2023 Volume 11 Pages 93-105
Soil compaction is one of the most harmful elements affecting soil structure, limiting plant growth and agricultural productivity. It is crucial to assess the degree of soil penetration resistance to discover solutions to the harmful consequences of compaction. In order to obtain the appropriate value, using soil cone penetration requires time and labor-intensive measurements. Currently, satellite technologies, electronic measurement control systems, and computer software help to measure soil penetration resistance quickly and easily within the precision agriculture applications approach. The quantitative relationships between soil properties and the factors affecting their diversity contribute to digital soil mapping. Digital soil maps use machine learning algorithms to determine the above relationship. Algorithms include multiple linear regression (MLR), k-nearest neighbors (KNN), support vector regression (SVR), cubist, random forest (RF), and artificial neural networks (ANN). Machine learning made it possible to predict soil penetration resistance from huge sets of environmental data obtained from onboard sensors on satellites and other sources to produce digital soil maps based on classification and slope, but whose output must be verified if they are to be trusted. This review presents soil penetration resistance measurement systems, new technological developments in measurement systems, and the contribution of precision agriculture techniques and machine learning algorithms to soil penetration resistance measurement and prediction.
The structure of the plowed layer of cultivated soil varies over time due to tillage, soil compaction, and natural processes. Negative effects on soil conditions are owing to annual tillage in the absence of other soil management techniques [1]. By preventing topsoil loss due to wind erosion and preserving soil moisture during frequent summer droughts, conservation tillage can increase agricultural production. As a result of soil compaction caused by mechanized agriculture, the soil becomes fragmented, compacted, and displaced. The cumulative effects of these processes change how the blocks are arranged, how large they are, and how they are shaped, which changes how porous these units are both inside and between them [2]. The world’s agricultural sector must be able to utilize sufficient arable soil to feed the expanding population, which is expected to reach 9 billion by 2050 [3]. These data indicate that crop cultivation must be increased by 70% to increase the food processing sector. Although this will necessitate the use of heavy machines and equipment for harvesting crops and other agricultural tasks, it increases the risk of soil compaction, which results in soil deterioration, and negatively affects the 68 million hectares of arable land [4].
The physical, chemical， and biological makeup of the soil are adversely impacted by soil compaction, which also restricts plant growth and slows the rate of water intrusion, which lowers productivity. To address the harmful effects of compaction, it is crucial to ascertain the degree of soil resistance. High soil compaction is caused in agricultural fields by the instead of tractors, cultivators, and combines that are heavier now that they have more equipment attached to them. Other factors that contribute to soil compaction include plowing under unsuitable conditions. In addition to these outside factors, considerable soil compaction can also be caused by natural factors such as intense rainfall and drought [5, 6]. The specific weight and moisture of the soil can be used to describe soil compaction. Air spaces close up between soil particles as they get better to one another during soil compaction. As a result, soil compaction increases bulk density and resistance to soil penetration [7, 8]. Machines that operate on compacted soil also require more power [9, 10]. In addition to these negative agricultural and environmental effects, soil compaction can also be the root cause of surface water runoff, flooding, chemical leaching, and soil erosion. Influences the soil’s basic environmental processes, including its porosity (proportion of larger pores), air-holding capacity, water infiltration capacity, and hydraulic conductivity [11]. Also, crop yields drop by 40% to 90% [12].
This review aims to clarify the technological development in equipment and devices for measuring soil penetration resistance and the extent of the contribution of precision farming techniques in facilitating measurement operations and locating soil compaction, as well as the possibility of drawing digital soil maps through descriptive geostatistics and the possibility of predicting soil compaction based on machine learning algorithms.
Soil penetration resistance can be measured by measuring the resistance of the vertical soil, as this method is considered one of the simple methods [14]. Soil penetration resistance is measured using two methods: First, an open-ended tube is used to collect soil samples from the field at a specific depth, and the samples are then tested in the lab to determine the penetration resistance. The second method, a conical head of a particular size is lowered into the ground vertically and at a set depth, beginning at the soil’s surface [15]. The cone’s tip has a 30-degree angle, and the force needed to push the cone into the ground at a normal speed of 30 mm per second is divided by the cone’s base area to determine the resistance to soil penetration. The cone index (CI), which represents this estimated value of the cone, is expressed by the following equation [9].
in which, CI is cone index [MPa], F is force [N], and A is base area [mm2].
Cone penetrators are used because they are quick, economical, and simple to use when measuring soil penetration resistance. Since spatial heterogeneity has a significant impact on penetration resistance, accurate data collection is necessary to establish the relationship between soil penetration resistance and other characteristics [15]. There are also different types of soil resistance measuring devices: Proctor penetrometer (a), Penetrograph (b), Pocket penetrometer (c), Cone penetrometer with dial indicator (d), Penetrologger (e) (Fig.1) [16].
Figure 1: Types of cone penetrometers [16]
(a) proctor penetrometer, (b) penetrograph, (c) pocket penetrometer, (d) cone penetrometer with dial indicator, (e) penetrologger
However, penetration resistance measurement using cone penetrometers is difficult and time-consuming, especially on highly hard soils. Therefore, soil penetration measuring devices have been developed in order to take measurements accurately and quickly, as they are installed on tractors and generally use hydraulic power to penetrate the cone into the soil [17]. Randy L Raper et al. [18] created a hydraulically powered multi-probe conical soil penetration meter that can be mounted on a farm tractor and can be used to quickly and easily measure soil pressure (Fig.2).
Figure 2: Multi sensors soil penetration resistance meter [18]
Tekin and Okursoy [5] also developed a tractor hydraulic system penetration meter to measure soil penetration resistance (Fig.3). The developed penetrometer is installed on the hydraulic device of the tractor and the standard conical penetrometer is pushed into the soil at a constant speed by means of a hydraulic piston. Using the tractor’s hydraulic system, the cone penetration rate can be kept at a record value with the penetrometer, the analog data of the load, which will determine soil compaction, can be converted into digital forms and the user can access the data instantly from the computer in the cabin. In addition, the system offers great convenience to the user by easily obtaining data on soil compaction accurately, quickly, and effortlessly without leaving the tractor compartment.
Figure 3: A tractor hydraulic system drives a penetrometer [5]
With the development of technology, devices, and software appeared to facilitate agricultural operations and provide alternative solutions to problems. With the availability of methods, models, technological tools, and laptops, their use in field applications has increased. During the period of agricultural production development, there was rapid development in information technologies after mechanization, automation, control, and informatics. As a result, today intelligent machines and machine-controlled production systems have begun to replace traditional production methods [19, 20, 21]. These techniques have a place in the precision agriculture strategy, which guarantees that field performance can be tracked, planned, and evaluated down to the square meter level so farmers can know how well or badly each portion of the field is performing [22].
Precision agriculture is expressed using different terms around the world. These include site-specific farming, site-specific management, computer-aided farm, prescription farming, variable rate application (VRA), etc. Regardless of the term used for a precision agriculture system, the system connects the controllers, electronics, computer, database, and account information. Components of precision agriculture technology: Global Positioning Systems (GPS), Geographic Information systems (GIS), VRA, and Remote Sensing [23, 24, 25].
Variables in precision agriculture are divided into three sections: spatial, temporal, and economic. The variance must first be determined and then an actionable management decision taken. Accordingly, adapting and developing the right strategies and practices will ensure the success of precision agriculture. The sole purpose of precision agriculture is to increase production, but also to include practices that will allow for savings in the use of inputs without causing a loss of yield. In crop production management, different perspectives are put forward in order to understand and explain the physical and geographical variation of the land. In order to put these views into practice and achieve VRA, a decision support system is needed. In addition, sensing, monitoring, control, and data transmission systems are required technologies for precision agriculture applications [23].
It is simple for researchers to test soil parameters in an instant and dynamic method as a consequence of the development of technology in the field of precision agricultural applications [26]. Numerous studies highlight the need to accurately map soil parameters and assess soil resistance levels. Emphasis is concentrated on the development of horizontal penetrometers that can take real-time measurements while using GPS systems and field mapping activities to accomplish this. Additionally, studies show that the use of real-time measuring methods, tools, GPS systems, and soil maps of soil qualities play a significant role in boosting yields and enabling quicker and easier field operations [27]. In recent years, penetrometers (horizontal) have also been developed that can be used to continuously measure soil compaction. With this penetrometer, continuous reading can be done from different depths. The main elements of this type of penetrometer are the chassis, penetration force sensor, speed control, and depth sensor (Fig.4) [23].
Figure 4: Horizontal penetrometer and its components [23]
Sirjacobs et al. [28] developed a sensor that can measure soil compaction changes in real-time and can be connected to hydraulic system in tractor (Fig.5). A soil resistance map of the region was produced by the researchers by pulling the created sensor with a tractor fitted with a Differential Global Positioning Systems (DGPS) receiver at intervals of 5 m. When the sensor is pulled into the soil, the sensor gives different signals according to the resistance of soil penetration. These signals are converted into values that can be used to map the variance in soil penetration resistance. They also stated that the results obtained in field conditions show that the system is a useful innovation for precision agriculture, that allows the physical state of the soil to be characterized in real-time. In order to understand and quantify soil penetration resistance, soil compaction variance maps are produced.
Figure 5: General schematic diagram of ground resistance sensor (Fx)horizontal force; (Fz)vertical force; (My)moment [28]
Geostatistics has played an important role in changing the way that researchers think about soil diversity and soil mapping, and thus a greater understanding of soils. Mapping in this way gave rise to other problems, the solutions of which accelerated the development of spatial statistics [29]. The need for digital soil maps is growing due to three primary factors: (1) the lack of manpower and financial resources to develop soil maps, (2) the information these maps provide in real-world applications, and (3) the need for quantitative, repeatable results [30]. In traditional soil maps, the main information is taken from measurements made in the laboratory or at sampling sites. The search for ways to prepare digital soil maps from random data is very important because of the high cost and time required for soil sampling. This is why geostatistics has entered into digital soil mapping [30]. Where it is possible to predict the soil properties in unmonitored sites using statistical inference. Statistical models of the link between environmental factors and soil properties are frequently developed using digital soil maps. With the development of geostatistics, researchers are now able to precisely interpolate spatial patterns of soil attributes, making such maps more helpful [31].
Despite the fact that traditional maps rely on human vision, it is impossible to match the accuracy and efficiency of making maps based on the knowledge of knowledgeable soil surveyors [32, 33]. Compared with traditional soil maps, the quantitative approach of digital soil maps for spatial prediction also provides better spatial and objective consistency of information.
Many predictions or interpolation methods have been applied in the digital mapping of soil properties, especially statistical methods such as multiple linear regression [34, 35]. such as normal kriging [36], co-kriging [37], and hybrid techniques such as regression-kriging [38]. In geostatistics, the spatial coordinates of soil properties are used to predict values at unmapped sites [39]. The choice of a model for predicting soil characteristics depends on several factors, including the availability of soil data, the size and environmental characteristics of the area, the time of computer operation, the ease of implementing the model and interpreting the results, and the accuracy required for mapping [40]. In this sense, Minasny and Hartemink [41] tested several continuous variables prediction methods, based on criteria such as ease of use and prediction efficiency.
The important feature is that environmental variables help explain the spatial variability of soils. Incorporation of important external covariates into models can improve prediction accuracy, while unnecessary and redundant covariates would worsen the model performance. Thus, more field data must be collected if precision agricultural technology is to succeed. These operations require a lot of time and effort. However, with the ability to quickly and accurately measure data and map soil compaction, it will be possible to plow at different depths in certain areas of the region depending on the depth of soil compaction [42]. Thus, fuel, time, and labor are saved by plowing at different depths specific to the area. Using the penetrometer developed by [9], the soil penetration resistance values were mapped at a depth of 40 cm (Fig.6).
Figure 6: Soil resistance map [9]
With the help of penetration resistance maps, the parts of the field and the depth of plowing can be determined. In this type of map, the dark areas are the parts that show high soil compaction. Deep plowing of the entire field is carried out only in the area where the compaction is high. In this way, less energy is consumed (Fig.7) [23].
Figure 7: Penetration resistance map [23]
The methodology of quantitative soil extrapolation is very important because it aims to assess the similarity between known soils to other unknown soil types [43]. It may be able to extrapolate the data or model from the reference area, for instance, if one chosen area has extremely detailed soil maps and has soil-forming components similar to another area with few or no soil maps, these probabilities are combined to give predictions for soils outside the reference area [44]. Data extrapolation is a general concept, as it can be applied to variables other than soil traits. This accurate data is very important in digital mapping and precision agriculture [45]. However, its application to large areas is rare because it is difficult to repeat sampling at the same levels so that spatial mapping can be relied upon as a first step and not as a final result [46].
3.2 Machine learningThe advancement in remote sensing and the increase in the availability of environmental variables contribute to increasing the accuracy of prediction based on machine learning in digital soil mapping [47, 48]. Therefore, machine learning is a way to analyze data in order to learn and build a model to recognize data patterns. Most machine-learning approaches don’t need to be explicitly coded; they can operate with little to no human involvement [49]. This approach undoubtedly has important advantages, but it should be used with caution as it can be sensitive to overfitting and lacks transparency [50]. Several papers have studied different machine learning methods on different soil datasets during the past 10 years. In terms of digital soil map, we specifically compare the advantages and disadvantages of multiple linear regression (MLR), k-nearest neighbors (KNN), support vector regression (SVR), cubist, random forest (RF), and artificial neural networks (ANN). There are benchmarks to compare these Algorithms: hyperparameter quantity, sample size, covariate choice, learning duration, and final model interpretability. If the training time is a concern, the MLR, KNN, SVR, and Cubist algorithms should be considered because they contain less model parameters and hyper parameters. If data collection is extensive (thousands of samples) and calculation time is unimportant, results using ANN are likely to be the greatest [51]. Cubist, KNN, RF, and SVR are expected to perform better than ANN and MLR when there are fewer than 100 samples [52, 53]. Large data sets might not lessen how uncertain Cubist, KNN, RF, and SVR forecasts are. The best algorithms to utilize are Cubist, MLR, and RF when the user’s comprehension of the final form is crucial [54]. Table 1 shows the main factors for selecting a machine-learning algorithm for digital soil mapping [55].
| Algorithm | Quantity of hyper parameters | Not sensitive to data size | Covariate selection | Low computation time | Interpretability |
|---|---|---|---|---|---|
| MLR | 0 | X | X | √ | √ |
| KNN | 1 | √ | X | √ | X |
| SVR | 2 | √ | X | √ | X |
| Cubist | 2 | √ | √ | √ | √ |
| RF | 3 | √ | √ | X | X† |
| ANN | 8 | X | X | X | X |
* The checkmark (√) indicates that the algorithm has the capacity to apply covariate selection to decrease the number of variables that need to be processed by the algorithm, whereas the checkmark (X) indicates that ML methods use all covariates.
† RF is not entirely unintelligible. The model that RF created is a semi-black box.
A generalized model with predictive power outside of the domain in which it was trained is the ultimate, ideal aim. Because of its complexity, this idea might not be as useful when the research region is big. For instance, it would be helpful to know if a digital soil map of one area may be applied to another neighboring or nearby field. To complete this level of model evaluation, a second independent collection of points outside of the training dataset’s geographic range would be kept in reserve [54].
A lot of research has been done to predict soil penetration resistance using different machine learning algorithms. Zhu et al. [55] use the SVR algorithm to predict the soil shear resistance (cohesion and internal friction angle). While Rauter and Tschuchnigg [56] evaluated the machine learning model on soil classification from cone penetration tests using the RF algorithm. Abrougui et al. [57] evaluated the ANN model’s performance using a single neural network architecture similar to that shown in Figure 8. It is possible to forecast soil penetration resistance using input data including the tillage system, date, soil bulk density, and water content. The prediction performance of each model is evaluated using the root mean square error (RMSE). The closer the RMSE to 0, the greater the accuracy of the estimate, the lower the error, and the better the model works [58].
Figure 8: ANN topology test [57]
Prediction results in lack of transparency, so their output must be verified if they are to be trusted. However, there is a limitation when applied to spatial data, which spatial statisticians can help overcome. Spatial statisticians in rudimentary measurements still have much to do to incorporate uncertainty into numerical predictions, averages, and spatial totals across regions and to account for measurement errors and spatial positions of sample data. They must also communicate their understanding of these uncertainties to the end users of the soil maps, by whatever means they are made.
Precision agriculture practices show a rapid development trend as a new technological step in agriculture. Using satellite technologies, electronic measurement, and control systems, and computer software, accurate and practical data suitable for precision agriculture techniques are obtained. It is clear that systems developed to obtain soil penetration resistance are effective in protecting soil and water resources. Depending on the quantitative relationship between soil properties and the factors influencing them, numerical variance maps of soil penetration resistance can be made. To evaluate the performance of the model and the resulting map, three steps must be followed: In the first step, the robustness of the model must be tested by training on the initial dataset. In the second step, an independent validation within the same geographic range is required to test the predictive power of the model. The third and final step is to conduct another independent validation test to assess the portability of the model outside the area in which the model is being trained. Due to the development of machine learning algorithm, time and money spent on experimental implementation are now saved throughout the estimating phase. Algorithms include MLR, KNN, SVR, cubist, RF, and ANN. Prediction of soil conditions from environmental data are sources for producing digital soil maps based on classification and slope, but their output must be verified if they are to be trusted. Therefore, spatial mapping can be relied upon as a first step and not as a final result because data extrapolation is not guaranteed.
