (Moussaid et al., 2022) confirm the strong influence of factors such as crop variety, age, and the type of remote yield indicator used on yield forecasting effectiveness.

Despite the relatively good results for forecasting and estimating apple yields obtained using the combined use of satellite and UAV data (Pan and Liu, 2026), the use of these approaches is limited by the availability of UAV data over large areas. Improving methods based on high- and medium-resolution satellite data is possible by improving the quality of detection of plantations of different ages and varieties, as well as by searching for new satellite indicators of crop yield.

4 Conclusion

If all yield measurements for all years, varieties, and plots were combined into a single sample set, it was not possible to construct a significant regression model.

Dividing the sample by apple variety also did not produce good models, indicating the lack of a reliable relationship between the yield of individual varieties and NDVI.

Grouping the data by planting age revealed that, overall, the relationship exists for middle-aged plantings (13–16 years old) and is not observed for either older or younger plantings.

The results obtained showed that the relationship between yield and NDVI in individual plots was generally higher than for the entire sample, grouped by both age and variety.

Unlike similar approaches successfully used to predict yield for annual crops, in the case of perennial apple plantings, it is likely to be that comparable results cannot be obtained for large areas due to stronger effect of age and cultivar on NDVI.

Slightly more robust models can be built using the seasonal mean NDVI compared to using the seasonal maximum NDVI.

Research results have shown that cultivar and age of apple orchards must be taken into account when constructing models for estimating apple yield based on Sentinel-2 satellite data. Consequently, forecast accuracy can be improved by improving the recognition of apple orchard cultivars and ages using satellite data. Existing Sentinel-2 satellite data archives are already sufficient to obtain consistent observation series for the status of orchard plots (not individual trees) from planting to peak fruiting. The search for new satellite indicators of apple orchard yield is also a promising direction of investigations.

Generative AI statement

The author(s) declare that no Generative AI was used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Xiao Yang, Chinese Academy of Agricultural Sciences, China

Reviewed by: Yingisani Chabalala, University of South Africa, South Africa

Emilio Ramírez-Juidías, Universidad de Sevilla Instituto Universitario de Arquitectura y Ciencias de la Construccion, Spain