Real-time risk assessment framework for tractors: part 1
— Risk point estimation using machine learning based on accident factors —
2026 年 19 巻 3 号 p. 144-152
詳細
2026 年 19 巻 3 号 p. 144-152
Tractor rollovers remain a critical safety issue. We propose a risk assessment framework (Part 1) that estimates quantitative risk points (RP) from accident-factor variables available during or prior to operation. Using 1,164 de-identified tractor accident records from Japan and nine accessible variables, we trained a deep neural network regressor. The optimised model (five hidden layers) achieved r = 0.87 and R2 = 0.71, with errors decreasing as RP increased. This model outperformed the baseline models in accuracy. The framework estimates potential risk points of the tractor operational environment and generates a numerical index to inform operators of the prevailing risk level, enabling proactive safety management. Part 2 incorporates a dynamic risk index derived from vehicle-behaviour signals to provide a comprehensive assessment.
Agricultural accidents remain a serious global concern. According to the International Labour Organization (ILO) (2015), at least 210,000 agricultural workers die annually from work-related incidents, representing a fatality risk thrice higher than that in other industries. Millions more sustain severe injuries, the majority involving agricultural machinery. In the United States (US), the Bureau of Labour Statistics (BLS) (2023) recorded a 2023 fatality rate of 20.3 deaths per 100,000 full-time workers in agriculture, forestry and fisheries—the highest among all industries. In European Union (EU) countries, the 2020 agricultural fatality rate was 5.0 per 100,000 workers, ranking second to the construction sector (Schmitz-Felten, 2023).
Among machinery-related incidents, tractor accidents are frequent and severe. In Spain, Azparren et al. (2010) analysed 388 fatal agricultural accidents between 2004 and 2008 and found that 272 (approximately 70 %) involved tractor rollovers. In India, Khadatkar et al. (2022) conducted a survey across 1,703 villages, revealing that approximately 36.2 % of fatalities were caused by farm machinery, including tractors. In the US, Gilblom et al. (2023) reported that of 292 patients treated for agricultural injuries at Sanford Medical Centre Fargo between 2010 and 2020, 106 cases (36 %) were machinery-related, with tractors representing the largest share (34 cases). These patients exhibited higher injury severity scores (ISS) and longer hospital stays than those with non-machinery-related injuries.
Japan is no exception, with a 2023 report by the Ministry of Agriculture, Forestry and Fisheries (MAFF) (2023) reporting that 236 fatalities were linked to farming machinery, 61 of which involved tractors. Tractor-related deaths accounted for 25.8 % of all machinery-related fatalities—the highest proportion among equipment types.
The aforementioned statistics highlight the severity of tractor accidents. Beyond causing injury or death, such accidents can result in considerable economic losses.
1.1. Causes of tractor accidentsA major factor contributing to tractor accidents is their mechanical design, as high centres of gravity and limited suspension make them unstable on uneven terrain.
The ageing agricultural workforce, notably in developed countries, also plays a crucial role. A 2025 briefing by the European Parliament reported that 57 % of EU farm managers are over 55 years old, while only 12 % are under 40 (EP, 2025). In Japan, the average age of farmers reached 69.2 in 2024 (MAFF, 2023). This demographic trend increases risk, as older individuals often experience reduced cognitive and motor functions, thereby limiting their ability to detect and respond to hazards.
Previous studies have linked ageing to reduced driving performance. Horswill et al. (2008) found that older drivers exhibited longer hazard perception response times. Huisingh et al. (2018) reported that the injury and fatality risks in motor vehicle crashes increase with age, despite driving helping older adults to maintain their independence.
The Food and Agriculture Organization (FAO) (2024) also identified rural ageing as a global issue. In its recent documentation, it emphasised the declining physical ability and social marginalisation of elderly agricultural workers and highlighted the need for inclusive rural development.
1.2. Limitations of current active safety technologiesActive safety technologies are crucial for preventing accidents. Currently, the most common safety feature in tractors is a tilt-angle detection device that triggers an audible alarm when the vehicle exceeds a defined tilt threshold. However, these systems are not universally installed, lack international regulatory standards, and are not mandatory—even under the national safety guidelines of Japan. Additionally, rollovers often occur rapidly, leaving little time for response once the alarm is triggered, thereby limiting the effectiveness of the devices.
More advanced technologies, such as AI camera-based obstacle detection systems, have been integrated into autonomous tractors (John Deere’s Autonomous 8R [Hartmann, 2025] and AGCO’s Unlimited VIEW [AGCO, n.d.]). Although promising, these systems are expensive and have yet to be widely adopted, particularly among small-scale farmers.
1.3. Need for new safety systemGiven the aforementioned challenges, existing safety technologies remain insufficient for addressing the evolving demographic and operational conditions in agriculture. Hence, a more intuitive and accessible risk notification system is required.
Studies have shown that humans respond effectively to visual information than to auditory cues (Alyamani et al., 2024; Cobus et al., 2019). Hence, complementing sound-based alarms with a system that continuously presents quantitative visual risk information would enable operators to recognise danger during routine and emergency situations. Unlike conventional alarms, which trigger only in critical moments and often require interpretation, a real-time risk visualisation system supports continuous situational awareness.
This study aimed to develop a real-time, visually interpretable risk notification system for tractors using quantifiable accident indicators derived from historical data. Based on a risk assessment index (risk points) introduced in previous studies, the system is designed to help operators to identify and avoid hazards before accidents. By presenting risk levels numerically, the proposed approach enables proactive decision-making, notably for elderly or vulnerable users in modern agricultural settings.
1.4. Related work: risk points as quantitative risk indexRisk points are quantitative indicators introduced by Aoyagi et al. (2019) to assess the severity of tractor-related accidents. The index is calculated using accident factors and their occurrence probabilities, including injury severity scores derived from 72 documented tractor accidents across 18 regions in Japan between 2001 and 2015.
Although the aforementioned concept has been investigated in previous studies, no prior study has integrated it into a system for accident prevention. Accident factors were categorised into three groups—human, environmental and machinery—each comprising 40 detailed sub-factors. These factors represent common contributors to accidents and are consistently observed across multiple cases. Effective implementation of real-time risk assessment using risk points would enable operators to identify hazardous conditions using numerical indicators. Because agricultural tractor operation typically involves long working hours and repetitive seasonal tasks, operators may have difficulty perceiving gradual changes in daily temperature or humidity, as well as long-term declines in physical condition. Accordingly, accidents may be prevented by informing operators of the current potential risk, expressed as a risk point. When combined with auditory or visual alerts, this approach could enable the timely avoidance of dangerous situations.
Consequently, systematising and applying risk points in practical, real-time settings holds promise for improving human safety and reducing accidents during agricultural operations. A real-time risk assessment based on dynamic tractor data enables direct accident prevention by predicting rollover accidents, which account for a large proportion of serious incidents. This approach is reported in Part 2 of this study.
Figure 1 shows an overview of the proposed system framework, which evaluates risk points using regression and classification models.
The regression model estimates risk points based on accident factors, while the classification model assesses risk points based on tractor behaviour. The final output integrates both sets of risk points to provide an objective risk assessment. Part 1 of this study describes the regression model, while Part 2 the classification model. Risk points are expressed using Eqs. (1) and (2) with RP units as follows.
| (1) |
| (2) |
Aoyagi et al. (2019) evaluated and defined accident risk using a collection of tractor accident cases. Equation (1) expresses the risk associated with an accident factor as the product of its occurrence probability, pj, and the corresponding degree of injury, ej. The occurrence probability of each accident factor, pj, is derived using fault tree analysis, in which the top event is defined as the probability of tractor accidents in Japan, and the basic events—referred to as accident detail factors—are structured using AND/OR logic. The probability of tractor accidents in Japan is calculated by multiplying the annual number of fatal tractor accidents by five to account for injury cases, and then dividing the result by the total number of tractors in use. The degree of injury associated with each accident factor, ej, is calculated using Eq. (2). For each accident case, an injury severity score based on the level of damage is assigned. The scores for cases in which a given accident factor is present are then summed and divided by the total number of relevant cases, n. This procedure provides a probabilistic estimate of the expected severity of damage attributable to each accident factor. Thus, the risk point serves as a risk assessment index that statistically evaluates the hazard associated with tractor accidents. In addition, accident detail factors are characterised as situational conditions, such as ‘narrow working spaces’ or ‘deterioration of protective equipment’. These factors represent inherent risks within the working environment and contribute indirectly to accident occurrence. Therefore, although recognising the risk point does not in itself prevent accidents, it can support prevention efforts by prompting operators to take appropriate actions, such as taking breaks according to the perceived level of risk or shifting to tasks with lower physical demands.
The regression model (Fig. 2) systematises risk point calculation by converting detailed accident factors into numerical data as measurable explanatory variables. It considers the likelihood of occurrence and severity of injury to determine risk points. The risk points for case Ri serve as the response variable, denoted as ‘risk points based on accident factors’. By modelling risk point generation, real time calculation becomes possible—something otherwise unachievable after accidents.
The explanatory variables used in the regression model are obtained by converting numerical data that reinterpret selected factors from the set of 40 into representative values. These factors were selected based on two criteria: (1) they can be measured in real time during tractor operation or predefined in advance, with factors that cannot be directly measured described using proxy variables; and (2) they are associated with high risk points. With this selection, risk points conventionally calculated from a wide variety of factors can be estimated using only representative values of the selected factors.
The descriptions of each factor and their numerical conversions are as follows:
1) High age: Applies when the accident involves an operator over 70 years old. The numerical value represents the age of the operator as recorded in the accident record (e.g., age or date of birth entered prior to tractor operation).
2) Anxiety: Applies when the operator is anxious before the accident, as recorded in interviews. It is numerically represented by two variables: the ‘remaining time until sunset (h)’ and ‘remaining time until estimated precipitation (h)’, thereby reflecting common concerns about approaching darkness or forecasts of rain or snow.
Sunset times were obtained from historical data published by the National Astronomical Observatory of Japan (NAOJ) (n.d.) based on the date, region, and time of the accident. Past meteorological data from the Japan Meteorological Agency (JMA) (n.d.) were used to determine the timing of rain or snow, while the differences between the accident and sunset or precipitation on the same day were used as the numerical values for these variables (e.g., calculating the remaining time at the point of operation based on online meteorological information and the sunset time on the day of operation).
3) Bad temperature and humidity conditions: Applies when extreme heat/cold, including high humidity/dryness, caused physical strain or cognitive judgment impairment of the worker. As in 2), the temperature (°C) and humidity (%) at the time of the accident were obtained by referencing weather data from the JMA based on accident records (e.g., measuring ambient temperature and humidity using a thermo-hygrometer).
4) Bad precipitation conditions: Applies when rain, snow or strong winds resulted in slippery roads or reduced visibility, thereby leading to an accident. Precipitation (mm) and wind speed (m/s) at that time were obtained from historical JMA data (e.g., obtaining data from a wind speed sensor, a rain gauge, or publicly available real-time meteorological information).
5) Insufficient brightness: Applies to accidents caused by poor lighting conditions in the work environment. Using the sunset time (2)), a score of 0–3 was assigned based on the elapsed time after sunset: 0 (before sunset), 1 (within 30 min), 2 (between 30 and 60 min) and 3 (after 60 min). These time bands correspond with ‘civil’, ‘nautical’, and ‘astronomical twilight’, reflecting progressively diminishing natural light as defined in astronomy. For example, a twilight-related index (e.g., elapsed time after sunset) can be calculated based on the sunset time on the day of operation and the current time.
6) Inclination, steps, and location unevenness: Applies to accidents caused by slopes, steps or surface irregularities (e.g., obtaining the inclination angle of the machine body and vibrational acceleration from an inertial measurement unit (IMU) with six-axis acceleration and angular velocity sensors).
7) Unsafe or inoperable machine: Applies when tractors (outdated models) had safety or operability problems (e.g., entering the manufacturing year [or model year] in advance and assigning a binary value when the elapsed years exceed a predefined threshold).
8) Poor maintenance or unsafe equipment condition: Applies to accidents caused by inappropriate tractor maintenance (e.g., entering the date of the most recent maintenance in advance and assigning a binary value when the elapsed days exceed a predefined threshold).
9) Not equipped with rollover protective structure (ROPS): Applies to accidents involving tractors lacking a safety frame or cab. Although not considered in previous studies as an accident factor, it was included in this study owing to its considerable impact on accident severity (Tomita et al., 2009).
Variables 6)–9) were rated on a 0–4 scale to indicate the ‘danger level of the tractor’, with 6) assessing the inclination of the tractor and 7)–9) evaluating the inherent risks associated with the tractor (e.g., entering the presence or absence of ROPS in advance as a binary (0/1) variable).
Variables 2), 7), and 8) are quantified as proxy variables because they are difficult to measure directly. These variables are derived from corresponding factors documented in accident case records, such as statements indicating operator anxiety due to approaching rainfall or reduced operability associated with older, second-hand tractors. It should be noted that the aim of this paper is to propose a conceptual framework rather than to address its implementation as an operational system.
3.2. Regression model designAcross prior studies on agricultural machinery and on-road vehicles, deep neural networks (DNNs) have achieved higher estimation accuracy than other learning algorithms numerical regression tasks (Badgujar et al., 2022; Jalilnezhad et al., 2023; Lee et al., 2025; Rahimi-Ajdadi et al., 2011). Hence, deep learning was employed for the regression analysis using a DNN model. Numerous conditions, referred to as hyperparameters, were configured to optimise model performance.
3.2.1. Number of learning cyclesThe number of cycles required for the output error of the model to converge was assessed. Based on empirical estimation, convergence was expected between 800 and 1,000 cycles. Hence, all models were trained for up to 1,000 cycles.
3.2.2. Validation methodA ten-fold cross-validation method was employed by dividing the entire dataset into ten subsets, with nine-tenths assigned as the training/validation set and one-tenth reserved as the test set. Within the training/validation set, ten-fold cross-validation was further performed, with nine-tenths used for training and one-tenth used for validation. For the model structure achieving the highest estimation accuracy, independent tests were conducted across all four parameter sets, and the parameter set yielding the highest accuracy was selected as the final regression model configuration.
3.2.3. Output layerIdentity and absolute functions were used in the output layer. The identity function is generally employed in numerical regression DNN models, while the absolute function was added during estimation tests to convert any negative outputs from the identity function into positive values, as the response variable—risk points—is always positive. The absolute function was not applied during back-propagation of learning phase.
3.2.4. MiscellaneousLeaky ReLU was employed as the activation function, the squared error function to measure error, and Adam optimiser to update the weight coefficients.
Based on the aforementioned design conditions, numerous combinations of hidden layers and nodes were prepared for training and validation, while the model accuracy was evaluated to identify the optimal DNN structure and parameters. The input layer comprised nine nodes, corresponding to the number of variables in the training data. Regarding the hidden layers, the number of nodes was set as multiples of 9 (1, 9, 18, 27, 36, 45, 54, 63, and 72), with the number of layers ranging from 1–10.
Starting with one hidden layer, the number of nodes was adjusted to identify the configuration with the highest estimation accuracy. The process was then repeated as additional layers were added, with the number adjusted at each step.
3.3. Training data generationThe training data was created from 1,164 tractor accident cases across 44 prefectures in Japan (2012–2017), provided by the JA Mutual Insurance Federation (Zenkyoren), as presented in Table 1. Risk points based on accident factors, used as target output values, were calculated for each case using Eqs. (1) and (2). The converted numerical data representing detailed accident factors, along with the corresponding pre-calculated risk points, were compiled into a dataset for deep learning training.
| Explanatory variables | Minimum value | Maximum value | Mean | Standard deviation | Related factors |
|---|---|---|---|---|---|
| Age (years old) | 18 | 90 | 59.8 | 16.3 | 1) |
| Precipitation (mm) | 0 | 7 | 0.12 | 0.37 | 4) |
| Time difference between precipitation and accident (h) | 0 | 24 | 13.6 | 10.9 | 2) |
| Time difference between sunset and accident (h) | 0 | 16.4 | 5.46 | 3.5 | 2) |
| Temperature (°C) | −7.3 | 34.9 | 18.7 | 9.2 | 3) |
| Humidity (%) | 12 | 100 | 63.8 | 16.8 | 3) |
| Wind speed (m/s) | 0.10 | 20.3 | 3.6 | 2.2 | 4) |
| Level of twilight (points) | 0 | 4 | 0.29 | 0.99 | 5) |
| Danger level of tractor (points) | 0 | 5 | 0.72 | 1.28 | 6),7),8),9) |
| Risk point (target variable) (RP) | 0 | 0.25 | 0.012 | 0.040 | – |
Table 1 presents the generated training data statistics.
Japan comprises 47 prefectures, and the dataset used in this study represents 93.6% of all regions. Further, because farming follows an annual cycle, data spanning six years provide a sufficient period. Although previous studies (Aoyagi et.al 2019) used only 72 detailed tractor cases, this study expands it to 1,164 cases, ensuring an adequate sample size. Hence, the regression model developed using the dataset is expected to perform effectively for real-time risk assessment.
The model achieving the highest accuracy, determined via training and validation, comprises five hidden layers with a node configuration of 63-54-36-63-18 from input to output. Figure 3 shows the output results of this model, wherein the horizontal axis represents the accident case numbers arranged in ascending order of the risk points, while the vertical axis represents the risk points transformed by the sixth root to compress the dynamic range.
The estimated output achieved correlation coefficient r and determination coefficient R2 of 0.47 and 0.68, respectively. The average absolute error across all outputs was 0.011 RP, which is lower than the average target risk point of 0.015 RP. This regression model captured more accurately the upward trend of target values within the range of 10−2 to 10−1 RP than under 10−3 RP. This performance likely stems from the model using only detailed accident factors associated with high risk points as explanatory variables. Because factors with low risk points were excluded, the model struggled to reproduce target values below 10−3 RP.
4.1.2. Independent test resultsTable 2 presents the results of the independent tests conducted using ten weight coefficient patterns.
| R2 | r | MAE | P95 (Absolute error) | |
|---|---|---|---|---|
| Validation | 0.47 | 0.68 | 0.011 | 0.072 |
| Pattern 1 | 0.63 | 0.83 | 0.012 | 0.076 |
| Pattern 2 | 0.64 | 0.86 | 0.011 | 0.069 |
| Pattern 3 | 0.71 | 0.87 | 0.011 | 0.063 |
| Pattern 4 | 0.67 | 0.84 | 0.011 | 0.072 |
| Pattern 5 | 0.67 | 0.86 | 0.011 | 0.070 |
| Pattern 6 | 0.66 | 0.84 | 0.011 | 0.058 |
| Pattern 7 | 0.64 | 0.82 | 0.012 | 0.071 |
| Pattern 8 | 0.65 | 0.84 | 0.011 | 0.067 |
| Pattern 9 | 0.63 | 0.82 | 0.011 | 0.061 |
| Pattern 10 | 0.67 | 0.84 | 0.011 | 0.061 |
The mean absolute error (MAE) was nearly identical across patterns; however, r and R2 improved for all patterns compared with the validation phase. Pattern 3 yielded the lowest MAE and the highest R2 and r. In contrast, Pattern 4 achieved the lowest 95th percentile of MAE (P95), at 0.058. The standard deviation of P95 across all coefficient patterns was sufficiently small compared with the other patterns (0.0055), and the difference between Pattern 3 and Pattern 6 was 0.0053. Therefore, Pattern 3, which performed best on the other evaluation indices, was selected as the optimal weight parameter set for the DNN.
Figure 4 presents a scatter plot of the independent test results for the DNN model using weight parameter Pattern 3.
As a baseline model for comparison, multiple linear regression (MLR), Gaussian process regression (GPR) and Least-squares boosting (LSBoost) were performed. The resulting MLR equation is given in Eq. (3).
| (3) |
Table 3 presents the results of the independent tests for each baseline model after training and validation. For an accurate comparison, the same training and test data were used for each model. Based on these results, the DNN model demonstrated superior regression performance compared with the other baseline models.
| R2 | r | MAE | P95 (MAE) | |
|---|---|---|---|---|
| GPR | 0.60 | 0.80 | 0.013 | 0.075 |
| LSBoost | 0.62 | 0.79 | 0.012 | 0.077 |
| MLR | 0.48 | 0.67 | 0.016 | 0.091 |
Correlation coefficient is sensitive to outliers, while determination coefficient tends to increase with more explanatory variables. These factors cannot be dismissed given the insignificant risk points exceeding 10−1 RP in the training data and use of nine explanatory variables. To assess whether Pattern 3 is optimal, the numerical metrics and estimation accuracy of each weight coefficient pattern were evaluated against the key requirements of the system.
Two criteria were evaluated: the (1) output accuracy of 10−2 RP or more and (2) accuracy of the output error digits.
4.2.1. Output accuracy of target values over 10-2 RPRisk Points based on accident factors ranged from 10−6–10−1 RP, spanning between the lowest 0 RP and highest 10−1 RP.
Figure 5 shows the percentage of accident cases within each risk point range, categorised by injury severity. Approximately 60 % of accidents below 10−6 RP involved minor or no injuries, a proportion that remains similar at 10−3 RP. For cases between 10−2 and 10−1 RP, it decreases to approximately 40 %. A similar trend is observed for fatal cases: the percentage remains relatively low (a few percent up to 10 %) but increases sharply to 30 % between 10−2 and 10−1 RP. Medium-to-severe injuries exhibit a steady rate of 30–40 % across all risk point ranges. This indicates that injury severity is stable from ‘below 10−6 RP’ to ‘10−3RP’, while fatal and minor or no-injury cases change notably between 10−2 to 10−1 RP. Hence, the regression model must accurately estimate risk points over 10−2 RP (after case number 800), wherein injury severity intensifies.
Rollover and falling accidents occurred more frequently in cases with high-risk points (Table 4).
| Digits of risk points (RP) | Number of accident cases | Of which: rollover/falling accidents (cases [%]) |
|---|---|---|
| Under 10−5 | 388 | 120 (31) |
| 10−4 | 395 | 62 (16) |
| 10−3 | 191 | 68 (36) |
| 10−2 | 137 | 86 (63) |
| 10−1 | 53 | 52 (98) |
The occurrence rate of rollover/falling accidents was 63 % (86 out of 137 cases) at 10−2 RP and 98 % (52 out of 53 cases) for 10−1 RP. Overall, cases with risk points above 10−2 accounted for 73 % (138 out of 190 cases), representing the majority of accident situations. In such cases, tractor operators often survived, frequently sustaining permanent injuries and psychological trauma, coupled with considerable economic losses from tractor damage.
Given the aforementioned, the estimation accuracy for target values above 10-2 RP was used as a key evaluation criterion, as it reflects the severity of physical injuries and associated psychological and economic impacts.
Table 5 presents the output accuracy for target values over 10−2 RP. In the test dataset, only 15 samples had RP values of 10−2 or higher. Because the correlation coefficient and the coefficient of determination can be overly influenced by a small number of outliers under such a limited sample size, they may lack reliability; therefore, these metrics were excluded from the evaluation at this stage. Pattern 3 achieved the lowest MAE.
| MAE | |
|---|---|
| Pattern 1 | 0.057 |
| Pattern 2 | 0.061 |
| Pattern 3 | 0.053 |
| Pattern 4 | 0.054 |
| Pattern 5 | 0.057 |
| Pattern 6 | 0.058 |
| Pattern 7 | 0.058 |
| Pattern 8 | 0.057 |
| Pattern 9 | 0.058 |
| Pattern 10 | 0.054 |
The mean risk point was 0.10 for the target values and 0.063 for the estimated values, while the median risk point was 0.036 for the target values and 0.062 for the estimated values. The difference in the mean was 0.037, and the difference in the median was 0.028. Considering that the risk point spanned a wide range from 10−6 to 10−1, both differences can be regarded as nearly the same order of magnitude. On the other hand, as is evident from the scatter plot for values of 10−2 RP or higher (Fig. 6), outliers were observed in data points No. 9 to No. 13, which likely lowered the mean of the estimated values. In particular, No. 13 showed a large error. Overall, as shown in Fig. 6, the general trend of the target values was reproduced, although some underestimation remains and should be improved in future work.
Risk points based on accident factors span a wide range of magnitudes: 10−6–10−1 RP. Consequently, the exponential part (the power of 10) serves a key indicator than the integer part. If the model accurately estimates the exponential component, errors in the integer part are considered negligible. Hence, an output error of a similar order of magnitude as—or less than—the target risk point is considered an excellent estimation.
Figure 7 shows the estimation output error results for parameter Pattern 3. The output error is defined as the absolute difference between the target and estimated values. For target risk points of 10−4 RP or lower, the error remained consistently within the 10−4 to 10−3 RP. As the target value increased, the error decreased. Although an outlier appears near case number 900 for target values of 10−3 RP or higher, the errors for all other cases were equal to or smaller than their corresponding target values.
Except for the low risk points of 10−4 RP and below, the error and target values exhibited a similar distribution trend in the 10−3 RP and above range, indicating that the errors were of a similar order of magnitude to the target values. For elevated risk points, the output error was negligible (less than or equal to the order of the target risk points).
To summarise (1) and (2), Pattern 3 yielded high estimation accuracy, as indicated in its r and R2 for the overall estimation test and target values over 10−2 RP. Although a certain degree of output error was observed for low-risk point target values, the error decreased as the target value increased, while the output error for high-risk points remained within acceptable and negligible limits. Hence, Pattern 3 was selected as the optimal weight coefficient parameter.
4.3. Validity of regression modelThe regression model using coefficient Pattern 3 demonstrated high accuracy, achieving r and R2 of 0.87 and 0.71, respectively. Herein, the output error decreased as the risk point value increased.
In the low to medium risk point range (below 10−6–10−3 RP), the output values were concentrated around 10−4–10−3 RP, thereby resulting in large errors. Within this range, minor and no injuries were most frequent, followed by moderate to severe injuries, with fatal cases being rare. From a damage severity viewpoint, outputting similar risk points within 10−4–10−3 RP is not considered a serious issue for system operation. The objective of this paper (Part 1) is not to predict dynamic behaviour immediately prior to an accident, but rather to monitor and estimate potential risk based on situational conditions. Therefore, as shown in Fig. 5, the fact that the estimated risk points do not vary continuously with high resolution in the low-risk range—where the distribution of damage severity is nearly uniform—does not necessarily impair operational decision-making. However, if an application requires detailed risk assessment within the low-risk range, the use of the proposed method should be considered with caution. Based on the aforementioned, the developed regression model is considered valid.
In this study, a regression model was developed to estimate tractor risk points in real time using deep learning, based on measurable data collected during tractor operation. The model was trained on a dataset generated from recorded tractor accidents. Key accident factors were identified and converted into numerical data, enabling the model to accurately estimate risk points from these factors with reasonable accuracy.
The regression model (DNN) comprised five hidden layers with a node configuration of 63-54-36-63-18, and its weight coefficients were optimised over 1,000 training and validation cycles. Using the optimal parameter set, the model achieved high accuracy, with r and R2 of 0.87 and 0.71, respectively. MAE was 0.011, although the output error was large for low target risk points, it decreased as the target value increased, indicating the inclusion of only high-risk accident factors as explanatory variables and exclusion of low-risk factors. Considering the relationship between injury severity and risk point levels, the variability in outputs for low to medium-risk values is acceptable for system operation. Hence, the regression model is deemed valid as a framework for estimating risk points based on internal validation. The implementation performance of this framework as a system should be verified in future studies.
However, because dynamic behaviour data are not incorporated, additional measures are required for decisive prevention of serious accidents. Following this study, Part 2 of the research develops a risk assessment and rollover estimation model using tractor behaviour data.
The reproducibility and validity of the proxy variables used to represent latent constructs (e.g., psychological states) were not evaluated in this study. This issue should therefore be examined in future implementation-oriented research.
6.2. Accident casesThis study could not verify whether the interview and recording processes were conducted using a standardised evaluation rubric or a clearly defined rater protocol, nor to what extent inter-rater agreement was ensured. As a result, the possibility of label noise in the training data cannot be entirely excluded, which represents a limitation of the data generation process. In future studies aimed at practical implementation, the standardisation of recording procedures (e.g., rubric and protocol design) and the validation of inter-rater agreement using a subset of the data should be considered. In addition, validation using non-accident operational data and time-at-risk data will be necessary at a later stage of implementation-oriented research.
6.3. External validationIn this study, the proposed framework was not evaluated using an independent external dataset. Large-scale tractor accident investigation data are not systematically collected by public institutions; therefore, such datasets are difficult to obtain. The dataset used in this study, comprising more than 1,000 accident cases, represents one of the largest available datasets in Japan. Nevertheless, the absence of independent external validation is recognised as a limitation of the present study.
6.4. Evaluation of estimations in the ≤10−3 RP rangeAs shown in Fig. 5, for the explanatory variables used in this study, the composition ratios of damage severity levels (e.g., no injury, minor injury) do not change substantially within the low-risk range up to approximately 10−3 RP. This range can therefore be interpreted as corresponding to similar levels of accident outcome severity. On the other hand, increases in the estimated risk value can still be observed within this low-risk range. This behaviour is attributable to the influence of combinations of multiple accident factors, particularly the inclusion of factors with relatively high occurrence probabilities. However, the proposed estimation model incorporates only nine accident factors associated with higher risk as input variables. As a result, increases in occurrence probability arising from variations in accident factors outside these nine may not be fully captured. Consequently, the model may not continuously reflect fine-grained variations in risk within the low-risk range.
This study is a retrospective analysis based on past accident-case data, and non-accident data under the same location and date conditions as those at the time of the accidents were not available. Therefore, external validation incorporating non-accident data could not be conducted.
Accordingly, a one-class SVM (support vector machine) was additionally constructed to learn the inclusion region using only accident data.
Here, the underlying population is assumed to be divisible into accident and non-accident data. This assumption is introduced solely to model the accident-region boundary under the constraint that only accident data are available.
Using the nine explanatory variables from all accident data (combined training and test datasets), the inclusion region was estimated with a Gaussian kernel.
As a result of training, the SVM decision function was obtained, as in Eq. (4).
| (4) |
The coverage rate for the accident data was 98.0 %, indicating that most samples are included within the model region, whereas the remaining 2.0 % could be judged as non-accident. The obtained support vectors and their weight coefficients αi are shown in Table S1 in the Appendix.
The authors express their sincere gratitude to the staff of Zenkyoren for providing access to the accident case data, and Mr. Agbor Mbi for his English language editing.
| Symbol | Description |
|---|---|
| α i | Weight coefficient of the SVM model |
| e j | Degree of injury with each accident factor (point) |
| E i | Degree of injury with each accident case (point) |
| i | Index number of accident cases (–) |
| j | Index number of accident factors (–) |
| K | Kernel of SVM model |
| n | Number of accident cases corresponding to each accident factor (case) |
| p j | Probability of each accident factor (cases/[unit year]) |
| r i | Risk point of accident factor (RP) |
| x age | Age (years old) |
| x danger | Danger level of tractor (points) |
| x humid | Humidity (%) |
| x i | A given set of explanatory-variable data |
| x prec | Precipitation (mm) |
| x T_prec | Time difference between precipitation and accident (h) |
| x T_sun | Time difference between sunset and accident (h) |
| x temp | Temperature (°C) |
| x twilight | Level of twilight (points) |
| x wind | Wind speed (m/s) |
| y rp | Risk point (outputs of multiple linear regression model) (RP) |
| Z | Standardised feature components of support vectors selected by SVM |
The supplementary data, Table in this article is published in J-STAGE Data.
The authors declares that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
(URLs on references were accessed on 26 June 2026.)
