Machine Learning for Precision Agriculture: Predicting Persimmon Peak Harvest Dates and Yield Using Meteorological Data

Abstract

The acute shortage of agricultural labor necessitates the development of predictive models to optimize farm operations. This study investigates the effectiveness of artificial-intelligence-driven models to accurately forecast the timing and yield of persimmon harvests, using meteorological data alongside historical harvest records. An artificial neural network was designed to estimate peak harvest dates by analyzing key meteorological variables. The model was trained and validated using data from the JA Nara Prefecture Nishiyoshino Sorting Facility and Nara Prefecture Agriculture Research and Development Center. Its reliability was confirmed based on mean absolute error, demonstrating the ability to make predictions with an accuracy of approximately three days. Additionally, extreme gradient boosting models were developed to predict yields, incorporating elevation data to refine predictions at the field scale. The model was trained and validated using data from fields cultivated in the Gojo-Yoshino region. The effectiveness of these models was evaluated using root mean square error, demonstrating an improvement in prediction accuracy of up to 20% with the inclusion of elevation data, illustrating their capability to effectively capture yield variations across different orchards. These models can significantly improve labor management, harvest scheduling, and overall productivity within the realm of smart agriculture.

1. Introduction

The agricultural sector is facing an intensifying labor shortage, which exacerbates the workload on farmworkers and necessitates the development of sophisticated predictive tools. Smart agriculture, which integrates information and communications technology, big data, and intelligent devices, has emerged as a viable solution for optimizing farm operations and increasing productivity. Globally, various studies have underscored its efficacy in enhancing agricultural practices, including improved water management through Internet of Things (IoT)-based environmental monitoring [1] and yield enhancement via IoT-integrated machine learning models [2]. In Japan, field experiments have been instrumental in visualizing cultivation conditions using environmental data and in enhancing agricultural operations with smart devices [3].

Fruit harvest timing prediction models have been investigated in previous studies. de Souza et al. [4] developed an artificial neural network (ANN) model for harvesting bananas, using the average temperature, maximum temperature, minimum temperature, precipitation, and sunshine duration as input variables. Their results indicated that the model could predict harvest timing with an error margin of 0.3%. In a previous study [5], we developed an ANN model to predict the peak harvest dates for ‘Tonewase’, ‘Hiratanenashi’, and ‘Fuyu’ persimmons. This model utilizes meteorological variables that influence fruit coloration and can predict the peak harvest date with an accuracy of approximately three days. It was trained using peak harvest date data derived from arrival volume records at the JA Nara Prefecture Nishiyoshino Sorting Facility (JA Peak data) and cultivar adaptability test data from the Nara Prefecture Agriculture Research and Development Center in Japan (CT Peak data). Although the model demonstrated promising performance, it requires additional validation at the orchard level to confirm its robustness under varying growing conditions.

Image-based harvest prediction methods have been investigated for fruit crops. A method that integrates image processing with convolutional neural networks has been proposed for grape yield estimation, achieving an error rate of 11.8% [6]. Furthermore, backpropagation neural networks have been used to predict apple yields based on fruit canopy images, with a root mean square error (RMSE) of 2.34 kg per tree [7]. For citrus yield predictions, a long short-term memory (LSTM)-based model has been developed, which achieved error rates of 4.53% per tree and 7.22% for total yield prediction [8]. These studies highlight the potential of image-based deep learning techniques in yield prediction. However, applying these methods to other fruit trees presents challenges owing to the scarcity of historical image datasets and the extended maturation period of fruit trees, which complicates data accumulation.

Alternatively, numerical data such as meteorological records present a robust alternative for yield prediction. Unlike image-based methods, numerical datasets boast extensive historical records and are amenable to predictive modeling. For example, prior research on blueberries has successfully utilized numerical data, incorporating variables like bee species composition and weather conditions, to predict yields. In this context, several models—multiple linear regression, boosted decision trees, random forests, and extreme gradient boosting (XGBoost)—were evaluated, with XGBoost demonstrating the highest accuracy and achieving an error rate of 5.444% [9].

This study focuses on persimmon production in the Gojo-Yoshino region of Nara Prefecture, a prominent persimmon-growing area in Japan [10]. To address labor shortages, previous studies have explored the use of smart devices [11] and cultivation data [12,13] to enhance agricultural efficiency. Within the framework of the “Development and Improvement Program of Strategic Smart Agricultural Technology Grants”, demonstration experiments are underway to assess remote and automated irrigation systems operated via a dedicated communication network. Moreover, predictive models for harvest timing and yield are being developed to aid in more effective planning and decision making.

In this region, seasonal workers are conventionally hired to reduce labor shortages during the harvest season. Early planning of seasonal employment contributes to the stable securing of labor. To achieve this, it is essential to obtain prior information on factors such as harvest timing and yield. Conventionally, workforce planning has depended on empirical knowledge to predict harvest timing and yield. However, the escalating impacts of climate change [14,15] and the diminishing availability of agricultural labor are rendering conventional forecasting methods increasingly unreliable. Consequently, advanced predictive methodologies that utilize machine learning and meteorological data analysis are urgently needed.

Harvest timing is chiefly influenced by two physiological factors, namely, fruit enlargement and coloration, both of which are affected by meteorological conditions. For example, the ‘Fuyu’ persimmon grows optimally within a temperature range of 20–25 °C but shows growth suppression when temperatures rise above 25 °C. Additionally, temperatures exceeding 30 °C can adversely affect proper fruit coloration [16]. In contrast, ‘Hiratanenashi’ exhibits a secondary enlargement phase in response to the cooler temperatures of autumn [17], while ‘Tonewase’ is notably sensitive to temperature fluctuations in late August [18]. Prior studies have shown that summer irrigation promotes fruit enlargement in ‘Fuyu’, whereas exclusion of rainfall during this period inhibits it [19]. Furthermore, inadequate sunlight has been documented to negatively impact both fruit growth and coloration [20,21].

Several studies have investigated predictive models for the timing of persimmon harvests. For ‘Fuyu’, the initial occurrence of daily average temperatures below 23 °C is linked with the onset of peak harvest, indicating that harvest timing can be estimated by late September [22]. For ‘Hiratanenashi’, a multiple regression model utilizing mid-July average temperature and mid-September maximum temperature has achieved a prediction error margin of 2–3 days [23]. However, these models only provide predictions starting in September, which is often too proximal to the harvest period to be effectively used for workforce planning. To optimize labor allocation, developing models that can predict harvest timing several months in advance is essential. To the best of our knowledge, no prior studies have developed yield prediction models specifically for persimmons.

The authors are currently conducting a smart agriculture project in Japan’s hilly and mountainous regions. This paper reports on a predictive model for estimating the harvest timing and yield of fruit trees within this project. In particular, this study aims to refine the harvest peak date prediction model and develop a yield prediction model utilizing meteorological data. Whereas previous studies have often been limited to validation at a single site due to difficulties in collecting data from multiple locations, this study utilizes harvest data collected from multiple sites. The harvest peak date model will be enhanced with a more comprehensive dataset and its performance will be assessed at the field level. The yield prediction model will be devised into two variants: one for estimating total yield at the regional level and the other for calculating yield for individual orchards. The primary cultivars targeted for harvest timing prediction are ‘Tonewase’ and ‘Fuyu’, while ‘Tonewase’, ‘Hiratanenashi’, and ‘Fuyu’ are selected for yield prediction. By integrating sophisticated predictive modeling techniques into smart agriculture, this study aims to improve the efficiency of harvest scheduling and resource allocation in persimmon production.

2. Materials and Methods

2.1. Study Area

The study area is the Gojo-Yoshino region of Nara Prefecture (Figure 1), a prominent persimmon-growing area in Japan. The region has seen substantial development in agricultural land and enhancements in irrigation through national and prefectural initiatives [10]. The primary cultivars are the astringent persimmons ‘Tonewase’ and ‘Hiratanenashi’, along with the non-astringent ‘Fuyu’. Greenhouse-grown persimmons are harvested as early as July, with field-grown fruits being harvested subsequently in the order of ‘Tonewase’, ‘Hiratanenashi’, and ‘Fuyu’. The harvest begins in late September for ‘Tonewase’, late October for ‘Hiratanenashi’, and early November for ‘Fuyu’. The cultivation cycle includes pruning (January–February), bud thinning (late April), and fruit thinning (July–August), progressing through sprouting and leaf expansion in March and April, flowering in May, and culminating with the harvest from September onward.

2.2. Prediction Model for Harvest Peak Date

2.2.1. Data Collection

The construction of the model involved the use of both meteorological and harvest data. Meteorological data were obtained from the “Mesh Agriculture Meteorological Data System” provided by the National Agriculture and Food Research Organization [24,25] (https://amu.rd.naro.go.jp/wiki_open/doku.php?id=start, accessed on 1 December 2024). This dataset encompasses six crucial variables: daily average temperature, daily maximum temperature, daily minimum temperature, duration of sunshine, total precipitation, and precipitation frequency (defined as the number of days receiving at least 1 mm of precipitation).

The harvest data utilized in this study comprised three distinct datasets: (1) JA Peak data, documenting arrival volumes at the JA Nara Prefecture Nishiyoshino Sorting Facility; (2) Field Peak data from Field A within the study region; and (3) CT data from the Nara Prefecture Agriculture Research and Development Center. The JA Peak data span from 2001 to 2023, covering ‘Tonewase’ and ‘Fuyu’ cultivars, whereas the Field Peak data (2008–2023, except for 2014) concentrate on ‘Tonewase’. The CT data, covering the years 1991–2022 (excluding 1993–1998 and 2015), focus on ‘Fuyu’. These datasets were pivotal for the training and evaluation of the models. Figure 2 illustrates the distribution of the harvest peak data utilized in this study, detailing the breakdown by cultivar and source dataset.

2.2.2. Model Development

The prediction model adheres to the methodology outlined in Okayama et al. [5] and is predicated on an ANN. ANNs are capable of modeling nonlinear relationships within data, making them well-suited for capturing complex patterns and features. Furthermore, ANNs are widely utilized and are expected to offer high predictive accuracy. The ANN used in this study was configured with five hidden layers, trained for 500 epochs, and employed a batch size of 32. Since fruit coloration—an important determinant of harvest timing—is affected by meteorological conditions [16,17], this study considered the same variables employed in Okayama et al. [5] as candidate input features. Furthermore, two additional variables were introduced to improve the predictive performance of the model. To enhance the predictive accuracy of the model, additional input variables were incorporated, as detailed in

Table 1. Potential input variables for predicting harvest peak dates.

Input Items	Symbol
Mean temperature for a period	MTP
Cumulative mean temperature	CMT
Cumulative maximum temperature	CMXT
Cumulative minimum temperature	CMNT
Cumulative mean temperature for a period	CMTP
Cumulative diurnal temperature range	CTR
Cumulative effective temperature	CET
Cumulative sunshine duration	CSD
Cumulative precipitation	CPR
Cumulative precipitation frequency *	CPF

* number of days with precipitation ≥1 mm.

. The newly added variables were the cumulative precipitation frequency (CPF) and cumulative effective temperature (CET), the latter reflecting the sum of temperatures conducive to plant growth. For persimmons, the optimal temperature threshold is estimated at 10 °C for astringent varieties and 13 °C for non-astringent varieties [26]. Consequently, the effective temperature was calculated using the formula “daily average temperature—10/13 °C”.

Feature selection was performed using the permutation importance (PI) method [27,28], a technique that assesses the significance of a feature by evaluating the decrease in prediction accuracy when that feature is randomly shuffled. A higher PI value signifies a more crucial feature. This method is particularly advantageous for interpreting complex machine learning models, including those based on deep learning. In this study, input features that demonstrated low PI values were systematically excluded, and the model was retrained to optimize feature selection. The refined model was then evaluated using the mean absolute error (MAE).

2.2.3. Data Partitioning and Evaluation

The dataset was divided into training, validation, and test subsets. The validation set was employed to fine-tune input feature selection, while the test set was designated for the final model assessment. For the ‘Tonewase’ cultivar, the test dataset encompassed 11 years, including 4 years of data from JA Peak and 7 years from Field Peak. The validation dataset for ‘Tonewase’ included data from 5 years (3 years from JA Peak and 2 years from Field Peak), with the remainder allocated for training.

For the ‘Fuyu’ cultivar, the test dataset included 10 years of data (5 years each from JA Peak and CT data), and the validation dataset comprised 5 years (3 years from JA Peak and 2 years from CT data). Each dataset was meticulously prepared in 20 sets for both cultivars to ensure a robust evaluation process.

The model’s performance was quantified using the MAE and the prediction error (the difference between the true and predicted values) as of 1 May and 1 June. Given the sensitivity of deep learning models to initial conditions, each dataset was trained 10 times to reduce variability [29]. Additionally, to further scrutinize prediction performance, harvest peak dates were classified into three categories—“early”, “average”, and “late”—based on their respective mean and standard deviation (

Table 2. Categorization of harvest peak dates.

Cultivar	“Early”	Total	“Average”	Total	“Late”	Total
‘Tonewase’	10/1–10/4	4	10/5–10/11	30	10/12–10/14	4
‘Fuyu’	11/15–11/18	4	11/19–11/27	36	11/28–12/6	7

). Comparative analyses were conducted against the Okayama et al. [5] model, providing a comprehensive evaluation of the predictive capabilities.

2.3. Prediction Model for Yield

2.3.1. Data Collection

The meteorological data utilized for yield prediction were identical to those employed for predicting harvest peak dates. Yield data were sourced from two primary locations: (1) JA yield data (2001–2023), which documents annual arrival volumes at the JA Nara Prefecture Nishiyoshino Sorting Facility (Figure 3), and (2) Field Yield data (2008–2023) from 14 fields within the study area (Figure 4). The Field Yield data were standardized to yield per 10 a (ten ares, which is equivalent to 1000 m²) considering the cultivation ratio for each variety. The dataset comprises 95 data points for ‘Tonewase’, 79 for ‘Hiratanenashi’, and 47 for ‘Fuyu’ (

Table 3. Detailed field yield data.

Field Number	‘Tonewase’	Year	‘Hiratanenashi’	Year	‘Fuyu’	Year
A	Tone_A	2014–2023	Hira_A	2014–2023
B	Tone_B	2023	Hira_B	2023
C	Tone_C	2008–2023 * 2010, 2014
D	Tone_D	2008–2023	Hira_D	2008–2023
E	Tone_E	2008–2023	Hira_E	2008–2023	‘Fuyu’_E	2008–2023
F	Tone_F	2013–2023
G	Tone_G	2013–2023
H	Tone_H	2019–2023
I	Tone_I	2013–2023
J			Hira_J	2008–2023 * 2017	‘Fuyu’_J	2008–2023
K			Hira_K	2013–2023 * 2017	‘Fuyu’_K	2013–2022
L			Hira_L	2013–2023
M					‘Fuyu’_M	2020–2023
N					‘Fuyu’_N	2021
Total	9	95	7	79	5	47

‘*’ denotes years with missing data.

).

2.3.2. Model Development

The analytical model employed was XGBoost [30], configured with default parameter settings as specified in

Table 4. Hyperparameters for the XGBoost model.

Parameter	Description	Default Value
learning rate	Step size shrinkage to prevent overfitting	0.3
max_depth	Maximum depth of a tree	6
subsample	Fraction of training data randomly sampled for each boosting round	1.0
n_estimators	Number of boosting rounds (trees)	100

. Due to its high computational efficiency, XGBoost is considered well-suited for practical applications. This study developed two principal model types: the large-scale yield prediction model (LS-Model) and the field-level yield prediction model. The LS-Model, utilizing JA yield data, estimates total yield at the regional level, whereas the field-level yield prediction model forecasts yield per unit area at the individual field level.

Within the field-level yield prediction model, two variations were explored:

• Field-level model using regional input variables (FL-Model1)—This model adopted the same input features as the LS-Model but tailored them for field-level yield predictions.
• Field-level model with elevation factor (FL-Model2)—This variation of FL-Model1 incorporated elevation data to accommodate inter-orchard variability, with elevation determined from the latitude and longitude of each field.

The candidate input variables are outlined in

Table 5. Potential input variables for yield prediction, spanning January to September for ‘Tonewase’ and January to October for ‘Hiratanenashi’ and ‘Fuyu’.

Data	Processing Method	Symbol
Mean Temperature	Monthly Average	MT
Maximum Temperature	Monthly Average	MXT
Minimum Temperature	Monthly Average	MNT
Temperature Range	Monthly Average	TR
Sunshine Duration	Monthly Cumulative	CSD
Precipitation	Monthly Cumulative	CPR

. Fruit enlargement, a key factor influencing yield, is affected by meteorological conditions such as temperature and precipitation [18,19]. Therefore, the daily mean temperature (MT), maximum temperature (MXT), minimum temperature (MNT), and temperature range (TR) were computed on a monthly average basis, while sunshine duration (CSD) and precipitation (CPR) totals were accumulated monthly.

Feature selection was conducted in a two-stage process: Initially, correlation analysis identified and excluded features with weak correlations to yield (where −0.2 < r < 0.2). To further address multicollinearity, variables exhibiting high inter-correlations (|r| > 0.7) as well as those weakly correlated to yield were eliminated. The final feature set was determined using a wrapper method [31] (forward selection), wherein features were added iteratively to minimize the RMSE.

2.3.3. Data Partitioning and Evaluation

The JA yield data were employed to construct the LS-Model, while the field yield data facilitated the development of the field-level models FL-Model1 and FL-Model2. The JA Yield dataset was divided into training, validation, and test subsets. The validation subset played a crucial role in feature selection, while the test data from 2021 to 2023 were earmarked for model evaluation. The remaining data spanning from 2001 to 2020 were allocated for training and validation, with various combinations where 17 years were designated as training data and 3 years as validation data.

The field yield data were divided into training and test datasets, with cross-validation procedures implemented to ensure a robust model evaluation. For this process, ‘Tonewase’ and ‘Hiratanenashi’ were subjected to 20-fold cross-validation, whereas ‘Fuyu’ underwent 15-fold cross-validation.

3. Results and Discussion

3.1. Prediction Model for Harvest Peak Date

3.1.1. Selected Input Items

The results of the input feature selection via the PI method are depicted in Figure 5. For ‘Tonewase’, three critical input features were selected: CMT, CMXT, and MTP. Regarding ‘Fuyu’, six input features were identified: CMT, CMXT, CMNT, CTR, CET, and MTP. For both cultivars, all selected features were temperature-related, with CMT demonstrating the highest PI value. Notably, the PI value for CMNT in ‘Fuyu’ was more pronounced than in ‘Tonewase’. Features such as CPR, CPF, and CSD were excluded for both varieties as their PI values fell below the significance threshold of 5. This finding corroborates earlier studies that emphasized the predominant influence of temperature on fruit coloration [16,18,32]. Moreover, the pronounced correlation between minimum temperature and ‘Fuyu’ harvest timing has been well-documented [22]. These results highlight the dominant impact of temperature on harvest timing over precipitation or sunlight, consistent with the conclusions presented by Okayama et al. [5].

The selected input features for ‘Tonewase’ align with those reported in a study by Okayama et al. [5]. However, for ‘Fuyu’, the selected input features diverged, which may be attributed to differences in harvest timing trends across the datasets employed in this study. A detailed analysis comparing harvest timing across datasets is illustrated in

Table 6. Comparison of harvest peak date data, with ⃝ marking data used by Okayama et al. [5] and the proposed model.

Cultivar	Type	Okayama et al. [5]	Proposed Model	Average ± std	Max	Min
‘Tonewase’	JA	⃝	⃝	10/7 ± 3.0	10/14	10/1
	Field		⃝	10/9 ± 3.0	10/13	10/2
‘Fuyu’	JA		⃝	11/20 ± 2.5	11/26	11/15
	STD	⃝	⃝	11/26 ± 3.7	12/6	11/17

. While the variation trends for ‘Tonewase’ remained consistent across datasets, the JA Peak data for ‘Fuyu’ showed an earlier harvest timing compared to the CT data, influencing the selection of minimum temperature as a critical input feature.

Additionally, the increase in selected input features for ‘Fuyu’ compared to Okayama et al. [5] indicates that a broader range of features was necessary to effectively capture the varied trends in harvest peak dates.

3.1.2. Evaluation of the Model

The performance of the prediction models is detailed in Figure 6. For ‘Tonewase’, the MAE was 3.9 days on 1 May and 1 June. In the case of ‘Fuyu’, the MAE was 4.0 days on 1 May and reduced to 3.3 days by 1 June. The lower error rate for ‘Fuyu’ on 1 June suggests that the inclusion of additional temperature-related variables, such as CET, may have enhanced the model’s accuracy as the growing season advanced.

Further analysis of prediction errors across test datasets is presented in Figure 7. The model registered errors exceeding 5 days in 3 out of 11 years for ‘Tonewase’ and in 4 out of 10 years for ‘Fuyu’. These larger discrepancies predominantly occurred in years where the harvest peak dates notably deviated from the average. Conversely, when the prediction errors were within ±2 days, the actual harvest peak dates typically fell within 3 days of the average, demonstrating that the model performed effectively under typical conditions but faced challenges with more extreme variations.

The model’s accuracy was further assessed by classifying harvest peak dates into three categories: “early”, “average”, or “late”, based on historical distributions (Table 2). The classification analysis, detailed in Figure 8, indicated that prediction errors were generally skewed toward the mean harvest timing. Specifically, early harvest dates were frequently predicted to occur later than they actually did, while late harvest dates were predicted to be earlier. This pattern suggests that the model struggles to accurately capture extreme variations, likely owing to a scarcity of training data for such conditions, as confirmed by a multiple comparison test (p < 0.001).

Notably, the harvest peak dates for ‘Fuyu’ in the 2012 JA Peak dataset and the 2021 CT dataset showed larger than expected errors, even though they were categorized as average. This discrepancy may stem from regional differences in meteorological conditions that were not adequately considered in the model. Given that the JA Peak dataset reflects regional averages and the CT dataset is specific to the Nara Prefecture Agriculture Research and Development Center, disparities in temperature trends between these locations could have impacted the predictions. To reduce such discrepancies, future models could benefit from incorporating location-specific meteorological features to enhance prediction accuracy.

3.1.3. Comparative Analysis

The comparative results between the current model and the model proposed by Okayama et al. [5] are depicted in Figure 9. For ‘Tonewase’, the MAE of the Okayama et al. [5] model was 4.2 days on 1 May and 4.3 days on 1 June. In contrast, the current model reduced this error to 3.9 days for both dates, indicating a modest improvement in prediction accuracy. Conversely, for ‘Fuyu’, the Okayama et al. [5] model recorded an MAE of 5.0 days on 1 May and 5.1 days on 1 June, while the current model significantly enhanced accuracy to 3.9 days and 3.3 days, respectively. A t-test validated that these enhancements were statistically significant (p < 0.001), confirming that the current model outperforms the earlier model in predicting harvest peak dates.

Table 6 offers an exhaustive comparison of harvest peak date data across the datasets utilized in both studies. The analysis reveals that the timing for ‘Tonewase’ harvests was relatively consistent across datasets, whereas for ‘Fuyu’, the JA Peak dataset showed an earlier harvest timing by approximately six days compared to the CT dataset. This deviation was statistically significant (p < 0.001) and likely contributed to the improved prediction accuracy of the current model due to the inclusion of additional data capturing these variances.

In summary, the enhancements in prediction accuracy for ‘Fuyu’ were more pronounced than those for ‘Tonewase’, possibly owing to the greater variability in the harvest timing data. These outcomes suggest that amassing a diverse dataset with a broad range of harvest peak dates significantly contributes to the robustness of the model and the enhancement of its predictive accuracy.

3.2. Prediction Model for Yield

3.2.1. Selected Input Items

The selection results for input variables in the yield prediction model are delineated in Figure 10. A total of 16 variables were chosen, based on JA yield data and meteorological conditions. For ‘Tonewase’ and ‘Hiratanenashi’, several indicators related to MXT and CSD were identified as significant. Notably, MXT in May exhibited strong correlations with yield (p < 0.01), and similarly, CSD demonstrated significant correlations (‘Tonewase’: p < 0.01, ‘Hiratanenashi’: p < 0.05). In contrast, ‘Fuyu’ showed a greater responsiveness to MNT, TR, and CPR. Among these, CPR in May was statistically significantly correlated with yield (p < 0.05).

The influence of CPR on ‘Fuyu’ yield exhibited variability across different periods of the growing season. Negative correlations were observed from May to July, indicating detrimental impacts, while positive correlations appeared from August to October. This pattern corresponds with previous research, which suggests that excessive rainfall in early summer can escalate the incidence of anthracnose disease [33] and angular leaf spot [34], adversely affecting fruit development. Moreover, rainfall during this period can disrupt the application of pesticides, further diminishing yields. Conversely, irrigation and precipitation in the later months, from August to October, are known to promote fruit enlargement in ‘Fuyu’ persimmons [19]. These findings highlight that the impact of precipitation on yield is highly dependent on the timing within the growing season, with early summer precipitation having negative effects and late summer to autumn precipitation offering benefits.

The selected input features were refined using a wrapper method, as depicted in Figure 11. For ‘Tonewase’, the refined model incorporated TR in April, MNT, and CSD in May. For ‘Hiratanenashi’, critical variables included CPR in January, MNT in February, TR and MXT in April and May, CSD in May, and TR in October. For ‘Fuyu’, the most influential features were TR in February and May, CPR in July and August, and MT in August.

For ‘Tonewase’ and ‘Hiratanenashi’, temperature-related features during April and May were predominant, indicating that temperature conditions during early fruit development are crucial for astringent persimmons. Conversely, for ‘Fuyu’, precipitation during the later stages of fruit growth played a more significant role, highlighting the importance of water availability for non-astringent persimmons.

3.2.2. Evaluation of the Model

The performance of the LS-Model was evaluated using RMSE, summarized in

Table 7. Prediction accuracy (%), based on the RMSE of the LS-Model across validation and test datasets.

Cultivar	Validation	Test
‘Tonewase’	12.62 ± 6.9	12.93 ± 7.7
‘Hiratanenashi’	11.30 ± 5.1	9.63 ± 8.5
‘Fuyu’	10.65 ± 5.8	10.47 ± 6.9

. The RMSE for the validation dataset was 12.62% for ‘Tonewase’, 11.30% for ‘Fuyu’, and 10.65% for ‘Hiratanenashi’. For the test dataset, the RMSE values were 12.93% for ‘Tonewase’, 9.63% for ‘Hiratanenashi’, and 10.47% for ‘Fuyu’. These results demonstrate that the LS-Model effectively captured regional yield characteristics, maintaining an error margin of approximately 10%.

The results for FL-Model1 and FL-Model2 are outlined in

Table 8. Prediction accuracy results (kg/10 a) for the FL-Model1 and FL-Model2, highlighting the impact of incorporating elevation data on RMSE.

Cultivar	FL-Model1	FL-Model2
‘Tonewase’	48.94 ± 17.47	26.56 ± 9.87
‘Hiratanenashi’	30.21 ± 9.60	26.44 ± 9.93
‘Fuyu’	38.45 ± 18.61	33.77 ± 12.14

. The RMSE of FL-Model1, which utilized regional-level input variables, registered at 28.82% for ‘Tonewase’, 23.34% for ‘Hiratanenashi’, and 42.06% for ‘Fuyu’. In comparison to the LS-Model, FL-Model1 exhibited reduced accuracy, with RMSE values increasing by approximately 30% for ‘Tonewase’ and ‘Fuyu’. This discrepancy suggests that regional input variables alone are insufficient for precise field-level yield prediction, as field yield data are more influenced by localized conditions than regional yield data.

FL-Model2, which incorporated elevation data, exhibited improved accuracy over FL-Model1. The RMSE decreased by approximately 10% for ‘Tonewase’, underscoring the significant influence of elevation on inter-orchard yield variability. However, the improvement for ‘Fuyu’ was less significant, likely due to the limited number of field datasets available.

Further analysis of yield variation within and between orchards is summarized in

Table 9. Intra-orchard yield variation for each cultivar, represented by standard deviations (SDs) that reflect yield differences within orchards due to factors such as microclimate. The area of 10 a is equivalent to 1000 m².

Cultivar	Filed	Mean Yield (kg/10 a)	Intra-Orchard Variation (SD, kg/10 a)
‘Tonewase’	Tone_A	2017	472
	Tone_B	2351	―
	Tone_C	2862	450
	Tone_D	2083	596
	Tone_E	2444	610
	Tone_F	3020	635
	Tone_G	4567	962
	Tone_H	1284	224
	Tone_I	4876	988
‘Hiratanenashi’	Hira_A	4719	938
	Hira_B	4282	―
	Hira_D	3306	724
	Hira_E	3413	732
	Hira_J	2574	803
	Hira_K	4359	760
	Hira_L	5053	999
‘Fuyu’	‘Fuyu’_E	2816	641
	‘Fuyu’_J	2742	1152
	‘Fuyu’_K	2484	668
	‘Fuyu’_M	4556	1868
	‘Fuyu’_N	1119	―

and

Table 10. Inter-orchard yield variation for each cultivar, represented by SDs that indicate the influence of orchard-specific factors such as elevation and management practices. The area of 10 a is equivalent to 1000 m².

Cultivar	Mean Yield (kg/10 a)	Inter-Orchard Variation (SD, kg/10 a)
‘Tonewase’	2834	1185
‘Hiratanenashi’	3958	883
‘Fuyu’	2743	1225

. Table 10 shows the inter-orchard yield variation for each cultivar, with the standard deviation (SD) being 1185 kg/10 a for ‘Tonewase’, 883 kg/10 a for ‘Hiratanenashi’, and 1225 kg/10 a for ‘Fuyu’. These figures indicate that ‘Fuyu’ displayed the highest inter-orchard yield variability, likely attributable to differences in orchard-specific conditions such as soil quality and irrigation practices. Intra-orchard variability was most pronounced for ‘Fuyu’ (1868 kg/10 a), and least for ‘Tonewase’ (988 kg/10 a).

The dataset size for ‘Fuyu’ was marginally less than 50, marking it as the smallest dataset among the three varieties. It included data from nine orchards for ‘Tonewase’, seven for ‘Hiratanenashi’, and five for ‘Fuyu’. The smaller number of ‘Fuyu’ orchards may have contributed to the limited improvement in prediction accuracy observed for this cultivar. Conversely, ‘Tonewase’, which had the largest number of orchards, showed a significant improvement in accuracy with the inclusion of elevation data. ‘Hiratanenashi’ demonstrated lower inter-orchard variation, while ‘Fuyu’, with fewer orchards, experienced a less pronounced effect from the addition of elevation data.

These findings emphasize that meteorological data alone may not adequately capture both inter-orchard and intra-orchard variability. The integration of additional data, such as soil conditions and management practices, could further enhance model accuracy. Future research should focus on incorporating additional environmental factors, such as soil quality and irrigation patterns, to improve the robustness of yield prediction models.

3.3. Novel Contributions and Comparisons with Previous Studies

The harvest timing prediction model exhibited a notable improvement in accuracy compared to the model developed using the dataset of Okayama et al. [5]. This improvement is attributable to the increased volume and variability of the dataset, suggesting that greater dataset diversity enhances model performance. In a previous study by Obsie et al. [9], blueberry yield was predicted using weather data and bee species composition from a single field, yielding a prediction error of 5.444%. In contrast, the present study incorporated not only weather data but also elevation data that represent the topographic characteristics of multiple fields. The integration of elevation data substantially improved model accuracy and facilitated the development of a prediction model adaptable to diverse field conditions. These findings underscore the importance of including topographic features and qualitative data, such as cultivation practices, to enhance the generalizability of yield prediction models. Overall, the results highlight the critical role of dataset diversity and the integration of qualitative data in constructing highly accurate field-level prediction models.

4. Conclusions

This study developed predictive models for persimmon harvest timing and yield, achieving high levels of accuracy by integrating meteorological data with machine learning techniques. For harvest timing prediction, increasing the dataset size significantly reduced the variability in prediction results. Moreover, capturing site-specific variations in peak harvest dates necessitated a more extensive array of input features. Future research should explore a more detailed analysis of the interrelationships among these features to further refine model accuracy. Additionally, this study utilized accumulated meteorological data to construct a pseudo-time-series dataset and developed predictive models using an ANN. The adoption of time-series-specific architectures, such as LSTM, is anticipated to enhance prediction performance. In this study, we focused on major cultivars with a relatively large dataset. Validation for other cultivars will expand the range of cultivars to which the model can be applied in the future. For yield prediction, the model that incorporated regional meteorological variables demonstrated a low error rate of approximately 10%, effectively capturing regional yield trends. However, the field-level models exhibited higher error rates, indicating that regional input variables alone do not sufficiently capture field-specific variations. Yield variability at the field level is influenced by both inter-orchard and intra-orchard factors. Integrating elevation data into the field-level model improved prediction accuracy by addressing inter-orchard variability. However, to further increase model robustness, additional factors such as soil properties, irrigation management, and cultivation techniques should be integrated. Developing a generalized model that remains effective across different regions will necessitate identifying key variables that encapsulate both inter-orchard and intra-orchard variations. No correlation was observed between the predictions of the two models.

Additionally, this study identified common meteorological factors that influence both harvest timing and yield. Maximum temperature emerged as a critical factor for ‘Tonewase’, while minimum temperature was more impactful for ‘Fuyu’. These factors are crucial for both harvest timing and yield, highlighting the potential to develop a unified model that simultaneously predicts both dimensions. Employing a multi-task learning framework capable of forecasting harvest timing and yield within a single model could lead to further enhancements in agricultural planning and decision-making.

By implementing AI-driven prediction models, this research advances smart agriculture, facilitating optimized labor management and harvest scheduling in persimmon production. The developed harvest prediction model demonstrates high accuracy at the regional level, making it valuable for optimizing production and distribution. However, the use of only meteorological data or a limited set of field observations constrains model accuracy, particularly in regional-scale applications and generalization. Incorporating qualitative data—such as soil conditions and cultivation practices—as well as data from a larger number of fields with diverse topographic characteristics, is expected to improve model accuracy and generalizability. Future studies should investigate additional environmental factors and pursue long-term dataset expansions to enhance the generalizability and practical applicability of these models.