Pasture Quality Assessment through NDVI Obtained by Remote Sensing: A Validation Study in the Mediterranean Silvo-Pastoral Ecosystem

Abstract

Monitoring the evolution of pasture availability and quality throughout the growing season is the basis of grazing management in extensive Mediterranean livestock systems. Remote sensing (RS) is an innovative tool that, among many other applications, is being developed for detailed spatial and temporal pasture quality assessment. The aim of the present study is to evaluate the potential of satellite images (Sentinel-2) to assess indicators of pasture quality (pasture moisture content, PMC, crude protein, CP and neutral detergent fiber, NDF) using the normalized difference vegetation index (NDVI). Field measurements were conducted over three years at eight representative fields of the biodiversity and variability of dryland pastures in Portugal. A total of 656 georeferenced pasture samples were collected and processed in the laboratory. The results show a significant correlation between pasture quality parameters (PMC, CP and NDF) obtained in standard laboratory methods and NDVI satellite-derived data (R² of 0.72, 0.75, and 0.50, respectively). The promising findings obtained in this large-scale validation study (three years and eight fields) encourage further research (i) to test and develop other vegetation indexes for monitoring pasture nutritive value; (ii) to extend this research to pastures of the other Mediterranean countries, building large and representative datasets and developing more robust and accurate monitoring models based on freely available Sentinel-2 images; (iii) to implement an extension program for agricultural managers to popularize the use of these technological tools as the basis of grazing and pasture management.

1. Introduction

Pastures are crucial terrestrial ecosystems [1], occupying roughly one-third of the Earth’s terrestrial area [2,3]. Besides their role in preserving biodiversity and supporting ecological processes, grasslands are also essential in ensuring global food security, serving as a key source for agricultural practices, such as livestock production [3]. In extensive Mediterranean livestock systems, pastures are the main feeding source for ruminants [4], especially from autumn to spring [5,6]. Dryland pastures in Portugal have a growth cycle that develops after the first autumn rains, lasts through winter and ends in late spring or early summer, when the temperature rises sharply and there is no rainfall, resulting in decreasing pasture nutritional value [7]. The spring is the period of greatest productivity, when the most favorable conditions are combined (temperature and precipitation) [8,9]. In addition to this marked seasonality, the Mediterranean climate is characterized by inter-annual variability, which has an impact on pasture availability and, consequently, on the animal body condition [5,10]. The critical phase, which requires supplementation with concentrated feed, usually takes place between the end of June and the beginning of October and, in years with a relatively dry autumn, can last until the beginning of winter (December–January) [8].

Monitoring the evolution of pasture availability and quality throughout the growing season is an essential pathway to increasing the efficiency of the management of extensive grazing systems, particularly in terms of stocking rates, animal rotation interval between grazing areas and the calculation of forage needs and supplementary feeding [11,12]. Traditional methods that rely on field sampling and laboratory determinations are destructive, time-consuming, and costly due to the large number of samples required to accurately represent a field [13,14]. Additionally, the delay between sampling and accessing the results makes laboratory-based methods impractical [15], preventing farmers from using them for daily assessment of pasture quality in Mediterranean grassland [16].

Evaluating spatial variability is a crucial and necessary initial step in developing strategies to manage field variability within the framework of Precision Agriculture (PA) [17]. PA is an innovative concept and one of the pillars of the Common Agricultural Policy and of various strategic national and international plans and programs. This concept involves the incorporation of proximal and remote sensing technologies in the farming system, making it possible to capture important data, such as crop vegetative vigor. Several studies have shown the benefits of proximal [7,11,18] and remote [19,20,21,22] optical sensors to obtain crop status and its evolution. However, data collection with proximal sensors has limited spatialization capabilities, requiring frequent trips to the field and travelling through areas that are representative of marked spatial variability [13]. This process is time-consuming, and the sampling pattern needs careful planning to capture the spatial variability of pasture parameters [23]. The narrow number of samples that can be processed using reference methods reduce the ability to assess the spatial variability of pasture availability [24]. From this perspective, satellite imaging systems (e.g., Sentinel-2, Landsat or MODIS) present an appealing alternative for detailed spatial and temporal vegetation monitoring, offering benefits like technical maturity, stability, open access to data and a wide field of view [13]. The participation of service providers that provide access to computerized platforms for permanent and comprehensive crop monitoring can contribute to establishing sustainable grassland management systems [25]. Remote sensing (RS) platforms are increasingly regarded as necessary tools for planners and decision makers [26]. These systems enable continuous monitoring of pasture quality, making them suitable for managing and conserving the valuable Montado ecosystem [15].

The mission “Sentinel-2”, from the Copernicus program of the European Space Agency (ESA), systematically (frequency of revisit of 5 days in Portugal, temporal resolution) acquires optical imagery at relatively high spatial resolution. One available index is the Normalized Difference Vegetation Index (NDVI), which is calculated by measuring the reflectance of the radiation emitted by plants at specific wavelengths (Equation (1)). In practice, this corresponds to the normalized difference between the near infrared (NIR; Band 8 of Sentinel-2; 842 nm) and visible red reflectance (RED; Band 4 of Sentinel-2; 665 nm), with a spatial resolution of 10 m. This index is sensitive to changes in both chlorophyll content and the intracellular spaces in the spongy mesophyll of plant leaves. The leaf’s anatomy and physiology absorb radiation around the red area of the spectrum and reflect the near infrared [27]. Several studies have shown that NDVI correlates significantly with the leaf’s chlorophyll content [1,7,21,28] and is therefore an index of great application for estimating vegetative vigor [29], with interest in studies related to the environment, agriculture, etc. [27]. Higher NDVI values indicate superlative vegetation vigor and photosynthetic capacity, while lower NDVI values signal vegetative stress, which leads to reductions in chlorophyll and changes in the internal structure of leaves due to wilting [27,30,31].

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}\mathrm{D}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{E}\mathrm{D}}}$$

(1)

Another possible application of NDVI is to estimate the nutritive value of pastures. According to Silva et al. [32], combining RS indices with field measurements provides an affordable approach to monitor pasture areas, enabling the tracking of spatiotemporal changes in pasture management indicators at various scales [32]. Pasture management with RS provides several advantages over traditional methods, including near real-time information delivery, coverage of expansive areas, frequent and repeatable measurements, and the ability to use various spectral bands [1]. Multispectral or hyperspectral imaging, in particular, has proven to be an efficient, cost-effective, and non-destructive tool for mapping pasture nutrient concentrations swiftly and across extensive areas [33,34]. Pasture quality can be evaluated through parameters such as crude protein (CP) or neutral detergent fiber (NDF) [35]. High-quality pastures usually exhibit elevated levels of CP and reduced fiber content [15,36]. This is usually observed in the initial phase of the pasture vegetative cycle (Autumn), when the plants are young, and the leaves and green color dominate, reaching maximum CP values (greater than 20% of dry matter, DM) and minimum values of fiber (NDF; approximately 40% of DM). As the vegetative cycle progresses, the proportion of leaves decreases and the proportion of stems and structural elements increases, resulting in a decrease in CP to values of approximately 5–7% of the DM in June, and the NDF tends to increase, and can reach values greater than 70% of the DM. Pasture quality, in this phase, may critically decrease to levels which would not meet the requirements of the grazing animals [22]. The pasture moisture content (PMC) can also be an indicator of pasture quality, presenting maximum values in the initial stages of the plants (approximately 80–90%) and decreasing, especially in the final stages of the vegetative cycle (minimum values of approximately 50%).

This study starts from a known problem—the information on pasture quality requires exhaustive field sample collection and laboratory analysis, which is an expensive and time-consuming process, burdening the farmer and leaving little time for decision making. The hypothesis is to assess whether NDVI, obtained from satellite images, can capture the spatial variability of pasture quality parameters in real-time and at no cost, in support of decision-making for grazing and supplementation management.

The objective of this study is to evaluate the capability of satellite images (Sentinel-2) to assess pasture quality parameters (PMC, CP and NDF) using NDVI in three years and across eight fields representative of both the biodiversity and variability of dryland pastures in Portugal.

2. Materials and Methods

2.1. Experimental Fields

This study was performed, throughout the vegetative cycles of 2018/2019, 2019/2020 and 2020/2021, at eight dryland pastures, six located in Alentejo region of Southern Portugal (AZI, GRO, MIT, MUR, PAD and TAP), one in the region of Beira Interior of Portugal (QF) and one in in the Spanish province of Extremadura (CUB) (Figure 1). The location (geographical coordinates) and characteristics (area, soil texture, pasture type, predominant trees, and animal species) of each experimental field are summarized in

Table 1. Main characteristics of the experimental fields used in this work.

Field Code	Coordinates	Area (ha)	Soil Texture	Pasture Type	Predominant Trees	Animal Species (Type of Grazing)
AZI	38°6.2′ N; 8°26.9′ W	22.3	Sandy loam	Permanent; biodiverse (predominance of composites)	Holm oak and Cork oak	Sheep (Rotational grazing)
CUB	39°10.0′ N; 6°44.6′ W	32.8	Sandy clay loam	Annual; biodiverse (mixture of grasses and legumes)	Holm oak and Cork oak	Cattle and Pigs (Rotational grazing)
GRO	37°52.3′ N; 7°56.7′ W	28.3	Sandy loam	Permanent; biodiverse (predominance of composites)	Holm oak	Cattle (Rotational grazing)
MIT	38°32.23′ N; 8°00.05′ W;	19.3	Sandy loam	Permanent; biodiverse (mixture of grasses and legumes)	Holm oak	Cattle (Rotational grazing)
MUR	38°23.4′ N; 7°52.5′ W	29.6	Loam	Annual; biodiverse (mixture of grasses and legumes)	Holm oak and Cork oak	Sheep (Permanent grazing)
PAD	38°36.4′ N; 8°8.7′ W	32.2	Sandy loam	Permanent; biodiverse (predominance of composites)	Holm oak	Cattle (Permanent grazing)
QF	40°16.38′ N; 7°25.14′ W	25.3	Loamy sand	Permanent; biodiverse (mixture of grasses and legumes)	Oaks and Eucalyptus	Horses, Sheep and Cattle (Permanent grazing)
TAP	39°9.5′ N; 7°31.9′ W	27.1	Sandy clay loam	Permanent; biodiverse (mixture of legumes)	Holm oak and Cork oak	Cattle and Pigs (Rotational grazing)

. The location of these fields is representative of the temperate climatic conditions of Portugal (“Csa: hot summer Mediterranean climate” according to the Köppen–Geiger climate classification), with a precipitation gradient: smaller amounts of rainfall in the southern districts (“Beja” and “Évora”; mean annual rainfall in the period 1981–2010 of approximately 600 mm), and greater in the northern districts (“Portalegre” and “Castelo Branco”; mean annual rainfall in the period 1981–2010 of approximately 800 mm) (source: Portuguese Institute of Sea and Atmosphere).

In each experimental field, eight ‘10 m × 10 m’ areas (spatial resolution of NIR and RED Sentinel-2 bands) were georeferenced with Global Navigation Satellite Systems (GNSSs) and materialized with stakes, in areas without trees (Figure 2 and Figure 3).

The data used in this study were collected from soil and pasture field samples and satellite remote sensing images. Figure 4 shows a flowchart of the methodology used in this study.

2.2. Soil Sampling

At the beginning, the eight soil composite samples were collected in each experimental field, at 0.30 m depth. Each composite sample consisted of five subsamples. Soil samples were dried and sieved before being processed for texture determination using a sedimentographer (Sedigraph 5100, manufactured by Micromeritics, Norcross, GA, USA). Fine components of the soil (diameter less than 2 mm) underwent standard reference analysis as follows [37]: pH (1:2.5 soil-to-water suspension) was determined using the potentiometric method; organic matter (OM) was measured by combustion and CO₂detection via an infrared cell; phosphorous (P₂O₅) was extracted using the Egner–Riehm method and measured using the colorimetric method; the cation exchange capacity (CEC) was determined using ammonium acetate.

2.3. Pasture Sampling

Three pasture vegetation cycles were monitored, with 3 pasture collection campaigns in the first year and 4 pasture collection campaigns in the second and third years. The dates of pasture sample collection were established based on the different stages of development throughout each vegetative cycle, specifically for each experimental field. The campaigns began when the grassland was already showing significant vegetation (average NDVI > 0.6), usually in early winter (late December or early January) and lasted until late spring. Due to limitations and constraints imposed by COVID-19, it was not possible to carry out all the harvests in two of the experimental fields: “QF” and “CUB”. Throughout the monitoring period, eight pasture composite samples were collected in each experimental field. Each composite sample is a result of five pasture subsamples. A metal quadrat (0.5 m × 0.5 m) and electric shears were used to collect the samples.

Table 2. Number of samples taken in each experimental field, in each vegetative cycle of the pasture and in all three vegetative cycles monitored.

Date	AZI	CUB	GRO	MIT	MUR	PAD	QF	TAP	Total
January	8	8	8	-	-	-	-	-	24
February	-	-	-	8	8	8	8	8	40
March	-	8	-	8	8	8	-	-	32
April	16	-	8	-	-	-	8	8	40
May		8	8	8	8	8	8	8	56
Total (W–S: 2019)	24	24	24	24	24	24	24	24	192
January	8	8	8	8	8	8	8	8	64
March	8	8	8	8	8	8	-	8	56
April	8	-	8	8	8	8	8	8	56
May	8	-	8	8	8	8	-	-	40
June	-	-	-	-	-	-	8	8	16
Total (W–S: 2020)	32	16	32	32	32	32	24	32	232
December	8	8	8	8	8	8	8	8	64
February	8	-	8	8	8	8	8	-	48
March	-	-	-	8	-	-	-	8	16
April	8	-	8	-	8	8	-	8	40
May	8	8	8	8	8	8	8	8	64
Total (W–S: 2021)	32	16	32	32	32	32	24	32	232
Total (2019–2021)	88	56	88	88	88	88	72	88	656

W—Winter; S—Spring.

shows the number of samples taken in each experimental field, in each vegetative cycle of the pasture and in all three vegetative cycles monitored.

Pasture composite samples were submitted to laboratory analysis to determine the reference values of pasture moisture content (PMC, in %), crude protein (CP, in %) and neutral detergent fiber (NDF, in %), all on a dry weight basis, using standard analytical methods [37].

2.4. Satellite Remote Sensing Data

Time-series of Sentinel-2 optical images (freely available from the European Space Agency, ESA), downloaded from Copernicus data hub from January 2019 to May 2021, were used. Sentinel-2 product level 2A (atmospherically corrected) images were used with no subsequent treatment. In each farm, the surface reflectance data (Band 8 and Band 4) were extracted in eight “10 m × 10 m” pixels without trees.

All images with the presence of clouds were not considered in the analysis. The use of filters in the selection of images to avoid dates with clouds was carried out using the cloud mask provided by product level 2A of Sentinel-2, complemented by visual analysis. When satellite images showed clouds (even thin) in the date of interest, over all or part of each experimental field, the image was discarded and the satellite image of the date closest to the ground sampling date (up to 5 days apart) was used. The dates when the pasture samples were collected, the dates when the satellite images were extracted and the difference in days between them are shown in the

Table 3. Dates of pasture field sampling, dates of satellite images extraction and difference in days between them (gap days), for each experimental field.

Event	Sampling	AZI	CUB	GRO	MIT	MUR	PAD	QF	TAP
	Pasture	24 01 2019	09 01 2019	24 01 2019	13 02 2019	12 02 2019	15 02 2019	21 02 2019	20 02 2019
1	RS	25 01 2019	10 01 2019	25 01 2019	14 02 2019	14 02 2019	24 02 2019	24 02 2019	14 02 2019
	Gap (days)	1	1	1	1	2	9	3	6
	Pasture	02 04 2019	19 03 2019	15 04 2019	29 03 2019	29 03 2019	29 03 2019	23 04 2019	12 04 2019
2	RS	16 03 2019	16 03 2019	05 04 2019	31 03 2019	21 03 2019	26 03 2019	20 04 2019	20 04 2019
	Gap (days)	17	3	10	2	8	3	3	8
	Pasture	30 04 2019	14 05 2019	09 05 2019	03 05 2019	06 05 2019	06 05 2019	22 05 2019	20 05 2019
3	RS	05 05 2019	12 05 2019	15 05 2019	05 05 2019	30 04 2019	05 05 2019	15 05 2019	25 05 2019
	Gap (days)	5	2	6	2	6	1	7	5
	Pasture	21 01 2020	29 01 2020	21 01 2020	20 01 2020	22 01 2020	20 01 2020	30 01 2020	22 01 2020
4	RS	20 01 2020	04 02 2020	20 01 2020	20 01 2020	20 01 2020	20 01 2020	04 02 2020	20 01 2020
	(gap days)	1	6	1	0	2	0	5	2
	Pasture	02 03 2020	10 03 2020	02 03 2020	03 03 2020	09 03 2020	09 03 2020	(*)	10 03 2020
5	RS	10 03 2020	10 03 2020	10 03 2020	10 03 2020	10 03 2020	10 03 2020		10 03 2020
	Gap (days)	8	0	8	7	1	1		0
	Pasture	21 04 2020	(*)	21 04 2020	14 04 2020	20 04 2020	20 04 2020	23 04 2020	24 04 2020
6	RS	19 04 2020		19 04 2020	19 04 2020	24 04 2020	19 04 2020	24 04 2020	24 04 2020
	Gap (days)	2		2	5	4	1	1	0
	Pasture	28 05 2020	(*)	28 05 2020	26 05 2020	29 05 2020	29 05 2020	09 06 2020	01 06 2020
7	RS	29 05 2020		29 05 2020	24 05 2020	29 05 2020	29 05 2020	08 06 2020	29 05 2020
	Gap (days)	1		1	2	0	0	1	3
	Pasture	02 12 2020	09 12 2020	02 12 2020	11 12 2020	14 12 2020	14 12 2020	15 12 2020	09 12 2020
8	RS	05 12 2020	05 12 2020	05 12 2020	05 12 2020	15 12 2020	15 12 2020	20 12 2020	15 12 2020
	Gap (days)	3	4	3	6	1	1	5	6
	Pasture	18 02 2021	(*)	18 02 2021	03 02 2021	22 02 2021	22 02 2021	25 02 2021	05 03 2021
9	RS	28 02 2021		23 02 2021	13 02 2021	19 02 2021	23 02 2021	28 02 2021	10 03 2021
	Gap (days)	10		5	10	3	1	3	5
	Pasture	08 04 2021	(*)	08 04 2021	30 03 2021	09 04 2021	09 04 2021	(*)	13 04 2021
10	RS	09 04 2021		04 04 2021	04 04 2021	09 04 2021	09 04 2021		09 04 2021
	Gap (days)	1		4	5	0	0		4
	Pasture	13 05 2021	06 05 2021	13 05 2021	11 05 2021	13 05 2021	14 05 2021	18 05 2021	17 05 2021
11	RS	14 05 2021	04 05 2021	14 05 2021	14 05 2021	14 05 2021	14 05 2021	19 05 2021	14 05 2021
	Gap (days)	1	2	1	3	1	0	1	3

RS—Remote sensing; (*) Dates not sampled at the experimental farms “CUB” and “QF”.

.

2.5. Statistical Analysis

The IBM SPSS Statistics package for Windows (version 28.0, IBM Corp., Armonk, NY, USA) was used to perform statistical analyses. The mean, standard deviation (SD) and range of variation of each parameter (NDVI, PMC, CP and NDF) were determined for each pasture collection event, considering the eight sampling points of each experimental field. This information regarding descriptive statistics was grouped by pasture vegetative cycle (2018/2019; 2019/2020 and 2020/2021).

For each experimental field, the data obtained from the three pasture collection campaigns were subjected to linear regression analysis between NDVI and each of the pasture quality parameters (PMC, CP and NDF). The linear correlation coefficient (r) was used to assess the relationship between these variables and to evaluate the performance of the model. Chang et al. [38] indicated three levels of degree of correlation: |r| < 0.4 = low; 0.4 ≤ |r| < 0.7 = medium; 0.7 ≤ |r| < 1 = high. To avoid significant discrepancies between pasture ground sampling and cloud-free satellite imagery, a threshold of 5 days of difference was applied (if there were more than 5 days, then the observation would be discarded).

With the objective of evaluating whether three sampling years and eight experimental fields were significantly different, with a 95% significance level (p < 0.05), since the normality in the data cannot be assumed, the Kruskal–Wallis nonparametric test and the Dunn post hoc test were used. Thus, the differences in the mean values of pasture parameters were considered.

With the aim of testing the viability of a general model for predicting pasture quality parameters based on NDVI values obtained freely from Sentinel-2 images, regression analysis was carried out between averaged reflectance (NDVI) values and averaged values of pasture quality parameters (PMC, CP and NDF) obtained in each collection event and in each experimental field, over the three years of study. The coefficient of determination (R²) was used as a measure of how well the data fit the regression model [15]. Askari et al. [39], to assess predictive models built with Sentinel-2 data, classified as “excellent” if R² ≥ 0.8, “good” if R² ≥ 0.7, “moderate” R² ≥ 0.60 and “poor” accuracy if R² < 0.6.

3. Results

The initial analysis of soil parameters (

Table 4. Mean ± standard deviation of soil parameters (0–0.30 m depth) in each of the eight experimental fields.

Parameter	AZI	CUB	GRO	MIT	MUR	PAD	QF	TAP
Clay (%)	9.2 ± 3.0	23.5 ± 1.6	16.8 ± 7.2	6.8 ± 2.4	8.5 ± 4.7	6.6 ± 2.0	5.4 ± 2.2	7.0 ± 6.4
Silt (%)	17.0 ± 3.8	39.0 ± 0.6	25.5 ± 3.9	9.6 ± 2.5	15.9 ± 10.7	15.4 ± 2.2	13.7 ± 3.2	14.8 ± 9.9
Sand (%)	73.8 ± 5.6	35.5 ± 1.9	57.6 ± 8.9	83.7 ± 3.7	75.6 ± 14.6	78.0 ± 2.6	80.9 ± 2.0	78.2 ± 9.0
pH	6.7 ± 0.2	5.5 ± 0.3	5.8 ± 0.3	5.5 ± 0.3	6.0 ± 0.5	6.4 ± 0.5	5.6 ± 0.1	6.0 ± 0.3
OM (%)	1.9 ± 0.2	3.1 ± 0.2	2.5 ± 0.9	1.9 ± 0.4	2.7 ± 0.5	2.7 ± 0.2	2.5 ± 0.6	2.2 ± 0.8
P2O5 (mg kg−1)	8.5 ± 3.8	11.5 ± 2.9	24.3 ± 21.5	26.5 ± 13.0	29.2 ± 21.7	23.7 ± 6.7	34.2 ± 2.5	7.5 ± 3.2
CEC(cmol kg−1)	11.3 ± 3.9	15.2 ± 2.4	11.2 ± 1.8	9.4 ± 5.8	8.6 ± 2.8	15.5 ± 1.3	6.9 ± 0.6	7.2 ± 2.5

OM—Organic matter; CEC—Cationic exchange capacity.

) shows that these pastures integrated in the Montado ecosystem are generally installed on coarse-textured soils (six of the experimental fields have clay contents < 10% and sand contents > 70%; only two experimental fields have medium texture: “CUB” and “GRO”). The pH of these soils is generally acidic or slightly acidic and the average organic matter content (OM; 1.9–3.1%) is higher than the average content in other soils in the region with other production systems, particularly cereal production [8]. This is particularly relevant because the soil samples were taken in areas outside the effect of the tree canopy (Sentinel-2 pixels), where soil OM content tend to be significantly higher [6]. Table 4 also shows, on the one hand, very low phosphorus (P₂O₅) values with concentrations between 7.5 and 34.2 mg/kg and, on the other, relatively low CEC, except in two fields (“CUB” and “PAD”).

The descriptive statistical analysis of the pasture parameters (

Table 5. Descriptive statistics (mean ± standard deviation and range) of the pasture parameters in each of the eight experimental fields throughout the three vegetative cycles monitored.

Field	Sampling Year	W–S 2019		W–S 2020		W–S 2021
	Parameter	Mean ± SD	Range	Mean ± SD	Range	Mean ± SD	Range
AZI	DM (kg ha−1)	1678 ± 488	800–3200	1300 ± 234	300–2267	1394 ± 329	600–2983
	PMC (%)	65.4 ± 6.8	41.0–78.7	72.3 ± 4.3	46.9–86.1	76.0 ± 3.3	60.0–87.3
	CP (%)	9.9 ± 1.3	6.0–15.4	11.8 ± 1.7	5.2–17.7	10.8 ± 1.5	6.0–17.2
	NDF (%)	56.4 ± 5.3	45.8–66.7	57.7 ± 3.7	47.3–67.2	57.6 ± 4.0	50.2–69.0
	NDVI	0.56 ± 0.06	0.34–0.68	0.57 ± 0.07	0.35–0.77	0.56 ± 0.04	0.38–0.70
CUB	DM (kg ha−1)	3372 ± 1197	400–8000	1269 ± 375	733–2300	3311 ± 782	867–8193
	PMC (%)	75.5 ± 3.3	60.2–86.5	84.0 ± 1.7	78.0–88.4	76.2 ± 4.4	48.8–88.8
	CP (%)	14.5 ± 3.0	6.5–28.3	18.6 ± 2.2	15.4–25.3	13.8 ± 1.5	5.2–23.3
	NDF (%)	44.0 ± 3.9	26.6–69.1	41.9 ± 3.3	35.0–49.6	56.6 ± 3.7	39.6–68.5
	NDVI	0.66 ± 0.05	0.43–0.84	0.73 ± 0.05	0.59–0.78	0.65 ± 0.07	0.49–0.89
GRO	DM (kg ha−1)	2201 ± 753	1100–5000	1710 ± 386	367–3880	936 ± 338	327–2740
	PMC (%)	59.7 ± 8.7	38.5–80.0	70.0 ± 4.5	41.9–84.1	69.6 ± 4.7	37.7–81.1
	CP (%)	11.2 ± 2.0	7.0–16.0	11.6 ± 0.9	5.8–17.5	13.2 ± 1.6	6.4–21.7
	NDF (%)	58.9 ± 5.2	46.5–71.8	56.5 ± 3.5	40.3–70.6	52.0 ± 3.8	37.3–67.6
	NDVI	0.51 ± 0.07	0.31–0.79	0.59 ± 0.06	0.28–0.83	0.54 ± 0.06	0.24–0.78
MIT	DM (kg ha−1)	2683 ± 1018	1100–5800	2963 ± 1100	667–10,767	1685 ± 632	400–5420
	PMC (%)	80.5 ± 4.5	57.7–86.8	80.4 ± 4.3	58.8–90.1	84.2 ± 3.0	74.1–92.4
	CP (%)	14.7 ± 3.2	7.2–24.6	15.6 ± 2.3	6.7–21.4	16.1 ± 3.0	5.9–27.2
	NDF (%)	43.1 ± 5.9	30.7–57.6	48.5 ± 6.1	32.9–67.4	46.0 ± 5.5	31.2–61.2
	NDVI	0.67 ± 0.04	0.56–0.76	0.68 ± 0.08	0.33–0.83	0.69 ± 0.04	0.50–0.83
MUR	DM (kg ha−1)	4104 ± 1706	1100–13,000	1305 ± 369	333–4200	1444 ± 289	367–3243
	PMC (%)	76.0 ± 5.1	63.0–84.4	78.7 ± 3.0	68.3–85.8	83.9 ± 2.2	66.2–94.4
	CP (%)	11.6 ± 2.5	5.6–19.2	13.0 ± 3.1	7.0–25.8	14.8 ± 1.8	6.6–23.9
	NDF (%)	48.1 ± 5.2	32.0–66.0	57.8 ± 3.5	45.0–67.2	51.4 ± 4.1	36.3–63.2
	NDVI	0.72 ± 0.06	0.57–0.88	0.60 ± 0.04	0.35–0.75	0.65 ± 0.06	0.43–0.82
PAD	DM (kg ha−1)	2321 ± 549	1400–3800	1462 ± 242	667–3367	946 ± 220	320–1667
	PMC (%)	75.0 ± 3.8	63.9–82.5	77.6 ± 2.6	63.8–89.1	78.2 ± 4.0	56.4–91.7
	CP (%)	13.5 ± 3.1	8.4–22.4	15.4 ± 2.0	7.2–21.9	16.1 ± 2.3	10.6–25.7
	NDF (%)	47.4 ± 5.7	35.2–60.4	50.9 ± 2.5	40.4–62.7	54.4 ± 4.4	46.7–63.7
	NDVI	0.68 ± 0.03	0.59–0.73	0.67 ± 0.03	0.43–0.82	0.66 ± 0.04	0.54–0.77
QF	DM (kg ha−1)	1179 ± 437	300–3100	1084 ± 475	410–3003	668 ± 368	103–2353
	PMC (%)	73.1 ± 6.0	57.9–84.5	76.4 ± 2.7	57.9–87.7	80.1 ± 2.4	69.1–88.1
	CP (%)	11.9 ± 3.1	6.8–22.0	15.4 ± 2.1	6.8–26.1	18.1 ± 3.0	8.0–25.6
	NDF (%)	51.0 ± 7.7	32.4–67.3	50.5 ± 3.9	40.4–61.8	48.9 ± 6.1	32.1–62.6
	NDVI	0.58 ± 0.07	0.37–0.70	0.57 ± 0.11	0.32–0.90	0.68 ± 0.06	0.54–0.76
TAP	DM (kg ha−1)	2396 ± 722	600–4700	1833 ± 454	567–5133	1238 ± 325	483–3397
	PMC (%)	74.9 ± 4.3	64.7–83.5	75.0 ± 5.2	55.7–83.1	79.6 ± 2.9	70.0–87.2
	CP (%)	9.6 ± 2.3	5.9–16.9	10.2 ± 2.1	5.7–17.8	12.6 ± 2.5	7.3–19.5
	NDF (%)	50.6 ± 6.0	35.1–64.0	55.5 ± 6.2	41.4–71.1	56.3 ± 4.1	42.3–70.3
	NDVI	0.59 ± 0.07	0.37–0.78	0.59 ± 0.05	0.41–0.70	0.65 ± 0.06	0.49–0.76

W–S—Winter–Spring; SD—Standard Deviation; DM—Dry matter; PMC—Pasture moisture content; CP—Crude protein; NDF—Neutral detergent fiber; NDVI—Normalized difference vegetation index.

) shows significant spatial variability (high coefficient of variation) over the three years of the study, in all the experimental fields and in all pasture parameters considered (productivity and quality). This spatial variability is also reflected in the NDVI, which shows reasonable ranges of variation, where the maximum values are twice the minimum values in each experimental field and in each vegetation cycle (see, for example, in the ‘AZI’ field, variation between 0.339 and 0.682 in 2019, between 0.351 and 0.773 in 2020, or between 0.381 and 0.695 in 2021), and therefore has the potential to capture the spatial variability of the pasture’s nutritional value.

Table 5 also shows significant inter-annual variability in all the experimental fields. For example, in terms of productivity (DM, in kg ha⁻¹), year 2019 generally showed higher mean values than the rest, with a very significant drop in the year 2021 in practically all the experimental fields (e.g., from 2201 to 936 kg ha⁻¹ in “GRO”; from 4104 to 1444 kg ha⁻¹ in “MUR”; from 2321 to 946 kg ha⁻¹ in “PAD”; from 1179 to 668 kg ha⁻¹ in “QF”; or from 2396 to 1238 kg ha⁻¹ in “TAP”). This inter-annual variability is demonstrated in terms of significance in

Table 6. Inferential analysis of pasture parameters: comparison between sampling years.

Sampling Year	DM (Kg ha−1)	PMC (%)	CP (%)	NDF (%)	NDVI
W–S 2019	2624 b	72.0 a	12.0 a	50.1 a	0.606 a
W–S 2020	1440 a	78.4 b	14.7 b	52.7 b	0.599 a
W–S 2021	1629 a	76.6 b	12.8 a	52.8 b	0.609 a

W–S—Winter–Spring; DM—Dry matter; PMC—Pasture moisture content; CP—Crude protein; NDF—Neutral detergent fiber; NDVI—Normalized difference vegetation index. Means in the same column followed by different letters indicate significant differences according to Dunnett’s test (p < 0.05).

, for all eight experimental fields. For example, productivity was significantly higher in the 2018/2019 growing cycle than in the others (2019/2020 and 2020/2021), while for PMC and fiber (NDF) this behaviour was reversed.

In addition to the coefficient of variation, the range of variation of pasture parameters in each experimental field and in each growing season also reflects this important spatial variability. For example, in “MIT”, PMC range between 58 and 87% in 2019, between 59 and 90% in 2020 and between 74 and 92% in 2021; on the other hand, CP ranges between 7 and 25% in 2019, between 7 and 21% in 2020 and between 6 and 27% in 2021; while NDF ranges between 31 and 58% in 2019, between 33 and 67% in 2020 and between 31 and 61% in 2021 (Table 5). The inferential analysis for all the pasture evaluations carried out in each pasture (

Table 7. Inferential analysis of pasture parameters: comparison between experimental fields.

Field	DM (Kg ha−1)	PMC (%)	CP (%)	NDF (%)	NDVI
AZI	1461 bc	71.0 a	10.9 a	57.2 d	0.535 a
CUB	2863 e	77.6 b	15.3 c	46.2 ab	0.672 de
GRO	1639 bcd	66.2 a	12.0 ab	55.9 d	0.509 a
MIT	2308 de	82.2 c	15.4 c	46.1 a	0.680 e
MUR	2502 de	79.2 b	12.7 bc	51.9 bc	0.622 c
PAD	1508 bc	77.1 b	15.2 c	51.2 bc	0.668 de
QF	977 a	76.5 b	15.1 c	50.2 b	0.554 ab
TAP	1770 cd	76.7 b	11.0 ab	54.5 cd	0.600 bc

DM—Dry matter; PMC—Pasture moisture content; CP—Crude protein; NDF—Neutral detergent fiber; NDVI—Normalized difference vegetation index. Means in the same column followed by different letters indicate significant differences according to Dunnett’s test (p < 0.05).

) confirms the significant variability of all the parameters (productivity, quality and NDVI) between experimental fields. This variability reflects the complexity of determining factors in each field, mainly in terms of soil, grazing management and interaction with the climate.

The seasonality of pasture availability, particularly its quality, reflects the effect of temperature and rainfall, characteristic of the Mediterranean region [8]. This pattern of seasonality is shown in Figure 5, taking as an example the average values of pasture quality parameters (PMC, CP and NDF) in “MIT”. This experimental field was chosen as an example because it has a meteorological station. Figure 6 shows the average temperature and monthly rainfall accumulated at this meteorological station during the experimental period. The sharp drop in PMC and CP and the significant increase in fiber at the end of the cycle reflect the significant rise in temperature and the absence of rainfall at the end of spring (May/June). The inter-annual variability of accumulated precipitation is very clear, with a value well below the average (314 mm) in 2018/2019, practically double in 2019/2020 (627 mm), corresponding to the average values in this region and a new increase in 2020/2021 (778 mm), proving to be a relatively rainy year.

Based on the valid data, i.e., 520 data pairs, which correspond to 79% of the total data collected (after discarding the cases with a difference between ground measurements and satellite images of more than 5 days; Table 3), the ability to capture the spatial variability of pasture quality from satellite NDVI measurements can be proven by the significant correlation obtained in all experimental fields (

Table 8. Linear correlation coefficients (r) between NDVI and pasture quality parameters.

Field	n *	NDVI vs. PMC	NDVI vs. CP	NDVI vs. NDF
AZI	64	0.34 *	0.69 **	−0.33 *
CUB	48	0.80 **	0.73 **	−0.76 **
GRO	64	0.68 **	0.75 **	−0.57 **
MIT	64	0.76 **	0.78 **	−0.61 **
MUR	72	0.33 *	0.45 **	−0.64 **
PAD	80	0.73 **	0.66 **	−0.47 **
QF	64	0.75 **	0.61 **	−0.29 *
TAP	64	0.65 **	0.67 **	−0.50 **
All	520	0.68 **	0.66 **	−0.58 **

n *—Number of samples considered; NDVI—Normalized difference vegetation index; PMC—Pasture moisture content; CP—Crude protein; NDF—Neutral detergent fiber; **—Statistically significant at 99% confidence level (p < 0.01); *—Statistically significant at 95% confidence level (p < 0.05); Yellow—Moderate relationship; Green—High relationship.

): a positive correlation in the case of CP and PMC and a negative correlation with fiber, stronger in the first two, with high correlation coefficients in seven of the eight fields in the case of CP (average value of 0.66 and range between 0.45 and 0.78) and in six of the eight fields in PMC (average value of 0.68 and range between 0.33 and 0.80 in PMC) and, in general, moderate in fiber, in all eight fields (average value of −0.58 and range between −0.29 and −0.76).

The inclusion in the analysis of average values obtained in each experimental field, each collection event and in the three years of study, to evaluate the possibility of developing a general estimation model based on the NDVI obtained by RS, is presented in Figure 7, Figure 8 and Figure 9, respectively, for PMC, CP and fiber (NDF). The best estimation model is given, in the case of PMC, by a linear function (R² of 0.72; Figure 7) and by polynomial models for CP (R² of 0.75; Figure 8) and for fiber (R² of 0.50; Figure 9), which suggests a saturation effect for CP > 15% or NDF < 40%, therefore, in the initial phases of the vegetative cycle, at the beginning of winter (December–January).

4. Discussion

4.1. Soil and Pasture Variability

In Portugal, extensive livestock production systems are part of the mixed Montado ecosystem, where soils with low productive potential predominate, with the best soils being used for irrigated crops. The most common limitations of soils used in the Montado are their coarse texture, acid pH, toxicity of some micronutrients, low phosphorus content, low CEC and reduced arable depth [8]. The significant spatial variability of soil parameters (particularly OM, P₂O₅ and CEC), as indicated by the high coefficient of variation (CV) observed in this study, across both intra- and inter-experimental fields, was also observed in other works conducted within this ecosystem [40,41,42,43,44]. Schellberg et al. [45] noted that this variability is a hallmark of the ecosystem, exacerbated by the presence of trees and by the dynamics of animal grazing. Uneven waste distribution by grazing animals (dung and urine) contributes to the increase in soil property variability in extensive livestock systems [46,47,48], in addition to the significant effect of trees [41,43,49].

On the other hand, soil fertility is a key factor that influences the productivity and quality of dryland pastures [5]. In our study, both productivity and quality show a high degree of variability. In terms of productivity, according to Huyghe et al. [50], average annual dry matter yields can range from 500–1000 kg ha⁻¹ in semi-natural grasslands, which prevail in marginal soils, to 6000–7000 kg ha⁻¹ in agriculturally improved grasslands [50], as in the case of the pastures monitored in this study.

In terms of pasture quality, the values obtained in our study also fall within the usual range presented by other authors under similar conditions (e.g., Fernández-Habas et al. [15], in the Dehesa in Spain). In Mediterranean ecosystems, the seasonality of rainfall (including the duration of the growing season and the timing of the precipitation) and air temperature are critical factors that affect vegetation variability during the growth period [35]. These factors can, in turn, influence the quality of available pasture biomass (such as CP and NDF contents) on spatial and temporal scales [22,35]. Lugassi et al. [35] reported that the spatial and temporal patterns of CP and NDF contents are strongly dependent on the season, shaped by climatic and topographical conditions. These influences could help explain why the availability of forage initially increases until the end of the growing season, after which it declines due to the drying out of the plants [22].

Another important factor in pasture quality variability is the diversity of the botanical species [15], characteristic of biodiverse pastures, such as those in this study. For example, a study of Serrano et al. [51], carried out on May/June 2021 at the “MIT” field, showed the existence of sixty-two botanical species, representing seventeen families. Pastures with diverse functional groups and varying phenological stages may result in significant spatial variability [15].

The extensive data collected in this study from eight farms, sampled at various times throughout the growing season, enabled us to accurately capture the full extent of pasture heterogeneity in the Montado ecosystem. The inclusion of three years of data also sought to capture the temporal variability (seasonal and inter-annual) of pasture quality. For example, the range of variation of CP was 21 points (between 6 and 27%) and 36 points for the NDF (between 31 and 67%), which are similar to values obtained by Fernández-Habas et al. [15] in a study conducted on pastures in the same ecosystem, in Spain (with a range of 24 points for CP and 46.5 points for NDF). These are crucial components of pasture, particularly CP, as it can fluctuate significantly during the grazing season, and may fall below the critical threshold of 7%, especially during the dry summer months [52]. Consequently, rapid, and effective monitoring of CP is highly beneficial for developing a successful supplementation program [22]. The inverse relationship between CP and NDF contents observed throughout the phenological cycles in our study was also reported by Lugassi et al. [35]. They found that the decrease in CP and increase in NDF as the vegetative cycle progresses, especially in its final phase, correlates with a decline in NDVI. As pasture matures, chlorophyll content decreases, leading to a reduction in NDVI [6].

4.2. Correlation between Pasture Quality Parameters and NDVI

The objective of this study is to assess the potential of using satellite images (Sentinel-2) to monitor pasture quality indicators through NDVI. However, pasturelands typically exhibit high spatial variability (as demonstrated by our study), resulting in a complex relationship between field and RS data, and making pasture monitoring at different scales challenging [32]. Grassland systems are highly variable in composition, structure, and age and they continually change in response to various factors, including grazing, fertility, and moisture status [34]. Silva et al. [32] also highlights the difficulty in monitoring intensively grazed pastures due to their highly dynamic changes throughout the forage development cycle.

Our study demonstrated the capability of NDVI obtained by Sentinel-2 to capture the spatial variability of pasture quality across all experimental fields and during each collection event. We observed a high [38] and positive correlation between NDVI and CP and PMC, and a significant [38] and negative correlation between NDVI and NDF (Table 8). The global analysis, based on a comparison between all average values of pasture quality parameters and NDVI, obtained in each experimental field and each collection event in the three years of study, improved the accuracy of estimation models (R² of 0.7198 for PMC linear model, Figure 7; R² of 0.7475 for CP polynomial model, Figure 8; and R² of 0.5031 for NDF polynomial model, Figure 9). Based on the classification of Askari et al. [39], our predictive models for PMC and CP are good and the prediction of NDF is poor. The relationships between spectral measurements and pasture quality parameters, such as CP content, can be attributed to the absorbance of visible light by chlorophyll, which is abundant in green vegetation [34].

The potential of NDVI obtained by Sentinel-2 images to estimate, monitor, or map pasture quality parameters has been explored in previous studies [15,22,35]. As in our study, they all obtained more robust models for estimating CP than fiber (NDF). For example, Lugassi et al. [35] have obtained R² values of 0.69–0.72 for CP and 0.63–0.71 for NDF, while Ileri and Koç [22] obtained R² values of 0.57 for CP and, although in general significant, the highest determination coefficient between NDVI and NDF was still poor (only 0.29), measured at the end of the grazing season. These promising results indicate that the seasonal variation in CP content could be monitored using NDVI derived from Sentinel-2 imagery [15,22].

The NDF content of available forage is as important as the CP because their ratio significantly impacts animal performance [22]. Nevertheless, the correlation of NDF pasture content with NDVI values explained only poorly to moderately the variation of NDF and, according to Ileri and Koç [22], this relation might disappear in vegetation rich in woody and non-photosynthetic tissues.

4.3. Limitations and Perspectives

The results of our study, although promising given the wide range of pastures considered over a period of three years, also show that further research is necessary to explore ways to improve these pasture quality predictions. Several authors have pointed out some constraints on this purpose. For example, the spatial consistency [32], which corresponds to the differences in sampling size (in our case, five sub-samples of 0.5 m × 0.5 m) and satellite pixel size (in our case, 10 m × 10 m), has a standard error associated. This is, according to Silva et al. [32] and Serrano et al. [7,53], one of the major modelling uncertainties that might have caused the low correlations. Increasing the sampling area (collecting a greater number of pasture samples or harvesting larger areas within the pixel) would make the process unpractical and very expensive. For future studies, a previous survey, based in soil electrical conductivity or in satellite imagery time-series, could be used as a starting point for the establishment of management zones and implementing an effective smart sampling design [15] in each experimental field. Prior knowledge of the variability patterns (soil and pasture) in each experimental field makes it possible to establish zonal sampling, which simplifies the process and significantly reduces the number of samples needed to capture and validate this variability. The establishment of management zone maps is the basis of differentiated management using variable rate technology (VRT), such as the application of fertilizers, amendments or seeds, closing the Precision Agriculture cycle.

Another aspect to consider going forward, to improve the relationship between Sentinel-2 data and predicted variables, is that it is worth exploring all possible spectral band combinations, multiple ratios and other indices, rather than relying solely on the NIR and RED bands used to calculate NDVI. This is a question that gains relevance when pasture is composed of a diverse mixture of grasses, legumes and other plant species [1], as in this case. Biodiverse pastures are composed of different botanical families and species and, consequently, different rates of phenological development, which can make it difficult to establish relationships between pasture quality parameters and reflectance [15].

In our study, to minimize the gap between pasture ground sampling and radiometric data, we ensured a maximum difference of 5 days between imaging date and sampling date. In future works, it will be no less important to program the date on which pasture samples are collected according to the date the satellite is capturing the images, so that there is no time lag between the two. This mismatch between sampling and Sentinel-2 data is another limiting factor referred to, when calibrating robust models [15]. This can be particularly important during the period of greatest vegetative growth, usually during spring, when the temperature and soil moisture combine favorably, or at the end of spring, when the temperature rises very quickly, accelerating the process of plant dehydration, with a very significant impact on the color of the leaves and, therefore, on their spectral response. On days when the satellite images coincide with cloud cover, the NDVI survey can be complemented by a proximal optical sensor, such as the ‘OptRx’. The results obtained by Serrano et al. [7] showed the potential of this technology. Proximal sensing can help to overcome the constrain of scattered trees, resolving the difficulty of access to the signal of reflectance of vegetation under tree canopy, a fundamental element in the Montado ecosystem [54]. This aspect was also emphasized by Fernández-Habas et al. [15] for Dehesa, in Spain.

Still, our results, reflected in the nonlinearity (polynomial equations) of the spectral reflectance (NDVI) for estimating CP and NDF, seem to show a saturation effect for high values of CP (>15%) and low values of NDF (<40%), and therefore in the initial phases of the vegetative cycle, at the beginning of winter. Several studies report a similar problem of saturation of NDVI for top biomass levels, but at an advanced stage of the growing season [11,55]: a decrease in NDVI as a consequence of decrease in the chlorophyll content, even though total biomass continues to increase until the season’s end [6]. In this regard, Punalekar et al. [1] suggest that pasture quantity parameters should be considered when developing RS-based pasture quality prediction models. This is due to the higher sensitivity of reflectance to productivity and the generally weaker and non-specific relationships between most quality parameters and the spectral data.

Our suggestion for the near future in the field of pasture monitoring is to develop a tool (proximal or remote) that makes it possible to quickly, and simultaneously, estimate pasture productivity and quality. A parameter such as CP, expressed not as a percentage of dry matter but in kg of CP per ha (i.e., CP in % of DM × DM in kg ha⁻¹), would be extremely useful [51]. This parameter, in terms of PS, can be calculated based on chlorophyll sensors, where optical sensors have proven to give good results in the “2D” monitoring of pasture quality. This information can be complemented by measuring a productivity indicator parameter, for example, the compressed height of the pasture obtained by the Rising Plate Meter, conferring to it the third dimension (3D) of the parameter presented above. Previous studies have also explored the complementarity between various sensors to predict productivity and pasture quality (such as leaf nitrogen concentration, CP, NDF or others). These studies include comparisons between proximal sensors [11]; between PS and satellite images (RS) [35]; between different spectral configurations of Sentinel-2 (for example, red-edge and NIR bands, which have significant potential for detecting changes in nitrogen, chlorophyl and fiber [15]); and between different RS sources (for instance, Sentinel-1 and Sentinel-2 data) [56].

As a final note, it is important that new studies based on emerging technologies (e.g., sensors, geographic information systems—GIS, GNSS, satellite images) and their development focus on issues with strong environmental impact, such as monitoring indicators of the risk of pasture degradation (risk maps) [57,58].

5. Conclusions

Mediterranean livestock systems are generally not very productive, with low profit margins. The challenge of climate irregularity accentuates the complexity of this production context. Therefore, it is of great interest to develop predictive models of pasture quality, based on easily accessible, open-source and low-cost tools that provide real-time information and facilitate effective decision making to improve the efficiency and sustainability of Montado and its ecosystem services.

The results obtained in this large-scale study (three years and eight experimental fields) show a significant correlation between pasture quality parameters (PMC, CP and NDF) obtained by standard laboratory methods and NDVI satellite-derived data (R² of 0.72, 0.75, and 0.50, respectively). These promising findings encourage further research (i) to test and develop other vegetation indexes for monitoring pasture nutritive value; (ii) to extend this research to pastures of other Mediterranean countries, building large and representative datasets and developing more robust and accurate monitoring models based on freely available Sentinel-2 images; (iii) to implement an extension program for agricultural managers to popularize the use of these technological tools as the basis of grazing and pasture management (e.g., to adjust stocking rates, spatio-temporal grazing, supplementary feeding or pasture improvements).