Knowledge Discovery in Databases (KDD) Applied to the Demand for Technologies for Pasture Conservation and Management System

1. Introduction

In 2021, the beef cattle chain, one of the pillars of agribusiness in Brazil, generated US$ 178 billion, which represented around 11% of the country’s total GDP. Extensive systems account for around 80% of beef cattle production systems in Brazil. Its main characteristic is the use of pastures as the primary source of food. The animals’ nutrition is supplemented with minerals to cover deficiencies of both macro and micro elements. All the production phases in extensive systems take place in the field, from cow-calf to finishing. These production phases present a high variation in performance indices, which are directly dependent on pasture conservation and management technologies [1]. In Brazil, the pasture area, both natural and planted, occupies approximately 160 million hectares [2], which correspond to approximately 45% of the country's agricultural area [3].

Many Brazilian livestock producers raise more cattle than the pasture can sustain [4]. These findings help to clarify the overgrazing paradox by providing a foundation for understanding how producers’ objectives influence their perceptions and decisions. They believe in the economic gain provided by a greater number of animals per area, but some degree of pasture degradation is often a consequence [5].

Due to the large climatic differences between biomes, climate is the main factor that influences the production of tropical forage [6], underscoring its importance in the regional decision-making process carried out by farmers. Thus, producers serve as the primary agents responsible for adapting technologies and implement production systems in the real world [7]. Increasing the efficiency of beef production systems does not necessarily mean greater environmental impact. On the contrary, as in the case of degraded pastures, their recovery increases productivity and contributes significantly to carbon sequestration and fixation in the soil. Based on IBGE data, Martha et al. [8] reported that the stocking rate increased from 0.51 head per hectare in 1970 to 1.08 in 2006, reflecting the widespread adoption and a shift in livestock practices toward the important ”land-saving” effect.

Most of the ruminant production in Brazil is based on pastures, which are present in every Brazilian biome. This brings competitiveness to local production systems, as the use of grass-lands leads to lower costs. In addition, the productive potential of forage plants in Brazilian pastures and the possibility of intensification of production systems also favor the reduction of environmental impacts currently attributed to cattle raising [9].

Farm monitoring encompasses: (1) climate and weather conditions, such as temperature, humidity, and wind; (2) resilience to external factors that impact farm operations; (3) productivity, focusing on controlling outputs and related processes; (4) human involvement, aimed at adjusting and enhancing the efficiency of production systems; and (5) sustainability, with an emphasis on producing high-quality food. The collected data encompass all these aspects and provide insights for adopting technologies aligned with producers’ objectives. This accumulated knowledge has great value for end users in supporting informed decision-making [10]. In conventional analysis workflows, data often lose value over time, eventually being aggregated or discarded after a specific period. To address this issue, data should be periodically reassessed to generate new insights [11].

Knowledge discovery in database (KDD) is an important analytical method for identifying relationships among the various elements of the collected data. It aims to uncover actionable insights and to support strategic decision-making to promote sustainable systems and advance animal production, the food industry, and public health [12].

KDD follows a series of structured steps: (1) Selection - identifying relevant data aligned with the project objectives; (2) Pre-processing - improving data quality by addressing inconsistencies or missing values; (3) Transformation - applying appropriate methods based on the selected models; (4) Data mining – the core of the process, where algorithms are used for sorting, clustering, or statistical analysis to extract patterns, rules, or insights from the transformed data; and (5) Evaluation/Interpretation –analyzing and interpreting the results. As KDD is both iterative and incremental, the entire process or specific stages may be repeated multiple times [13].

The objective of this study was to extract knowledge about the technological demands and decision-making processes of livestock producers (stakeholders) operating within pasture-based production systems.

2. Materials and Methods

A nationwide survey was conducted electronically between July and August 2019, receiving responses from 712 participants across all Brazilian regions and biomes. Data were collected using a structured online questionnaire and included responses from various stakeholders (producers, researchers, administrators etc.) engaged in livestock systems where pasture serves as the primary source of animal feed. Four techniques commonly used in pasture planning and management were selected by the pasture specialists at Embrapa Pecuária Sudeste (São Carlos, SP) to serve as dependent variables in the analysis:

a) Soil analysis (PAnlSolo);

b) Monitoring and control of pasture stocking rate (PTxLot);

c) Soil correction with limestone (PCorrection); and

d) Rotation grazing (PRotacio).

The interviews provided information about the techniques used on each property based on the forage production system adopted. These techniques were classified as either adopted (1) or not adopted (0), resulting in dichotomous responses.

Using the estimation method of Generalized Linear Models (GLM), the probability (π) of a practice being classified as implemented (1) was modeled in relation to the independent (or contextual) variables associated with different forage production systems in Brazil, similarly to the methodology employed by [14].

The analysis of significant combinations of technological practices enabled a deeper understanding of the factors influencing the adoption of specific forage production systems. GLMs were employed using a logistic link function and binomial distribution, with the stepwise procedure applied to identify significant contextual variables (p>0.05). It is important to note that the estimated regression coefficients do not directly represent the effect of changes in the explanatory variables on the resulting probability (Pi). Instead, they indicate the effect of individual explanatory variables on the logarithm of the odds ratio (ln [P_i / (1-P_i)]).

For each set of significant contextual variables associated with a given dependent production technique, dedicated databases were developed and submitted to a data mining process. The objective was to extract knowledge on pasture production systems, which are composed of several interacting practices.

The quality of the binomial logistic model fit was evaluated using the Hosmer-Lemeshow test, which compares observed and expected outcomes by dividing the data into 10 groups. The test evaluates whether there are significant differences between the observed and expected frequencies within each group [15]. If such differences are not statistically significant at a certain significance level, the model is considered to have an adequate fit.

Another method used to evaluate model performance was the Receiver Operating Characteristic (ROC) curve. This technique visualizes and assesses the model’s ability to distinguish between classes by plotting sensitivity (true positive rate) on the Y-axis and specificity (false positive rate) on the X-axis. Originally developed in signal detection theory, ROC analysis is widely used to evaluate diagnostic systems. A key metric is the Area Under the Curve (AUC), which ranges between 0 and 1 and represents the model’s probability of correctly ranking a randomly selected positive observation higher than a randomly selected negative one [16].

The decision tree induction technique employed in this study utilizes hierarchical models widely applied areas across various fields due to their predictive accuracy and problem-solving capabilities, offering a straightforward and intuitive approach to knowledge discovery and decision-making from large datasets [17]. One key advantage of this method is its suitability to handle both continuous and discrete qualitative variables, enabling the construction of decision trees that classify data and highlight the most relevant attributes within their branches [18]. Decision trees apply multiple algorithms to determine optimal node splits, creating sub-nodes that increase homogeneity (node purity) with respect to the target attribute. The evaluates all available variables for potential splits and selects the one that yields the most homogeneous sub-nodes. Details about the methodology can be found in [19].

3. Results

Table 1 shows that most respondents to the electronic questionnaire were from the Cerrado (35.11%) and Atlantic Forest (34.27%) biomes, followed by the Caatinga (13.06%), Amazon (9.27%), Pampa (7.16%) and Pantanal (1.12%). Respondents indicated whether the analyzed technological practices were used on their properties.

Approximately 49.44% of participants identified as rural producers. The remaining respondents included technical assistance and rural extension agents (20.00%), private consultants (12.92%), and technicians from private companies related to agriculture (10.95%).

Table 2 lists the significant technological and contextual variables (p>0.05) identified using the stepwise method, for each technology deemed important by the pasture specialists.

3.1. Soil Analysis (PAnlSolo)

Table 3 presents the estimates and standard errors for the intercept and the parameters of the contextual variables: Fanalsolo (performs soil analysis for pasture formation); PCorrection (soil correction with limestone); and PMacro (fertilization with nitrogen, phosphorus, potassium, or sulfur).

When producers perform soil analysis during pasture formation, they likely also apply limestone and fertilization. Under these conditions, the average probability of conducting soil analysis in already established pastures is 97.28%. In contrast, when these technological practices are not adopted, the probability drops to 2.81%.

In the Hosmer-Lemeshow test, performed with 4 degrees of freedom, the test statistic was 2.76 (P = 0.59), indicating no significant difference between observed and expected values and, therefore, a good model fit. This result is consistent with the area under the ROC curve.

Figure 1 shows the ROC curve for the model analyzing the dependent variable PAnlSolo, with an AUC estimate of approximately 95%., which can be considered excellent.

To asses knowledge acquisition, a decision tree (DT) was constructed using the dataset to guide the decision-making process through a tree-like structure. Decision tree models enhance interpretability and rule transparency, which can increase users' confidence in employing them as decision-support tools for systematizing the adoption of technologies [22]. The trees were generated using Weka 3.8 software and the J48 classifier, an implementation of the C4.5 algorithm [23].

Each survey response constituted a data instance, and a response variable (PAnlSolo) was defined as the target attribute. For this binary variable, the model recorded either occurrence (Y) or non-occurrence (N) of soil analysis in established pastures. The decision tree was built using the four attributes identified by the binary logistic regression: the target attribute (PAnlSolo) and the predictors Fanalsolo (performs soil analysis during pasture formation); PCorrection (soil correction with limestone); and PMacro (fertilization with N, P, K, or S).

Figure 2 presents the decision tree constructed using the information gain criterion, based on Shannon Entropy estimates [24]. Of the 712 instances, 80% were used to train the algorithm. The model’s performance is summarized as follows: (a) accuracy rate of 89.44%; (b) six leaves (rules), indicating a relatively simple and interpretable model structure; (c) Kappa statistic of 0.78, which reflects substantial agreement between predicted and observed classes. The Kappa coefficient accounts for the agreement expected by chance and provides a more robust measure of classification performance. This value represents a substantial level of agreement [25].

Table 4 presents the decision rules and their respective confidence levels obtained through decision tree induction using the J48 algorithm. As expected, there was a strong connection between the technological attributes identified as significant in the logistic regression analysis and the target attribute (PAnlSolo). The model achieved an accuracy rate of 85.67%, indicating that the adoption of the PAnlSolo process is closely linked to the use of PCorrection, FAnalSolo and PMacro, supporting producer’s decision-making regarding soil analysis. The result of the soil analysis, whether conducted during pasture formation or in established pastures, provide essential parameters for calculating the need for soil amendments and fertilizers.

3.2. Monitoring and Control of the Pasture Stocking Rate (PTxLot)

Stocking rate is a complex variable influenced by a wide range of environmental and zootechnical factors. Defined as the number of animals per unit area, it reflects the concept of carrying capacity, that is, the level of grazing pressure that maintains a dynamic balance between animal gain and forage availability. Forage supply regulates the stocking rate and directly affects pasture quality and quantity. Thus, decision-making around stocking rate adjustment must consider the interaction between forage availability and livestock productivity [26].

Optimal grazing intensity refers to the ideal level of pasture utilization that maximizes livestock production without disrupting the ecological balance among plant species. There is a critical threshold of forage availability, which is the minimum amount required to maintain maximum animal intake and performance [27]. Below this point, gain per animal remains steady as stocking increase, potentially triggering pasture degradation and a decline in cattle body condition.

Table 5 presents the estimates and the standard errors for the intercept and parameters of nine contextual variables: areapec (livestock area), FContInv (weed control during pasture establishment); PCurvaNivel (use of contour terraces); PPlan (planning of forage production sectors); PAnlSolo (soil analysis); Fanalsolo (performs soil analysis in the formation); PMicro (micronutrient fertilization); PInvHerb (herbicide-based weed control); PManagement (monitoring and control of grazing frequency and intensity).

Figure 3 presents the ROC curve for the model analyzing the dependent variable PTxLot, with an AUC estimate of 87%, indicating a very good model fit. Although the Hosmer-Lemeshow test result was not statistically significant (p > 0.05), it was near the threshold. The test, performed with 8 degrees of freedom, yielded a value of 14.49 (p = 0.0698), suggesting the model is adequately specified.

Producers with livestock areas between 1,001 and 2,500 hectares and who adopt the associated technological practices have a 97.54% probability of also implementing pasture stocking rate monitoring and control. In contrast, producers with the same land area who do not adopt these technologies have only a 13.89% probability of using stocking rate control as a management tool. Figure 4 illustrates that as the livestock area increases and technological practices related to stocking rate monitoring are adopted, the likelihood of implementing stocking rate also increases.

Figure 5 illustrates the decision tree induced using the J48 algorithm, with 70% of the instances allocated for training and PTxLot as the target attribute. Due to the large number of leaves (over 30), interpreting the resulting rules became complex. To address this, a tree pruning strategy was applied based on a calculated error estimate. The J48 algorithm was used to compute this estimate using the training dataset [28].

The Kappa statistics and accuracy were estimated at 0.47 and 76.63%, respectively. The Kappa value indicates a moderate level of agreement. To improve the interpretability of the decision rules, the pruning parameter was adjusted from the default 25% to 15%. This modification reduced the number leaves from 30 to 8, thereby facilitating a clearer understanding of the target attribute. However, two variables (Areapec and PCurvaNivel) were deemed non-informative in the final model.

Table 6 presents the decision rules associated with the adoption of the stocking rate control technology, generated through the J48 decision tree algorithm. The results highlight the complexity of stocking rate management, which is influenced by numerous factors, some of which are beyond direct control. Stocking rate is a fundamental concept in pasture management [29]. A proper understanding of related terms such as stocking rate, grazing pressure, and carrying capacity, is essential. This aligns with the first attribute, PManagement (monitoring and control of plant grazing frequency and intensity), which serves as the root node of the decision tree.

3.3. Soil Correction with Limestone (PCorrection)

The use of limestone and fertilizers is essential for the intensification of pasture management [30]. Limestone plays a key role in correcting soil acidity, as the neutralization process requires direct contact between soil particles and either the limestone itself or its by-products. Consequently, effective incorporation of limestone into the soil is necessary, although this is not always feasible, depending on field conditions [31]. According to [32], the availability of soil moisture required for the chemical reaction of limestone is more critical than the time interval between liming and sowing.

Table 7 presents the estimated coefficients and standard errors for the intercept and four independent variables: PCorrection (soil correction with limestone); PAnlSolo (soil analysis); FCorSolo (performs soil correction during formation); PMacro (fertilization with N, P, K, or S); FAnalSolo (performs soil analysis in the formation).

Producers who adopt all four technological processes have a 94.18% probability of also performing soil acidity correction using limestone. The Hosmer-Lemeshow test yielded a statistic of 6.5306 with five degrees of freedom, which was not statistically significant (P=0.2580), indicating a good model fit. This result is consistent with the model’s performance shown in the ROC Curve (Figure 5), which presented an AUC value of 93.12%, indicating a very good level of discrimination. The overall model fit is further illustrated in Figure 6.

Figure 7 presents the induced decision tree, generated using 70% of the instances for training. The target attribute was PCorrection, and, as expected, PAnalSolo was selected as the root node, indicating it as the most important predictor. Since nutrient availability is influenced by soil pH, correcting soil acidity enhancing fertilizer use efficiency, which supports the observed association between PCorrection and PMacro (macronutrient fertilization). Interestingly, although PMacro was statistically significant in the binary logistic regression model, it was excluded from the J48 decision tree due to its low information gain. Information gain quantifies an attribute’s ability to partition the data into distinct categories. The attribute with the highest gain is selected for each node, with the root node representing the most informative variable [33].

The Kappa index and accuracy were calculated at 0.6971 and 85.05%, respectively. The Kappa value indicates a substantial level of agreement. The decision tree generated six classification rules (i.e. leaves). Table 8 presents the analysis of these six rules, highlighting the importance of alignment between the target attribute (PCorrection) and the strategies of conducting soil analysis (PAnlSolo) and implementing soil correction during pasture formation (FCorSolo). The rules exhibited high confidence levels (over 90%), which indicates strong consistency among the technological practices. This alignment is to some extent expected, considering that rural producers increasingly have access to technical information from diverse sources. Moreover, due to the high cost of soil correction with limestone, producers tend to adopt more precise and efficient practices to ensure the effectiveness of this input.

3.4. Rotation Grazing (PRotacio).

Over time, pasture specialists have analyzed grass defoliation patterns to better understand how the frequency, intensity, and uniformity of grazing affect the development of forage plants and grazing animals. Research in this area has documented the physiological responses of grasses to grazing, demonstrating that increased defoliation intensity or frequency generally leads to reductions in biomass and plant productivity, though the magnitude of these effects varies accross species.

According to [34], who conducted experiments within real production systems, the adoption of rotational grazing using multiple paddocks increased spatial and temporal heterogeneity of the grass, as well as the variation in defoliation frequency among paddocks. This approach enabled producers to improve grazing management in temperate pastures. In contrast, in a scenario and risk analysis of a rotational beef cattle system, [35] found that although profitability increased, the system was characterized by low attractiveness and a high probability of economic losses. These findings underscore that the effectiveness of rotational grazing depends on a correct understanding of the mechanisms of biomass production under a wide range of environmental conditions [36].

Table 9 presents the estimated coefficients and standard errors for the intercept and three contextual variables: Biome; PAnlSolo (soil analysis); and FMacro (fertilization during pasture formation).

The results presented in the table suggest that a producer whose property is located in the Atlantic Forest biome and who conducts both soil analysis and fertilization during pasture formation has a 78.74% probability of adopting rotational grazing as a herd management strategy. Conversely, a producer in the same biome who does not implement these practices has only a 35.78% likelihood of adopting rotational grazing. Figure 8 illustrates the probabilities of rotational grazing adoption across different biomes, according to whether or not other technological processes are employed.

It is noteworthy that the probability of adopting the rotational grazing technological process is particularly high in the Amazon and Atlantic Forest biomes (over 80% among producers who also adopt other technological practices). Even among producers who do not implement the related technologies, the probability remains above 40%. Pastures possess significant potential for atmospheric carbon sequestration, with efficient storage occurring in the soil through the development of grass root systems. Additionally, well-managed grazing can reduce the age at slaughter, thereby contributing to the mitigation of methane emissions. Furthermore, ongoing research aimed at enhancing technological efficiency in fire and deforestation prevention, as well as pasture restoration, is expected to generate substantial positive impacts. Such advances will further optimize and strengthen the role of rotational grazing systems in sustainable livestock production [37].

Figure 9 presents the ROC curve for the model analyzing the dependent variable Protacio (adoption of rotational grazing), with an estimated AUC of 73%, which indicates a good level of model discrimination. The Hosmer–Lemeshow test yielded a statistic of 4.797 with seven degrees of freedom and was not statistically significant (P = 0.6929), suggesting a good fit of the model to the data.

Figure 10 presents the induced decision tree, generated using 80% of the instances for training. The target attribute was PRotacio, with PAnlSolo selected as the root node, indicating it as the most important predictor. Although the Biome attribute was statistically significant in the logistic regression analysis, it was excluded from the decision tree by the J48 algorithm due to its low information gain, that is, it did not contribute meaningfully to partitioning the data into distinct categories. The Kappa index was estimated at 0.5015, indicating a moderate level of agreement. Overall, 77.46% of the instances were correctly classified.

Three leaves (rules) were determined, as presented in Table 10.

The adoption of PAnlSolo guides the entire decision-making process related to the adoption of rotational grazing. These findings are consistent with [37], who emphasizes the need for further research to better understand and harness the interactions between soil analysis, nitrogen fertilization, and defoliation intensity, particularly in tropical species. Such studies are essential for identifying the most efficient grazing system strategies, with the dual goals of maximizing productivity per area and minimizing negative environmental impacts.

4. Discussion

Soil analysis is essential for understanding soil health and its nutrient composition, which directly influences pasture productivity. Tecniques for soil evaluation help identify nutrient deficiencies and excesses, allowing for informed decisions regarding fertilization and soil amendments. Furthermore, soil carbon dynamics indicate that improved pasture management, originating from thorough soil assessment, can augment soil carbon stocks, thereby contributing to environmental sustainability [38]. Effective soil management practices, including liming, lead to increased soil fertility and, consequently, higher forage yield and quality, thereby supporting improved stocking rates [39]. Thus, soil analysis serves as the foundation for all subsequent pasture management practices. Effective stocking rate management helps maintain a balance between animal demand and pasture supply, thereby preventing overgrazing and pasture degradation.

The adverse effects of excessive stocking rates, which can lead to reduced biomass and biodiversity in pastures, while optimal rates can enhance overall productivity and sustainability [40]. Findings support that lower stocking rates generally yield better pasture variability and productivity, reinforcing the need for careful management [41].

Soil correction with lime is a necessary practice for addressing soil acidity, commonly prevalent in many pasturelands. The applying lime significantly elevated pasture production, emphasizing that soil pH plays a crucial role in nutrient availability and, consequently, plant growth [42]. Therefore, limestone application not only aids in ameliorating soil chemistry but also enhances ecological conditions needed for robust pasture growth.Rotational grazing can optimize pasture utilization and animal weight gain compared to continuous grazing systems [43]. The adjusting grazing intervals in rotational systems allows for greater forage accumulation, positively impacting animal growth rates and pasture health [44].

This practice aligns with point towards optimal animal gain being achievable at suitable stocking rates, further strengthened by the integration of rotational grazing methods. Implementing these strategies ensures not only the health of the pasture but also optimizes livestock performance and contributes to long-term agricultural resilience.

5. Conclusions

The combination of parametric (inferential) and non-parametric (non-inferential) methodologies represents a promising and practical analytical strategy. In this article, the integration of binary logistic regression and decision tree induction proved to be a potentially effective strategy. Nevertheless, further applications across different areas of agribusiness are needed to consolidate this strategy as a robust framework for knowledge discovery and extraction.