Modeling techniques for simulating total forage biomass in rangelands of the Caatinga Biome, Brazil

ABSTRACT

From 2018 to 2019, total forage biomass (TFBM) harvests were carried out in four transects established in the rangelands of the Caatinga biome, which exhibited varying woody density. In northeastern Brazil, these transects are located in the municipalities of Tauá (Ceará state) and Ouricuri (Pernambuco state). The objective was to use artificial neural network (ANN) and phytomass growth (Phygrow) models to simulate rangeland production. ANNs acquired by Sentinel 2-A remote sensing data were used to obtain six vegetation indices (VIs): normalized difference vegetation index (NDVI); soil-adjusted vegetation index (SAVI); enhanced vegetation index (EVI); leaf area index (LAI); modified soil-adjusted vegetation index 2 (MSAVI₂); and normalized difference water index (NDWI). These indices were calculated in QGis software (version 2.18). The development of the ANN model incorporated both measured TFBM and the data collected by the aforementioned VIs. The Phygrow model was performed using the virtual platform Phyweb. The validation models summary revealed that the Phygrow model overestimated TFBM by more than 7.5 %. A comparison of the ANNs for all sites with the Phygrow model revealed that Phygrow had lower root mean square error (RMSE) values (88 vs 462 kg ha^–1 of dry matter) and a higher Willmott index (0.84 vs 0.52). However, the ANN model exhibited a lower percentage of bias (0.5 vs 7.5 %) and lower mean bias error (MBE of 6.9 vs 13.2). Phygrow demonstrated superior performance in simulating TFBM when compared to the ANN model, thereby substantiating its efficacy in estimating TFBM in rangelands of the Caatinga biome. In regions characterized by variable woody density, such as the Caatinga, the Phygrow model emerges as a useful tool for the optimization of rangeland management.

Artificial Neural Networks; Phygrow model; climate change; semi-arid regions; stocking rate

Introduction

The Caatinga biome in northeastern Brazil is considered to be among the most threatened, with a significant portion of the native forest having been converted for the creation of rangeland areas in response to the demand for livestock production. The irregular rainfall distribution in the region renders agricultural practices risky, compelling many farmers to opt for livestock production (Santos et al., 2010). However, the ongoing destruction of native forests to create rangeland areas for livestock production has become a major concern (Silva et al., 2020). Furthermore, the combination of intense droughts and high stocking rates has had a detrimental impact on these rangelands, exacerbating Brazil’s ongoing degradation of the Caatinga biome.

In rangelands, it is imperative to understand and forecast the quantity of forage accessible to livestock, otherwise known as total forage biomass (TFBM). The TFBM determines the carrying capacity of rangelands. Thus, stocking rates may be adjusted based on the rangeland condition, time of year, and animal condition (Díaz-Solís et al., 2009). Rangelands experience fluctuations in productivity throughout the year (Magliano et al., 2015), and a comprehensive understanding of these dynamics is imperative for determining the sustainable capacity of an area. The prevailing method for assessing available biomass involves directly extracting forage through cutting and subsequent weighing. However, this approach is often laborious and time-consuming, particularly in extensive areas and regions with limited accessibility. Consequently, there is a pressing need to develop and study reliable biomass monitoring systems that are practical and easily accessible to farmers (Sibanda et al., 2017).

In remote sensing, empirical models are predominantly characterized by selecting spectral bands or vegetation indices (VIs) that exhibit a high degree of correlate highly with forage biomass production (Verrelst et al., 2010). In contrast, mechanistic models are predicated on the laws of physics, chemistry, and biology to simulate a process or system, dividing it into components to understand the whole behavior in terms of the responses of the individual components and their interrelationships (Lara and Rakocevic, 2014). The objective of this study was to evaluate and compare the performances of an empirical model (ANN) incorporating select, remotely sensed VIs, and a mechanistic model (Phygrow) in simulating the biomass of rangelands in the Caatinga biome of Brazil.

Materials and Methods

Study site

To obtain field data for the study, four permanent cross-shaped transects, 25 m long from the center on each side, were established in the states of Ceará and Pernambuco (Brazil). The transects are located in the Caatinga biome (Figure 1A) and were monitored for two years. Two transects were established in the municipalities of Tauá, Ceará state (Figure 1B), on the Cachoeirinha do Pai Senhor Farm in the district of Barra Nova (5°37’10” S, 40°7’17” W, altitude 489 m). Two additional transects (7°57’43” S, 40°7’56” W, altitude 451 m) were established in the municipality of Ouricuri, Pernambuco state (Figure 1C).

Figure 1
– A) Biomes in northeastern Brazil, B) Municipalities of Tauá (Ceará state), and C) Ouricuri (Pernambuco state), where the permanent transects were established. Average monthly rainfall (mm) and maximum temperature (▼), average (○), and minimum (●) for D) Tauá and E) Ouricuri.

Tauá is a warm, semi-arid tropical region (IPECE, 2017), and the predominant vegetation is thorny deciduous forest and open shrubby Caatinga. Conversely, Ouricuri is situated in a region characterized by altitudinal variations in climate, resulting in distinct temperature patterns across different elevations. The climate is colder at higher altitudes, while the mountains and massifs’ foothills and slopes experience warmer temperatures. The vegetation in Ouricuri is predominantly deciduous forest and hypoxerophilic Caatinga (Beltrão et al., 2005). The rainy season typically extends from February to April in both Tauá and Ouricuri, while the dry season occurs from May to January.

ANN model: in situ vegetation assessment and measurement of TFBM development and training

After the transects were established, the polygon of the research area was delineated, with the limits obtained through land control points by GPS navigation Garmim Etrex 10. These limits were then processed in the Google Earth Pro tool (version 7.3.6.9345). The areas were evaluated for their shrub-tree density, with total density determined by counting the shrub and tree plants in each quadrant of the transects.

Total forage biomass was defined as the sum of the herbaceous and shrub-tree biomass up to a height of 2 m, since 2 m is the maximum height that local ruminants (i.e., cattle, goats, and sheep) can reach during grazing. The in situ measurement of TFBM was carried out using the destructive method. To sample and measure the biomass of the herbaceous layer, we established two frames measuring 0.25 m² (1 × 0.25 m) in each quadrant of the transect, resulting in a total of eight frames per transect. The herbaceous forage within each frame was cut close to the ground (Figure 2). For estimating the total biomass of the shrub-tree layer, we placed two frames measuring 20 m² (2 × 10 m) in each quadrant of the transect. The referential unit method was used to sample the biomass of each plant referring to that stratum. The method entailed the selection of three branches (reference unit) of each shrub-tree plant present in the frame, and the plant biomass was estimated by quantifying the number of reference units within the area extending to a height of 2 m.

Figure 2
– Flowchart of the steps to assess the total forage biomass (TFBM) model in rangelands of the Caatinga biome through remote sensing data and the artificial neural networks (ANNs) model. VIs = vegetation indices; NDVI = normalized difference vegetation index; SAVI = soil-adjusted vegetation index; EVI = enhanced vegetation index; LAI = leaf area index; MSAVI2 = modified soil-adjusted vegetation index 2; NDWI = normalized difference water index; RMSE = root mean square error; MBE = mean Bias Error.

The herbaceous and woody biomass samples were subsequently transferred to the laboratory for further analysis. They were placed in an oven with forced air circulation at 65 °C until they reached constant weight. Subsequently, the samples were then weighed on a digital scale, and the TFBM was calculated by summing the biomass values from the herbaceous layer and woody vegetation.

The transects were then assigned designations based on their total density of woody plants. For instance, in Tauá, the transect exhibiting a density of 144 plant ha^–1 was designated Tauá Density 144 (TD144, Figure 3A), and the transect with 280 plant ha^–1was designated Tauá Density 280 (TD280, Figure 3B). A similar approach was employed in Ouricuri, where the transect exhibiting 212 plant ha^–1was designated Ouricuri Density 212 (OD212, Figure 3C), and the transect with 372 plant ha^–1was designated Ouricuri Density 372 (OD372, Figure 3D).

Six VIs were utilized as input data to train the artificial neural network (ANN) model to simulate the TFBM. Images were obtained from the Sentinel-2A satellite and the Earth Explorer platform (United States Geological Survey) for the dates closest to the sample collection dates. The atmospheric correction procedure was followed to obtain the surface reflectance factor (ρ) from the extension of the Semi-Automatic Classification Plugin (Congedo, 2016), present as a complement to the processing of remote sensing data in the Qgis software (2.18). The correction method employed was dark object subtraction, as proposed by Chavez (1988, 1989), which corrects the interference of atmospheric effects from the digital numbers of the satellite image, ignoring the atmospheric radiation absorption. The VIs utilized in this study and their respective formulas are listed in Table 1.

Table 1
– Spectral vegetation indices (VIs) used in this study.

VIs	Equation	Reference
NDVI	$\frac{(\rho b8-\rho b4)}{(\rho b8+\rho b4)}$	Pearson and Miller (1972)
SAVI	$\frac{(1+L)\times (\rho b8-\rho b4)}{(\rho 8+\rho b4+L)}$	Huete (1988)
LAI	$\frac{-\ln \frac{(-0.69-SAVI)}{0.59}}{0.91}$	Allen et al. (2007)
MSAVI2	$\frac{(2\times \rho b8+1-\sqrt{(2\rho b8+1{)}^2-8(\rho b8-\rho b4)})}{2}$	Qi et al. (1994)
NDWI	$\frac{(\rho b5-\rho b11)}{(\rho b5+\rho b11)}$	Gao (1996)
EVI	$\frac{2.5\times (\rho b8-\rho b4)}{(1+\rho b8+C1\times \rho b4-C2\times \rho b2)}$	Huete et al. (1994)

• NDVI = normalized difference vegetation index; SAVI = soil-adjusted vegetation index; LAI = leaf area index; MSAVI₂ = modified soil-adjusted vegetation index 2; NDWI = normalized difference water index; EVI = enhanced vegetation index. r = surface reflectance factor; rb8 = infrared band; rb2 = blue band; rb4 = red band; rb11 = short-wave infrared 1 (SWIR1) band. Canopy background adjustment (L), C1 and C2 are the coefficients of the aerosol resistance term, which uses the blue band to correct for aerosol influences in the red band; L= 0.5 to 0.25; C1 = 7.5; C2 = 2.5.

In the implementation of soil-adjusted vegetation index (SAVI), the coefficient of soil adjustment exhibited the following variations: L = 0.50 was adopted for thin vegetation and L = 0.25 for the more conserved vegetation, given that thin vegetation has a more significant contribution of the soil element. The leaf area index (LAI) is calculated as leaf area per unit ground area (per m²). The VIs were subsequently paired with the biomass data for each quadrant (Figure 3E-H).

Artificial neural network models, including forecasting and other machine learning applications, are widely used in research on the rangelands. The utilization of remote sensing data proves advantageous in the construction of empirical models. Inspired by the functioning of the human brain (Gupta, 2013), ANN models consist of an interconnected set of processing units known as artificial neurons or nodes. These nodes are typically organized into an input layer, one or more hidden layers, and an output layer (Dongare et al., 2012). The most important input parameters of remote sensing in the ANN model are determined by analyzing the TFBM data with International Business Machines Corporation Statistical Package for Social Sciences software (version 20.0.0). This software is predicated on inferential statistics and employs recorded data to perform a sensitivity analysis to determine the importance of each input parameter (Rana et al., 2018).

The multi-layer perceptron (MLP) procedure was utilized for this analysis, constituting a supervised method that employs feedforward architecture. The MLP procedure comprises an input layer, one or more hidden layers, and an output layer. The input nodes receive the data values and pass them to the first nodes of the hidden layer, which, after multiplying, assigns a bias to that sum, and passes the results through a non-linear transformation such as the sigmoid transfer function (Kashaninejad et al., 2009). When a dependent variable is categorical, the neural network arranges cases into the “best” category based on the input predictors. The training of the ANN model was predominantly supported by 70 % of TFBM data, with the remaining 30 % allocated for model evaluation and testing.

Phygrow model calibration

The calibration of phytomass growth (Phygrow) was performed on Phyweb, the virtual platform of the Phygrow model (Angerer, 2017). The Phygrow model simulates biomass production under ideal growing conditions and then adjusts estimates of plant growth downward based on water levels, temperature stress, and grazing demands (Angerer, 2012; Stuth et al., 2003). In simulating TFBM production, the Phygrow model incorporates data on subcomponents of this process, including the types and coverage of plants and their characteristics, the climate, the soil and soil water balance variables, and grazing practices in the area. The interplay among vegetation, soil, climate, and grazing practices within a designated area collectively determines the annual soil water content, thereby directly influencing TFBM.

The parameterization of the Phygrow model was performed by collecting data on vegetation cover, and the botanical composition of plant species was collected (Rhodes et al., 2014). The percentage of vegetation and soil cover (e.g., the presence of monocots, dicots, shrub-trees, and litter or lack thereof) was determined by the visual method and entered the Phygrow model as a variable of plant community composition. Following an analysis of the vegetation types, the plant subcomponent was subdivided into three functional groups: monocotyledons, herbaceous dicots, and woody plants or shrub-trees. Growth parameters for specific functional groups were also entered, including LAI, relative growth rate, leaf renewal, root depth, plant height, and ideal temperatures for plant growth. Following the initial characterization of the two ecological sites, periodic forage biomass harvests were conducted at the sites for the model calibration and validation processes. The types and percentage of species found in each of the ecological sites, and the parameters for each of the three functional groups of plant species added during the model parameterization process are presented in Table 2.

Table 2
– Coverage of monocotyledons, dicotyledons, shrub-trees, soil, and litter. Additional information on three functional groups for the purpose of parameterization of the phytomass growth (Phygrow) model.

Transects	Coverage (%)
	Monocotyledons	Dicotyledons	Shrub-tree	Soil	Litter
TD144	30	30	15	10	15
TD280	20	20	30	10	20
OD212	10	38	20	30	12
OD372	26	44	40	30	10
Functional Groups	Phygrow Parameters
	LAI	Height	Biomass		Root depth
		cm	kg ha–1 of dry matter
Monocotyledons	1.80	50	1100		40
Dicotyledons	1.51	50	1000		40
Shrub-Tree	2.21	200	250		60

• TD144 = Tauá density 144; TD280 = Tauá density 280; OD212 = Ouricuri density 212; OD372 = Ouricuri density 372; LAI = leaf area index.

The climate subcomponent includes six variables: year, day, maximum temperature, minimum temperature, rainfall, and solar radiation. Phyweb receives daily updates of Climate Hazards Group InfraRed Precipitation with Station Data rainfall data (Figure 1D and E), with temperature and rainfall measurements obtained by entering the latitude and longitude of each monitoring location.

The soil subcomponent has 13 parameters, including soil depth, density, infiltration, and water retention variables. Samples were collected from the surface layer (0-20 cm) during the initial establishment of the transects to determine the soil characteristics and facilitate the parameterization of the soil subcomponent. The soil water balance accounting model encompasses seven variables that simulate the quantities of water available to plant roots: water interception, runoff, evaporation, infiltration, percolation, transpiration, and deep drainage. The Phygrow model can accommodate unlimited soil layers to delineate differences in soil texture, saturated hydraulic conductivity, field capacity, and rock factor (Stuth et al., 2003).

The grazing subcomponent of the model comprises 19 variables, which reflect the forage intake and stocking rate of each type of grazing animal, and their grazing preference for each plant species.

The parameters were entered for two model subcomponents in the calibration stage and subsequently verified through simulation. This process was undertaken to ascertain the discrepancies between the biomass values obtained through field collection and those derived from simulation. The model calibrations were carried out for each pair of subcomponents, with adjustments to each subcomponent made as necessary. The model calibration was deemed successful if the discrepancy between the simulated and field values fell below the standard error of the media. Subsequent to calibration, the model parameters are no longer adjusted, and the residual biomass data are utilized for model validation. During the calibration stage, 70 % of the data were utilized, corresponding to seven biomass data points per scenario. The remaining 30 % of the data were employed for validation, resulting in a single validation with the 12 data points from the four monitoring locations.

Evaluation of the simulation models

Given the absence of a normal distribution, a normality analysis of the data was conducted. The Spearman’s correlation analysis was selected to assess the relationship between TFBM and VIs. This analysis was performed using the CorrPlot package (Wei et al., 2017) from the RStudio Software version 4.4.1 (also utilized for plotting the correlation matrix). The Willmott index was employed to validate the models (Table 2). This index ranges from zero to one, with one representing perfect accuracy and zero representing no accuracy (Willmott, 1981).

The statistics used to validate the models are shown in Table 3. The mean bias error (MBE) was used to measure the accuracy of the simulated TFBM with respect to the actual measured values. The percent bias has an optimal value of zero, with small values indicating good accuracy, positive values indicating model overestimation, and negative values indicating model underestimation. The root mean square error (RMSE) verifies the magnitude of the difference between the simulated TFBM and the observed TFBM. A linear regression analysis was performed on the observed and simulated data to verify the model accuracy. Coefficient of determination (R²) values vary from 0 to 1, with an R² closer to 1 indicating a better model. The performance of the models for the evaluation process was obtained using the hydroGOF package of the R software (R Core Team, 2019) and presented by graphical and numerical methods using the SigmaPlot software (version 11.0).

Table 3
– Statistics used to measure the performance of models in simulating total forage biomass (TFBM) in rangelands of the Caatinga biome.

Statistics	Equation
Willmott index	$1-\frac{\Sigma (Si-Mi{)}^2}{\Sigma \left(\right[Si-M]+[Mi-M]{)}^2}$
MBE	$\frac{\sum _{i=1}^{N}Si-Mi}{N}$
Percentage of bias	$\frac{Si-Mi}{N}$
RMSE	$\sqrt{\frac{\sum _{i=1}^N(Mi-Si{)}^2}{N}}$

• For all the equations described above: S = i-th value from the simulation; M = i-th measured value; N = number of pairs of simulated and measured TFBM; MBE = mean bias error; RMSE = root mean square error.

Results

Data correlation and ANN model

For Tauá, the NDVI showed the highest Spearman’s correlation (r = 0.49) with TFBM, followed by SAVI and LAI (both r = 0.48); the normalized difference water index (NDWI) showed the lowest correlation (r = 0.31) with TFBM (Figure 4A). For Ouricuri (Figure 4B), SAVI showed the highest Spearman’s correlation (r = 0.48) with TFBM, followed by enhanced vegetation index (EVI) and modified soil-adjusted vegetation index 2 (MSAVI₂) (both r = 0.44). When the data from all sites were used (Figure 4C), SAVI had the highest Spearman’s correlation (r = 0.48), followed by LAI (r = 0.45).

Figure 4
– Spearman’s correlation between total forage biomass (TFBM) measured in rangelands of the Caatinga biome and spectral vegetation indices obtained from remote sensing data A) Tauá data; B) Ouricuri data; and C) All sites. NDVI = normalized difference vegetation index; SAVI = soil-adjusted vegetation index; EVI = enhanced vegetation index; MSAVI2 = modified soil-adjusted vegetation index 2; NDWI = normalized difference water index; LAI = leaf area index; Corr = Spearman’s correlation.

The normalized importance (expressed as percentages) of the VIs as features and their classification in estimating TFBM using ANN models are shown in Figure 5A-C. For Tauá rangelands, SAVI had the greatest normalized importance (0.274), followed by LAI (0.242). For Ouricuri, NDVI had the greatest normalized importance (0.285), followed by SAVI (0.283). When the normalized importance was analyzed for all sites, SAVI was the most important index (0.315), followed by LAI (0.244).

Figure 5
– Normalized importance (expressed as percentages) of spectral vegetation indices for each predictor A) Tauá data; B) Ouricuri data; and C) All sites. Linear regression of measured and simulated total forage biomass (TFBM) data obtained for artificial neural network D) Tauá data; E) Ouricuri data; and F) All sites. NDVI = normalized difference vegetation index; SAVI = soil-adjusted vegetation index; EVI = enhanced vegetation index; MSAVI2 = modified soil-adjusted vegetation index 2; NDWI = normalized difference water index; LAI = leaf area index. N = number of samples; R2 = coefficient of determination. (●) Data Points Regression.

When evaluating the performance of the ANN model in the different sites (Figure 5D-F), the best fit was found in the Tauá data, as indicated by the lower RMSE and higher R² values (0.40) as compared to the other sites (Ouricuri data, R² = 0.21 and all sites R² = 0.23). The Tauá data also had the highest Willmott index values (0.74), followed by the data from all sites combined (0.52) and the Ouricuri data (0.47).

Calibration and validation of the Phygrow model

Calibration of the Phygrow model to field data was generally successful (Figure 6B-D and F). The average TFBM harvested in rangeland TD144 was 1791 kg ha^–1, compared to a simulated average of 1930 kg ha^–1. The average TFBM harvested in rangeland TD280 was 1357 kg ha^–1, compared to a simulated average of 1535 kg ha^–1. In rangeland OD212, the average TFBM measured at the site was 1024 kg ha^–1, compared to a simulated average of 1204 kg ha^–1. In rangeland OD372, the average TFBM measured was 1002 kg ha^–1, while the simulated mean was 1021 kg ha^–1. In each of these sites, the simulated averages were higher than the measured averages, indicating a tendency for the Phygrow model to overestimate TFBM values.

Figure 6
– Soil water content in A) Tauá and D) Ouricuri and total forage biomass (TFBM) as measured and simulated by the phytomass growth model (kg ha–1) for each of the study transects in the Caatinga biome. Woody densities: B) Tauá density 144, C) Tauá density 280, E) Ouricuri density 372, and F) Ouricuri density 212. Data obtained from the Phyweb platform. (●) Data collected in the field (2018 and 2019).

Comparison of model validation: Phygrow vs ANNs

In the validation phase, the Phygrow model (Figure 7 and Table 4) overestimated the TFBM by more than 7.5 %. Nevertheless, the Phygrow model had an RMSE of 88 kg ha^–1, a Wilmott index of 0.84, and an R² of 0.68, indicating that the model can simulate the values of TFBM of the Caatinga rangelands. Comparing the ANN model for all sites with the Phygrow model (Table 4), it was observed that Phygrow had lower RMSE values (88 vs 462 kg ha^–1 of dry matter) and a higher Willmott index (0.84 vs 0.52); however, the ANN model had a lower percentage of bias (0.5 vs 7.5 %) and a lower MBE (6.9 vs 13.2). The ANN model exhibited significantly lower R² values for each site and all sites combined.

Figure 7
– Comparison between the total forage biomass (TFBM) as measured and simulated by the phytomass growth model (kg ha–1) in validation. (●) Data collected in the field and the bars correspond to the standard error. N = number of samples; R2 = coefficient of determination.

Table 4
– Summary of models’ performances.

Statistic	Phygrow	ANNs
	Phygrow simulated	ANNs simulated Tauá data	ANNs simulated Ouricuri data	ANNs simulated in all sites
Measured (kg ha–1 of dry matter)	1285	1382	1276	1338
Simulated (kg ha–1 of dry matter)	1391	1368	1289	1345
Bias (%)	7.50	–1.00	1.30	0.50
RMSE (kg ha–1 of dry matter)	88	409	463	462
MBE (kg ha–1 of dry matter)	13.20	–13.7	12.60	6.90
R 2	0.68	0.40	0.21	0.22
Willmott Index	0.84	0.74	0.47	0.52
N	12	56	40	96

• ANNs = artificial neural networks; RMSE = root mean square error; MBE = mean bias error; R² = coefficient of determination; N = number of pairs of simulated and measured total forage biomass.

Discussion

Remote sensing is a technique that can capture a variety of vegetation characteristics, thereby facilitating the study of vegetation dynamics. For instance, remote sensing can detect morphological changes in the plant canopy that cause variations in the absorbed, transmitted, and reflected fractions of the incident solar radiation. These variations in reflectance are captured by orbiting sensors, which produce spectral images of the canopy surface (Morais et al., 2021).

The normalized difference vegetation index (NDVI) is the most widely used VI to simulate the biophysical parameters of vegetation. However, in some situations, it tends to saturate depending on two atmospheric noises and overestimates biomass values (Lu et al., 2014). Consequently, various VIs, such as SAVI, have been developed to address these limitations, which has been shown to enhance the accuracy of biomass estimates by correcting for atmospheric influences and exposure (Karnieli et al., 2013). However, the efficacy of SAVI in estimating the biomass of natural grasslands, such as those found in the Caatinga biome, remains to be thoroughly evaluated through remote sensing methods. Nevertheless, research has demonstrated this index’s significance in understanding plants’ biological and physical processes in plants (Galvíncio et al., 2013), including photosynthesis, evapotranspiration, interception of precipitation, and carbon flux (Sha et al., 2019).

Conversely, no studies have been conducted on the performance of empirical models derived from remotely sensed VIs in simulating TFBM in Caatinga rangelands. However, some studies have evaluated the use of remotely sensed VIs for other purposes. A high correlation (R² = 0.75) between NDVI and total woody vegetation biomass has been reported, leading to the estimation of woody biomass using remotely sensed NDVI with high accuracy (Francisco et al., 2017). Similarly, the correlation between NDVI, SAVI, LAI, and precipitation in a Caatinga site (specifically, the municipality of Aiuabas in the Ceará state) has been evaluated, demonstrating that these indices are important tools for studying rangeland production in regions where precipitation is poorly distributed and concentrated in a few months of the year (Teixeira et al., 2016).

In the present study, the low R² values obtained from the ANN model suggest that it is not accurately simulating TFBM. Conversely, when the model underestimates the TFBM values, the stocking rate will likely be low, potentially resulting in under grazing. Under such conditions, the animals can select the most nutritious forage crops, achieving maximal weight gain (Reis et al., 2009). However, it is imperative to note that the optimal stocking should be designed to prevent forage wastage while ensuring efficient land use.

The Phygrow model was evaluated in regions of East Africa, and the results demonstrated a correspondence of 96 % in the linear regression analysis with a standard error of 161 kg ha^–1 of dry matter (Kaitho et al., 2007). In the same study, the validation stage demonstrated the model’s high capacity to simulate biomass, indicating its potential as a helpful farm management tool, particularly for setting and adjusting stocking rates. Validation is a critical step in assessing the precision and accuracy of model forecasts, thereby enhancing the reliability of the estimates obtained by the model (Tedeschi, 2006). However, the simulation curves also show that the Phygrow model tends to overestimate the true biomass values at the onset of the rainy season, potentially due to an overestimation of the precipitation parameter, as demonstrated in Figure 6B-C and E-F.

The biomass production of Caatinga rangelands is directly related to the amount and distribution of rainfall and soil water content throughout the year. The rangelands of the Caatinga biome are distinguished by prolonged periods of drought. The low soil water content impedes the development of plant root systems, and plants that cannot reach deeper soil layers may experience more significant stress during dry periods. Consequently, the seasonality of the soil water content (Figure 6A and D) directly influences plant growth in the Caatinga, as evidenced by the empirical measurements of TFBM depicted in Figure 6B-C and E-F.

Phygrow is a mechanistic model that simulates above-ground herb and shrub growth, forage consumption, and hydrologic processes to provide near real-time assessment of current and future forage conditions. Its utilization has been observed in early warning initiatives and in developing forage production monitoring systems in multiple arid and semi-arid regions worldwide, including countries such as Kenya (Matere et al., 2020) and Mongolia (Angerer, 2012). However, a paucity of studies have evaluated the capacity of Phygrow to simulate TFBM in rangelands of the Caatinga biome in northeastern Brazil.

The utilization of simulation modeling in the context of TFBM serves as a valuable instrument for elucidating productivity characteristics, exhibiting a multitude of applications. The development and refinement of methodologies based on modeling can assist livestock farmers in adjusting their livestock herds and serve as a foundation for the formulation of public policies (Zilverberg et al., 2017; Morais et al., 2023). Despite the recognized significance of disseminating knowledge regarding modeling in simulation, a paucity of studies has evaluated the application of models in rangelands of the Caatinga. This paucity of research can be attributed to a lack of understanding regarding the capabilities and limitations of the use of modeling, and to the dearth of experience in evaluating, calibrating, and utilizing models. Notwithstanding the challenges, future trends indicate an increased utilization of modeling tools in livestock, contingent upon ensuring the proper calibration of these models. These models are expected to provide precise and accurate information regarding biomass estimation.

The r coefficient between remote sensing data and TFBM (in situ production) demonstrated a medium-to-low correlation in the Caatinga, primarily due to woody plants in areas not grazed by domestic herbivorous herds. Conversely, the empirical ANN model, employing remotely sensed spectral VIs, demonstrated efficacy in estimating and identifying the oscillation of TFBM in rangelands of the Caatinga biome. This model serves as a viable method for simulating TFBM. However, the findings of this study suggest that the mechanistic Phygrow model outperforms the ANN model in terms of validation, thereby affirming its superiority in simulating TFBM in rangelands characterized by variable woody densities. The Caatinga biome exemplifies this condition. The Phygrow model emerges as a valuable instrument for formulating sustainable management decisions and policies in such regions.

Figure 3
– Total forage biomass (TFBM) obtained in rangelands of the Caatinga and spectral vegetation indices (VIs value) obtained from remote sensing data. Rangelands with woody densities: A and E) Tauá density 144; B and F) Tauá density 280, C and H) Ouricuri density 372, and D and G) Ouricuri density 212. NDVI = normalized difference vegetation index; SAVI = soil-adjusted vegetation index; EVI = enhanced vegetation index; LAI = leaf area index; MSAVI2 = modified soil-adjusted vegetation index 2; NDWI = normalized difference water index.

Acknowledgments

The authors thank the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES), the Universidade Federal do Ceará (UFC), and the Empresa Brasileira de Pesquisa Agropecuária (Embrapa) for scholarship granter. The authors thank researcher Jay P. Angerer (Livestock and Range Research Laboratory, United States Department of Agriculture – USDA) for offering suggestions and providing guidance during the calibration and validation phases of the Phygrow model.

References

• Allen RG, Tasumi M, Trezza R. 2007. Satellite-based energy balance for mapping evapotranspiration with internalized calibration (METRIC) - Model. Journal of Irrigation and Drainage Engineering 133: 380-394. https://doi.org/10.1061/ (ASCE)0733-9437(2007)133:4(380)
  » https://doi.org/10.1061/
• Angerer JP. 2012. Gobi forage livestock early warning system. p. 115-130. In: Coughenour MB, Makkar HPS. eds. Conducting national feed assessments. Food and Agriculture Organization, Rome, Italy.
• Angerer JP. 2017. Phyweb Simulation Modeling Tutorial. Texas A&M University, College Station, TX, USA.
• Beltrão BA, Mascarenhas JDC, Miranda JLF, Souza Junior LC, Galvão MJTG, Pereira SN. 2005. Projeto cadastro de fontes de abastecimento por água subterrânea, estado de Pernambuco: diagnóstico do Município de Ouricuri. CPRM/PRODEEM, Recife, PE, Brazil (in Portuguese).
• Chavez Jr PS. 1988. An improved dark-object subtraction technique for atmospheric scattering correction of multispectral data. Remote Sensing of Environment 24: 459-479. https://doi.org/10.1016/0034-4257 (88)90019-3
  » https://doi.org/10.1016/0034-4257 (88)90019-3
• Chavez Jr PS. 1989. Radiometric calibration of landsat thematic mapper multispectral images. Photogrammetric Engineering and Remote Sensing 55: 1285-1294.
• Congedo L. 2016. Semi-Automatic Classification Plugin Documentation. Release 6.0.1.1. https://doi.org/10.13140/RG.2.2.29474.02242/1
  » https://doi.org/10.13140/RG.2.2.29474.02242/1
• Díaz-Solís H, Grant WE, Kothmann MM, Teague WR, Díaz-García JA. 2009. Adaptive management of stocking rates to reduce effects of drought on cow-calf production systems in semi-arid rangelands. Agricultural Systems 100: 43-50. https://doi.org/10.1016/j.agsy.2008.12.007
  » https://doi.org/10.1016/j.agsy.2008.12.007
• Dongare AD, Kharde RR, Kachare AD. 2012. Introduction to artificial neural network. International Journal of Engineering and Innovative Technology 2: 189-194.
• Francisco PRM, Chaves IB, Chaves LHG, Lima ERV, Silva BB. 2017. Antecedent and humidity index of the normalized of difference vegetation mapping caatinga. Agropecuária Científica no Semiárido 13: 82-91 (in Portuguese, with abstract in English).
• Galvíncio JD, Moura MSB, Silva TGF, Silva BB, Naue CR. 2013. LAI improved to dry forest in semiarid of the Brazil. International Journal of Remote Sensing Applications 3: 193-202.
• Gao BC. 1996. NDWI-A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment 58: 257-266. https://doi.org/10.1016/S0034-4257 (96)00067-3
  » https://doi.org/10.1016/S0034-4257 (96)00067-3
• Gupta N. 2013. Artificial neural network. Network and Complex Systems 3: 24-28.
• Huete AR. 1988. A soil-adjusted vegetation index (SAVI). Remote Sensing of Environment 25: 295-309. https://doi.org/10.1016/0034-4257 (88)90106-X
  » https://doi.org/10.1016/0034-4257 (88)90106-X
• Huete A, Justice C, Liu H. 1994. Development of vegetation and soil indices for MODIS-EOS. Remote Sensing of Environment 49: 224-234. https://doi.org/10.1016/0034-4257 (94)90018-3
  » https://doi.org/10.1016/0034-4257 (94)90018-3
• Instituto de Pesquisa e Estratégia Econômica do Ceará [IPECE]. 2017. Perfil Básico Municipal. IPECE, Tauá, CE, Brazil (in Portuguese).
• Kaitho RJ, Jama AA, Stuth JW, Kariuki G, Abdirahman A, MacOpiyo L, et al. 2007. Livestock early warning information resource in the Horn of Africa: Forage and livestock marketing information analysis and forecasts. Outlook on Agriculture 36: 267-272. https://doi.org/10.5367/000000007783418525
  » https://doi.org/10.5367/000000007783418525
• Karnieli A, Bayarjargal Y, Bayasgalan M, Mandakh B, Dugarjav C, Burgheimer J, et al. 2013. Do vegetation indices provide a reliable indication of vegetation degradation? A case study in the Mongolian pastures. International Journal of Remote Sensing 34: 6243-6262. https://doi.org/10.1080/01431161.2013.793865
  » https://doi.org/10.1080/01431161.2013.793865
• Kashaninejad M, Dehghani AA, Kashiri M. 2009. Modeling of wheat soaking using two artificial neural networks (MLP and RBF). Journal of Food Engineering 91: 602-607. https://doi.org/10.1016/j.jfoodeng.2008.10.012
  » https://doi.org/10.1016/j.jfoodeng.2008.10.012
• Lara MAS, Rakocevic M. 2014. Uso de modelos matemáticos no estudo de pastagens. p. 333-346. In: Reis RA, Bernardes TF, Siqueira GR. eds. Forragicultura: ciência, tecnologia e gestão dos recursos forrageiros. Funep, Jaboticabal, SP, Brazil (in Portuguese).
• Lu D, Chen Q, Wang G, Liu L, Li G, Moran E. 2014. A survey of remote sensing-based aboveground biomass estimation methods in forest ecosystems. International Journal of Digital Earth 9: 63-105. https://doi.org/10.1080/17538947.2014.990526
  » https://doi.org/10.1080/17538947.2014.990526
• Magliano PN, Fernández RJ, Mercau JL, Jobbágy EG. 2015. Precipitation event distribution in Central Argentina: spatial and temporal patterns. Ecohydrology 8: 94-104. https://doi.org/10.1002/eco.1491
  » https://doi.org/10.1002/eco.1491
• Matere J, Simpkin P, Angerer J, Olesambu E, Ramasamy S, Fasina F. 2020. Predictive livestock early warning system (PLEWS): monitoring forage condition and implications for animal production in Kenya. Weather and Climate Extremes 27: 100209. https://doi.org/10.1016/j.wace.2019.100209
  » https://doi.org/10.1016/j.wace.2019.100209
• Morais LF, Cavalcante ACR, Aquino DN, Nogueira FHM, Cândido MJD. 2021. Spectral responses in rangelands and land cover change by livestock in regions of the Caatinga biome, Brazil. Scientific Reports 11: 18261. https://doi.org/10.1038/s41598-021-97784-5
  » https://doi.org/10.1038/s41598-021-97784-5
• Morais LF, Cavalcante ACR, Furtado RN, Gorayeb A, Aquino DN, Nogueira FHM, et al. 2023. Spatial and temporal analysis (2008-2017) of droughts and their effects on livestock in the Brazilian semi-arid region. Anais da Academia Brasileira de Ciências 95: e20220890. https://doi.org/10.1590/0001-3765202320220890
  » https://doi.org/10.1590/0001-3765202320220890
• Pearson RL, Miller LD. 1972. Remote mapping of standing crop biomass for estimation of the productivity of the shortgrass prairie, Pawnee National Grasslands, Colorado. Remote Sensing of Environment 7: 1355.
• Qi J, Chehbouni A, Huete AR, Kerr YH, Sorooshian S. 1994. A modified soil adjusted vegetation index. Remote Sensing of Environment 48: 119-126. https://doi.org/10.1016/0034-4257 (94)90134-1
  » https://doi.org/10.1016/0034-4257 (94)90134-1
• Rana A, Rawat AS, Bijalwan A, Bahuguna H. 2018. Application of multi-layer (perceptron) artificial neural network in the diagnosis system: a systematic review. IEEE - International Conference on Research in Intelligent and Computing in Engineering (RICE) 7: 1-6. https://doi.org/10.1109/RICE.2018.8509069
  » https://doi.org/10.1109/RICE.2018.8509069
• Reis RA, Ruggieri AC, Casagrande DR, Páscoa AG. 2009. Supplementation of beef cattle as strategy of pasture management. Revista Brasileira de Zootecnia 38: 147-159 (in Portuguese, with abstract in English). https://doi.org/10.1590/S1516-35982009001300016
  » https://doi.org/10.1590/S1516-35982009001300016
• Rhodes EC, Tolleson DR, Angerer JP, Kava JA, Dyess J, Nicolet T. 2014. A comparison of rangeland monitoring techniques for modeling herbaceous fuels and forage in central Arizona, USA. Fire Ecology 10: 76-91. https://doi.org/10.4996/fireecology.1002076
  » https://doi.org/10.4996/fireecology.1002076
• Santos MVF, Lira MA, Dubeux Junior JCB, Guim A, Mello ACL, Cunha MV. 2010. Potential of Caatinga forage plants in ruminant feeding. Revista Brasileira de Zootecnia 39: 204-215. https://doi.org/10.1590/S1516-35982010001300023
  » https://doi.org/10.1590/S1516-35982010001300023
• Sha Z, Wang Y, Bai Y, Zhao Y, Jin H, Na Y, et al. 2019. Comparison of leaf area index inversion for grassland vegetation through remotely sensed spectra by unmanned aerial vehicle and field-based spectroradiometer. Journal of Plant Ecology 12: 395-408. https://doi.org/10.1093/jpe/rty036
  » https://doi.org/10.1093/jpe/rty036
• Sibanda M, Mutanga O, Rouget M, Kumar L. 2017. Estimating biomass of native grass grown under complex management treatments using worldview-3 spectral derivatives. Remote Sensing 9: 55. https://doi.org/10.3390/rs9010055
  » https://doi.org/10.3390/rs9010055
• Silva MV, Pandorfi H, Lopes PMO, Silva JLB, Almeida GLP, Silva DAO, et al. 2020. Pilot monitoring of Caatinga spatial-temporal dynamics through the action of agriculture and livestock in the Brazilian semiarid. Remote Sensing Applications: Society and Environment 19: 100353. https://doi.org/10.1016/j.rsase.2020.100353
  » https://doi.org/10.1016/j.rsase.2020.100353
• Stuth JW, Schmitt D, Rowan RC, Angerer JP, Zander K. 2003. Phygrow Users Guide and Technical Documentation. Texas A&M University, College Station, TX, USA.
• Tedeschi LO. 2006. Assessment of the adequacy of mathematical models. Agricultural Systems 89: 2-3. https://doi.org/10.1016/j.agsy.2005.11.004
  » https://doi.org/10.1016/j.agsy.2005.11.004
• Teixeira DBS, Teixeira LMN, Costa CAG. 2016. Correlation between precipitation and vegetation indexes under preserved Caatinga condition. Journal of Hyperspectral Remote Sensing 7: 21-30. https://doi.org/10.29150/jhrs.v7.1.p21-30
  » https://doi.org/10.29150/jhrs.v7.1.p21-30
• Verrelst J, Schaepman ME, Malenovský Z, Clevers JG. 2010. Effects of woody elements on simulated canopy reflectance: Implications for forest chlorophyll content retrieval. Remote Sensing of Environment 114: 647-656. https://doi.org/10.1016/j.rse.2009.11.004
  » https://doi.org/10.1016/j.rse.2009.11.004
• Wei T, Simko V, Levy M, Xie Y, Jin Y, Zemla J. 2017. Package 'corrplot'. Statistician 56: 1-18.
• Willmott CJ. 1981. On the validation of models. Physical Geography 2: 184-194. https://doi.org/10.1080/02723646.1981.10642213
  » https://doi.org/10.1080/02723646.1981.10642213
• Zilverberg CJ, Williams J, Jones C, Harmoney K, Angerer J, Metz LJ, et al. 2017. Process-based simulation of prairie growth. Ecological Modelling 351: 24-35. https://doi.org/10.1016/j.ecolmodel.2017.02.004
  » https://doi.org/10.1016/j.ecolmodel.2017.02.004

Data availability statement:

Data will be available upon request via email.

Declaration of use of AI Technologies:

Artificial intelligence (AI) technologies were not used.

*Corresponding author

<[e-mail]>

Edited by:

Vinícius Nunes de Gouvêa

Conflict of interest:

The authors declare no competing interests.

Publication Dates

• Publication in this collection
  27 June 2025
• Date of issue
  2025

History

• Received
  13 Jan 2024
• Accepted
  19 Nov 2024