1.

Introduction

Kyrgyzstan is a mountainous country with rich biodiversity. Although it occupies only 0.13% of the Earth land surface, Kyrgyzstan has about 2% of the world’s floral and 3% of its faunal diversity.¹ Also many wild fruits and flowers originate from this region with a wide range of genetic diversity related to widely cultivated plants like tulips^2,3 and apples.⁴ Moreover, Kyrgyzstan is a poor country with more than 60% of its population residing in rural areas.⁵ Wild walnut-fruit forests in the south play an important role in the rural economy, which are effectively the main sources of income for many families in the villages, surpassing animal husbandry. Walnut (Juglans regia) is the main crop, which is being harvested on annual basis by local residents; however, wild apples also contribute to the incomes of rural families. Forest products contribute from 22% to 61% of local incomes with walnut collection being the most significant source followed by wild apples.^6,7

The most numerous wild apple species is Malus sieversii; red apples (Malus niedzwetzkyana) are far less abundant ($<10\%$ of apple trees). Both Malus species are included in the Red Data Book of Kyrgyzstan.¹ Malus niedzwetzkyana is rated as “vulnerable” in national Red Data Book and “endangered” by the IUCN Global Red List; Malus sieversii is categorized as “least concern” and “vulnerable,” respectively, on the two lists. Malus sieversii is considered an ancestor of Malus orientalis and Malus domestica, a domestic variety of apple trees,⁴ which makes it a valuable genetic resource for new varieties.

Livestock that browse freely in these forests destroy young trees and damage stems and branches of adult trees.^8–10 Overgrazing and selective cutting for firewood contributes to suppression of natural regeneration and loss of genetic diversity.^11,12 In addition to these anthropogenic influences, climate change is increasing drought occurrence frequency and thus changing natural habitats.^13–16 Kyrgyzstan is the third most vulnerable country in Eastern Europe and Central Asia to climate change, mainly owing to its climate-sensitive agricultural systems and a lack of an adaptive capacity.¹⁷

If no measures are taken, the sparse populations of wild apples can be severely damaged and further decreased. Thus accurate spatial and temporal monitoring of trees is important to protect natural ecosystems. In addition, accurate estimates of the extent of degraded forests and forest clearing are key components of international climate change initiatives, such as reducing emissions from altered and transformed forests in developing countries (REDD+) within the context of the UN framework convention on climate change.¹⁸ As selective logging and collection of nontimber forest products are widespread, it is imperative to develop a national or subnational system that accurately monitors forest canopy openings, intensity of use, and contributes to improved forest management to recover forest biomass (carbon stocks) and natural regeneration.¹⁹

Remotely sensed data provide a great potential for analyzing Earth surface dynamics at various spatiotemporal scales, particularly in areas that are challenging to access.²⁰ Knowledge of spatial variability among and within forest parcels is a key factor for the users of walnut fruit forests to estimate yield and quality.²¹ In this context, remote sensing (RS) has already shown its potential and effectiveness in spatiotemporal vegetation monitoring.^19,22–24 Additionally, many satellite platforms (e.g., Landsat, MODIS, Aster, SPOT, Sentinel-1, and Sentinel-2) are now providing free datasets, thus promoting satellite imagery for many agricultural applications,^{14,22,23,25–27} including those with multisensor data fusion approachs.²⁸ For example, remotely sensed images from Sentinel-2 used in our research offer decametric resolution in terms of space and time, with a ground sample distance of up to 10 m and a revisit time of 6 days. Misregistration of Sentinel-2 imageries was addressed in the processing baseline (version 02.04)²¹ and deployed by the European Space Agency on June 15, 2016.²⁹

However, when considering vegetation over complex terrain, such as mountainous Kyrgyzstan with 94% of the land occurring above 1000 m a.s.l.,³⁰ RS becomes more challenging. Indeed, steep and complex orography leading to varying lighting conditions may deeply affect the computation of the overall spectral indices leading to a biased vegetation status assessment. Therefore, approaches and algorithms using unmanned aerial vehicles (UAV) have been developed for satellite-based multispectral vegetation pixel calibration over mountainous regions. Low-altitude platforms, such as UAV and airborne sensors, provide imagery with high spatial resolution (up to a few cm) and flexible flight scheduling³¹ facilitating vegetation monitoring at high resolution in complex terrain.

Recent studies applied UAVs in assessing forest condition,^19,32–34 precision agriculture,^21,23,24,35 and mapping, modeling, and monitoring landslides.^36–43 Several approaches have been employed to calibrate satellite imagery with UAV multispectral data and field radiometric measurements. These focus on improving the satellite multispectral imagery to facilitate use in precise monitoring of large agricultural areas by calibrating spectral bands^44,45 or calibrating vegetation indices.^35,46 Some studies have improved satellite reconnaissance data on physical characteristics of vegetation (e.g., leaf area index) using UAV data.^47,48 Other studies solve the practical issue of mapping of invasive species with combinations of UAV and Sentinel imagery and application of machine learning approaches.⁴⁹ Our understanding of forest vegetation phenology has also improved by calibrating satellite-derived vegetation and imagery index time series with images derived from UAV and spectral reflectance sensors.⁵⁰ Another study used RGB (red, green, and blue) images derived from consumer grade UAVs to more precisely identify riparian forests from Sentinel image time series.⁵¹ Other recent investigations focused on the alignment among different spectral bands of Landsat, Sentinel, PlanetScope, and UAV (MicaSense) images on various surface categories in urban areas, providing an exhaustive comparison of spectral band matching,⁵² and atmospheric correction of satellite images with UAV-sensed normalized difference vegetation index (NDVI).⁵³ All these studies show the viability and practical applicability of employing satellite imagery correction with UAV imagery because UAV imagery is more accurate due to its proximity to the observation point and there is no need for atmospheric correction.

The scope of this work is to compare Sentinel-2 and UAV imagery from Western Tian Shan for monitoring applications in walnut fruit forests with a practical aim to improve forest management and species conservation. In this research, a detailed analysis and comparison of products provided by a decametric resolution satellite and a low-altitude centimetric resolution UAV platform is presented. We developed an RS-based approach to calibrate Sentinel-2 data at a study site in Western Tian Shan using UAV data. The effectiveness of Sentinel-2 data bias correction was evaluated by considering the well-known relation between the NDVI, vegetation condition index (VCI), and standard precipitation index (SPI).

2.

Materials and Methods

2.1.

Study Site

The study was conducted in the western Tian Shan (Fig. 1) on slopes of the Fergana and Chatkal ranges. More specifically, the investigated areas are: Sary-Chelek biosphere reserve, Padysha-Ata nature reserve, and Kara-Alma forestry unit. These areas are recognized as key biodiversity areas (KBA) by the Critical Ecosystem Partnership Fund (Table 1).

Fig. 1

Study site locations: (a) within Kyrgyzstan; (b) Jalal-Abad region with croplands and area covered by UAV; (c) UAV image example from Padysha-Ata nature reserve; and (d) Pacha-Ata meteorological station showing the study polygon.

Table 1

West Tian-Shan KBA.

Name	KBA_ID	Coordinates	Area (km2)	Description
Sary-Chelek	KGZ06	Lon: E71.933132°	237.96	UNESCO biosphere reserve on south slope of Chatkal range in the south of Kyrgyzstan
		Lat: N41.868115°
Padysh-Ata	KGZ05	Lon: E71.683163 deg	680.55	Padysha-Ata nature reserve on the south slope of Chatkal range
		Lat: N41.717878 deg
Kara-Alma	KGZ18	Lon: E73.341518 deg	267.6	Kara-Alma forestry unit has the largest population of wild apple trees, on the western slope of Fergana range
		Lat: N41.249517 deg

The western slopes of Fergana range and southern slopes of Chatkal range in the south of Kyrgyzstan are covered with unique fruit and nut forests, occupying elevations ranging from about 1000 to 2500 m a.s.l. These semiwild forests are dominated by walnut (Juglans regia) trees with patches of wild apple (Malus sieversii and M. niedzwetzkyana) and wild pear (Pyrus korshinskyi), which are globally threatened. Abies semenovii and Picea schrenkiana can be found in mixed forests on hillslopes. Other scattered tree species also grow in this forest, but they do not develop forest patches and are distributed sporadically within the walnut and apple forests. These scattered species include Acer turkestanicum, Pyrus turcomanica, Crataegus spp., Betula spp., Populus spp., Fraxinus spp., Prunus sogdiana, Lonicera spp., Berberis spp., Cotoneaster spp., Rosa spp., and among others.

2.3.

Methodological Approach

2.3.1.

Automatic photogrammetric adjustments

Photogrammetric adjustments and the development of orthomosaic UAV images were conducted using Agisoft Metashape 1.8.3 Professional Edition.⁵⁵ This software uses the coordinates of exposure centers of each image performing aerial triangulation in each cell and reconstructs the photogrammetric blocks.¹⁹ It also conducts automatic lightning adjustments of scenes based on the information from the UAV lightning sensor. These adjustments ensure data consistency among flights. Finally, we obtained 59 multispectral orthomosaics, which were resampled to a 2-cm spatial resolution and $100\times 100\mathrm{m}$ tile size. Specific information regarding the camera and image overlays of the UAV image processing is presented in Table 3.

Table 3

Parameters used in the Agisoft Metashape software for image orthomosaic generation and georeferencing the UAV image of the study site.

Attribute	Value
Number of scenes per tile	$\sim 410$
Number of tiles	59
Flight altitude	50 m
Image resolution	$0.02\mathrm{m}{\mathrm{pixel}}^{-1}$
Coverage area (tile size)	100 × 100 m
Spectral channels	Red, green, blue, red edge, and NIR

NDVI was calculated in Agisoft Metashape with a raster calculator as follows:

Eq. (1)

$$\mathrm{NDVI}=\frac{(\mathrm{NIR}-\text{Red})}{(\mathrm{NIR}+\text{Red})},$$

where NIR is the near-infrared band and Red is the red band. We used only NDVI images in the analysis. The UAV image composite [Fig. 2(a)] allowed us to identify individual trees, whereas the corresponding NDVI image sections derived from Sentinel-2 [Fig. 2(b)] imagery showed much less detail.

Fig. 2

One tile of the remotely sensed images of the study site (August 2021): (a) UAV NDVI image (0.02 m spatial resolution) and (b) Sentinel-2 NDVI image (10 m spatial resolution).

2.3.2.

Data processing

In this section, specific methods for data processing developed to compare and investigate the imagery derived from the two platforms with different spatial resolutions are presented and discussed in detail.

We calculated NDVI with Sentinel-2 data according to Eq. (1) in Google Earth Engine for the dates closest to the UAV flights with $<10\%$ of the scene covered with clouds. As a result, 59 Sentinel-2 NDVI tiles with areas of about $100\times 100\mathrm{m}$ were obtained [Fig. 3(a)], which were overlapped with UAV NDVI tiles. However, because the spatial resolution of Sentinel-2 images is 10 m and that of UAV images is 0.02 m, the pixels did not always overlap—i.e., not all UAV pixels fell entirely into Sentinel-2 pixels, and sometimes Sentitel-2 tiles were larger than UAV tiles because pixel size caused a fragmentation problem (Fig. 3). Thus we omitted these fragmented pixels from further analysis. For each, tile we generated a polygon grid with each cell exactly representing pixels of Sentinel-2 images overlayed on the UAV images [Figs. 3(a) and 3(b)]. For each unfragmented cell, the mean NDVI value of each UAV image was calculated.

Fig. 3

(a) Ordered grid of pixels from Sentinel-2 and (b) polygon grid example highlighted with black.

2.3.3.

Bias correction

To assess the spatial patterns on bias distribution, we calculated NDVI difference between Sentinel and UAV mean values for each Sentinel pixel as

Eq. (2)

$${\mathrm{NDVI}}_{\mathrm{diff}(x,y)}={\mathrm{NDVI}}_{\text{Sent}(x,y)}-{\mathrm{NDVI}}_{\mathrm{UAV}(x,y)},$$

where $(x,y)$ are the pixel coordinates in Sentinel-2 and the overlayed UAV NDVI, ${\mathrm{NDVI}}_{\mathrm{diff}}$ is the difference in NDVI values between Sentinel-2 and UAV images, ${\mathrm{NDVI}}_{\text{Sent}}$ are NDVI values calculated from Sentinel-2 images, and ${\mathrm{NDVI}}_{\mathrm{UAV}}$ are NDVI values calculated from UAV images. To assess potential errors, mean error (ME)

Eq. (3)

$$\mathrm{ME}=\frac{1}{n}\sum _{i=1}^n({\mathrm{NDVI}}_{\text{Sent}}-{\mathrm{NDVI}}_{\mathrm{UAV}})$$

and the square root-mean-squared error (RMSE) were calculated:

Eq. (4)

$$\mathrm{RMSE}=\sqrt[2]{\frac{1}{n}\sum _{i=1}^n{({\mathrm{NDVI}}_{\text{Sent}}-{\mathrm{NDVI}}_{\mathrm{UAV}})}^2}.$$

To assess systematic patterns in errors for overestimation and underestimation of NDVI, we selected all the pixels with absolute values of ${\mathrm{NDVI}}_{\mathrm{diff}}$ greater then RMSE (i.e., pixels with significant error) and divided them into pixels with positive (${\mathrm{NDVI}}_{\mathrm{diff}}>0$) and negative (${\mathrm{NDVI}}_{\mathrm{diff}}<0$) errors, comprising overestimation and underestimation of NDVI by Sentinel-2. For all pixels with significant negative ${\mathrm{NDVI}}_{\mathrm{diff}}$, we have calculated spatial maximum and minimum NDVI values (${\mathrm{NDVI}}_{\max }$ and ${\mathrm{NDVI}}_{\min }$), comprising the upper and lower NDVI envelopes of underestimated values [see Eqs. (5) and (6)]. Likewise, for all the pixels with significant positive ${\mathrm{NDVI}}_{\mathrm{diff}}$, we calculated spatial maximum and minimum NDVI values (${\mathrm{NDVI}}_{\max }$ and ${\mathrm{NDVI}}_{\min }$), comprising the upper and lower NDVI envelopes of overestimated values [see Eqs. (7) and (8)]:

Eq. (5)

$$\mathrm{UNE}={\mathrm{NDVI}}_{\max (x,y)}\forall (x,y)\in (|{\mathrm{NDVI}}_{\mathrm{diff}(x,y)}|>\mathrm{RMSE})\cap ({\mathrm{NDVI}}_{\mathrm{diff}(x,y)}<0),$$

Eq. (6)

$$\mathrm{LNE}={\mathrm{NDVI}}_{\min (x,y)}\forall (x,y)\in (|{\mathrm{NDVI}}_{\mathrm{diff}(x,y)}|>\mathrm{RMSE})\cap ({\mathrm{NDVI}}_{\mathrm{diff}(x,y)}<0),$$

Eq. (7)

$$\mathrm{UPE}={\mathrm{NDVI}}_{\max (x,y)}\forall (x,y)\in (|{\mathrm{NDVI}}_{\mathrm{diff}(x,y)}|>\mathrm{RMSE})\cap ({\mathrm{NDVI}}_{\mathrm{diff}(x,y)}>0),$$

Eq. (8)

$$\mathrm{LPE}={\mathrm{NDVI}}_{\min (x,y)}\forall (x,y)\in (|{\mathrm{NDVI}}_{\mathrm{diff}(x,y)}|>\mathrm{RMSE})\cap ({\mathrm{NDVI}}_{\mathrm{diff}(x,y)}>0),$$

where $(x,y)$ are the values at pixel coordinates $x$, $y$, UNE represents the upper negative envelope, LNE is the lower negative envelope, UPE is the upper positive envelope, and LPE is the lower positive envelope.

As previously noted, we developed different bias correction rates for values of NDVI $\in$ (LNE, UNE) and NDVI $\in$ (LPE, UPE). All the NDVI values outside of these ranges are assumed not to require bias correction, as the error is less than RMSE. We used RMSE and ME for bias correction to compare which of these metrics perform better. The following equation represents this approach with RMSE; the same was conducted with ME:

Eq. (9)

$${\mathrm{NDVI}}_{\text{Sent}}^{\mathrm{corr}}=\{\begin{array}{ll}{\mathrm{NDVI}}_{\text{Sent}}+\mathrm{RMSE},& {\mathrm{NDVI}}_{\mathrm{Sent}}\in (\mathrm{LNE},\mathrm{UNE}),\\ {\mathrm{NDVI}}_{\text{Sent}}-\mathrm{RMSE},& {\mathrm{NDVI}}_{\mathrm{Sent}}\in (\mathrm{LPE},\mathrm{UPE}),\\ {\mathrm{NDVI}}_{\text{Sent}},& {\mathrm{NDVI}}_{\text{Sent}}\neg \in (\mathrm{LPE},\mathrm{UPE})\cup (\mathrm{LNE},\mathrm{UNE}),\end{array}$$

where ${\mathrm{NDVI}}_{\text{Sent}}^{\mathrm{corr}}$ is the bias-corrected Sentintel-2 NDVI.

3.

Results

3.1.

Bias Correction

After our qualitative visual analysis of ${\mathrm{NDVI}}_{\mathrm{diff}}$ [Eq. (2)] on 59 UAV images, we determined that Sentinel-2 overestimates NDVI in the areas with open terrain and grass (low NDVI values), and underestimates NDVI in the areas with trees (high NDVI values). The error analysis [Eqs. (3) and (4)] resulted in the following: $\mathrm{RMSE}=0.1228$ and $\mathrm{ME}=-0.077$. The relation between ${\mathrm{NDVI}}_{\text{Sent}}$ and ${\mathrm{NDVI}}_{\mathrm{UAV}}$ is positive [Figs. 4(a) and 4(b)]. The linear regression model yielded an $R^2=0.59$ and a $p\text{value}<0.001$ [Fig. 4(a)] on uncorrected data. The envelope interval estimates were: $\mathrm{LNE}=0.6$, $\mathrm{UNE}=0.8$, $\mathrm{LPE}=0.5$, and $\mathrm{UPE}=0.6$. After implementation of the bias correction with RMSE and ME for Sentinel-2 using Eq. (9), the accuracy of the Sentinel-2 derived NDVI increased [Fig. 4(b)].

Fig. 4

Relation between NDVI derived from Sentinel-2 and NDVI derived from UAV imagery; number of pixels = 6486: (a) Sentinel-2 without bias correction and (b) Sentinel-2 with bias correction.

Verification with VCI indicates that using RMSE for bias correction of NDVI provides better results than ME (Table 4). After correction, $R^2$ increased to 0.88 with $p\text{value}<0.001$ [Fig. 4(b)]. Thus we describe the bias correction by the following conditional equation:

Eq. (11)

$${\mathrm{VI}}_{\text{Sent}}^{\mathrm{corr}}=\{\begin{array}{ll}{\mathrm{NDVI}}_{\text{Sent}}+0.1228,& {\mathrm{NDVI}}_{\text{Sent}}\in [0.6,0.8],\\ {\mathrm{NDVI}}_{\text{Sent}}-0.1228,& {\mathrm{NDVI}}_{\text{Sent}}\in [0.5,0.6),\\ {\mathrm{NDVI}}_{\text{Sent}},& {\mathrm{NDVI}}_{\text{Sent}}\neg \in [0.5,0.6)\cup [0.6,0.8]\end{array}.$$

Table 4

Correlation coefficients between Sentinel-2 derived corrected and uncorrected NDVI and VCI with 0- and 1-month lag, and SPI indices with depth from 1 to 12 months; bold coefficients are the significant ones with p<0.05.

	SPI1	SPI2	SPI3	SPI4	SPI5	SPI6	SPI7	SPI8	SPI9	SPI10	SPI11	SPI12
VCI	0.13	0.40	0.31	0.30	0.26	0.28	0.25	0.33	0.37	0.36	0.36	0.34
${\mathrm{VCI}}_{\mathrm{Corr}\text{-}\mathrm{RMSE}}$	0.13	0.41	0.31	0.31	0.27	0.31	0.27	0.35	0.39	0.37	0.37	0.35
${\mathrm{VCI}}_{\mathrm{Corr}\text{-}\mathrm{ME}}$	0.13	0.40	0.31	0.30	0.26	0.28	0.25	0.33	0.37	0.36	0.36	0.34
${\mathrm{VCI}}_{+1\text{month}}$	0.41	0.36	0.28	0.21	0.25	0.22	0.28	0.35	0.37	0.38	0.36	0.39
${\mathrm{VCI}}_{\mathrm{Corr}\text{-}\mathrm{RMSE}+1\text{month}}$	0.42	0.36	0.29	0.22	0.28	0.24	0.29	0.37	0.37	0.40	0.37	0.42
${\mathrm{VCI}}_{\mathrm{Corr}\text{-}\mathrm{ME}+1\text{month}}$	0.41	0.36	0.28	0.21	0.25	0.22	0.28	0.35	0.37	0.38	0.36	0.39
NDVI	0.05	0.09	0.12	0.11	0.11	0.07	0.10	0.08	0.10	0.17	0.13	0.13
${\mathrm{NDVI}}_{\mathrm{Corr}\text{-}\mathrm{RMSE}}$	0.02	0.11	0.14	0.13	0.12	0.10	0.13	0.11	0.13	0.19	0.16	0.15
${\mathrm{NDVI}}_{\mathrm{Corr}\text{-}\mathrm{ME}}$	0.05	0.09	0.12	0.11	0.11	0.07	0.10	0.08	0.10	0.17	0.13	0.13
${\mathrm{NDVI}}_{+1\text{month}}$	0.22	0.29	0.17	0.17	0.14	0.16	0.17	0.22	0.25	0.25	0.21	0.19
${\mathrm{NDVI}}_{\mathrm{Corr}\text{-}\mathrm{RMSE}+1\text{month}}$	0.20	0.27	0.16	0.16	0.14	0.17	0.18	0.22	0.25	0.25	0.20	0.21
${\mathrm{NDVI}}_{\mathrm{Corr}\text{-}\mathrm{ME}+1\text{month}}$	0.22	0.29	0.17	0.17	0.14	0.16	0.17	0.22	0.25	0.25	0.21	0.19

3.3.

Verification of Bias Correction with Crop Yields

Using yield productivity data for different crops in Jalal-Abad province together with correlating with SPI indices for September, we identified that wheat and barley are the crops most related to precipitation and drought (Table 5). Thus these crops were used for verification of NDVI correction.

Table 5

Correlation coefficients between yield productivity (metric tons per hectare from 1991 to 2021) and SPI indices; SPI with depth from 1 to 12 months; Bold coefficients are significant (p<0.05).

	SPI1	SPI2	SPI3	SPI4	SPI5	SPI6	SPI7	SPI8	SPI9	SPI10	SPI11	SPI12
Grains	0.15	0.09	−0.13	−0.09	−0.08	−0.02	−0.04	−0.09	−0.08	−0.21	−0.21	−0.21
Wheat	0.25	0.32	0.20	0.14	0.16	0.26	0.14	0.09	0.05	0.22	0.22	0.22
Barley	0.26	0.33	0.11	0.24	0.29	0.35	0.27	0.22	0.19	-0.01	-0.02	-0.01
Corn	0.15	0.14	0.07	0.05	0.04	0.13	0.11	0.07	0.08	-0.08	-0.09	-0.09
Rice	0.19	0.16	−0.13	−0.09	−0.12	−0.06	−0.09	−0.14	−0.10	−0.15	−0.16	−0.16
Cotton	0.11	0.13	−0.11	−0.02	−0.04	0.03	0.02	−0.03	−0.02	−0.08	−0.08	−0.08
Tobacco	0.14	0.21	0.30	0.02	0.04	0.00	0.02	-0.02	-0.04	0.10	0.10	0.10
Vegetable oils	0.18	0.18	0.05	−0.01	−0.05	−0.01	−0.02	−0.07	−0.03	−0.06	−0.06	−0.06
Potatoes	0.22	0.16	−0.01	−0.05	−0.05	−0.01	−0.01	−0.03	0.00	−0.12	−0.12	−0.12
Vegetables	0.15	0.16	−0.11	−0.03	−0.03	0.01	0.00	−0.05	−0.01	−0.16	−0.16	−0.16
Melons	0.13	0.13	−0.05	−0.05	−0.05	−0.03	−0.03	−0.09	−0.04	−0.17	−0.17	−0.17
Fruits and berries	0.11	0.00	−0.10	−0.15	−0.13	−0.11	−0.12	−0.17	−0.13	−0.16	−0.16	−0.16
Grapes	−0.21	−0.12	−0.04	−0.09	−0.12	−0.09	−0.14	−0.14	−0.18	0.15	0.15	0.16

The highest correlation coefficients between yield and SPI are between barley/wheat and SPI6. The significant correlation is between barley and SPI6 with $r=0.35$, $p<0.05$. A qualitative visual assessment of yield related to bias corrected NDVI was only possible due to insufficient sample size (Fig. 5).

Fig. 5

Crop yield and NDVI time series for Jalal-Abad region (spatially averaged).

The same tendency of the bias corrected NDVI and yield productivity is evident (Fig. 5). The corrected annual NDVI curve is smoother than the uncorrected one and is in closer agreement with the annual yield curve of barley and wheat (Fig. 5). The maximum yields for wheat and barley in 2018 were 2.7 metric tons per hectare and 2.1 metric tons per hectare, respectively, and NDVI with bias correction was 0.38. The minimum yields for wheat and barley in 2021 were 2.2 metric tons per hectare and 1.7 metric tons per hectare, respectively, and NDVI with bias correction was 0.31.

4.

Discussion

Although most of forest, cropland, and pasture datasets are based on RS observations, the analysis of vegetation succession still relies largely on fieldwork in a space-for-time framework without the benefit of important historical remotely sensed information, with some exceptions.²⁶ Nonetheless, challenges remain in the search of an optimal and reliable relationship between remotely detected changes in vegetation indices (e.g., NDVI and VCI) and actual vegetation changes on the ground.

We found that decametric resolution satellite imagery had limitations in directly providing reliable information on the status of walnut-fruit forests in the Western Tian Shan. This effect was confirmed by the improved relation found between the corrected satellite derived NDVI and the NDVI derived from UAV surveys in contrast to the original relation [Fig. 4(a)]. Sentinel-2 data overestimates NDVI in areas with open terrain and grass, and underestimates NDVI in areas with trees. However, the accuracy of Sentinel-2 derived NDVI increased by implementing a bias correction based on RMSE for different NDVI intervals: $0.5<\mathrm{NDVI}<0.6$ with negative RMSE and $0.6<\mathrm{NDVI}<0.8$ with positive RMSE [Fig. 4(b)], and $R^2$ increased to 0.88 with a $p\text{value}<0.001$ [Fig. 4(b)]. However, using UAV together with field spectroradiometry can yield a greater spectral band coherence of $\mathrm{RSME}<1\%$.⁴⁴ High LAI estimation accuracy ($R^2$ values up to 0.88) was also obtained in other studies.⁴⁷

In addition, we found that drought index derived from RS, data are well correlated with ground-based SPI drought index. Using our proposed bias correction method, the correlation between VCI and SPI increased by 3% on average (Table 4). Moreover, shifting VCI by 1-month ahead of precipitation data yielded the highest correlation with SPI indices. This shift in vegetation response relative to precipitation has also been observed by other studies.^58–61

The highest correlation relations between crop yield and SPI drought indices were obtained for barley and wheat. A significant correlation was found between barley yield and SPI6 ($r=0.35$). Barley and wheat are mostly cultivated on unirrigated lands; thus they tend to be more correlated with drought index SPI. According to WMO recommendations, SPI6 determines agriculture drought (soil drought) which is indicative of yield response to drought. Other crops, such as corn, rice, cotton, tobacco, vegetables, potatoes, vegetable oil crops, melons, fruits, and berries, are grown in flatter, irrigated lands and thus their yields are less affected by meteorological and soil droughts. Bias corrected NDVI replicated barley and wheat yield productivity well (Fig. 5).

We assumed that the extremely high spatial resolution due to the near ground surface UAV surveys provided a “ground truth” quality layer with respect to Sentinel-2 imagery. Nonetheless, true field data collection coupled with hyperspectral imagery would still be required to further investigate the effects of complex and steep topography on Sentinel derived NDVI values. Furthermore, the spectral bands of DJI Phantom 4 Multispectral and Sentinel-2 a/b are not identical (Table 2), which contributes some uncertainty. However, our findings are a clear step forward from previous studies that used Phantom quadcopters adapted and equipped with third-party cameras,⁴⁶ notwithstanding the results of this research.

Despite the lack of consideration of certain factors discussed herein, we are confident that our work is an advance in understanding how the combination of UAV and satellite RS methods can improve the monitoring of droughts, pastures, croplands, and walnut fruit forest conditions. Because other investigations obtained similar results in terms of Sentinel-2 sensitivity to vegetation cover,^23,24 we are confident that our analysis is relevant for other mountains regions and that the bias correction method can be used to correct the Sentinel-2 images to obtain improved results in other similar contexts. Our findings can thus be used to improve management of forest resources and contribute to conservation of rare and threatened tree species.

5.

Conclusions

In this study, Sentinel-2 satellite images were compared with imagery from a UAV equipped with a multispectral camera to evaluate both techniques based on NDVI and VCI indices.

The statistical comparison between Sentinel-2 and UAV imagery shows that the followings.

Both UAV and Sentinel-2 platforms provide important information for the vegetation cover of mountains, and they are important tools for the monitoring and managing walnut-fruit forests, pastures, and croplands in Kyrgyzstan and similar mountainous regions. The choice of the most appropriate technology (UAV or satellite) depends on the use and the aim of the data collection, as they have different spatial and temporal characteristics, costs, and requirements. But using proposed bias correction method, we can significantly increase accuracy of the Sentinel-2 data.