2023 was the world's warmest year on record, summer 2023 in Japan, the high temperature reduced rice quality (i.e. the ratio of top-grade rice). The crop grades are directly linked to rice farmers' incomes. For stable production of high-quality crops, UAV monitoring whose introduction cost is reasonable, is attracting attention and expectations. In previous our research, time-series clustering analysis was developed on vegetation indices obtained from UAVs. In this research, the analysis method of previous research was applied to high-temperature years and analyzed changes in vegetation indices and additional fertilizer efects. As a result of time-series analysis, it was possible to determine which rice fields were fertilized appropriately. The fields diagnosed as adequate got the highest yields. In addition, although it was not at a statistically significant level, rice applied with chemical fertilizers had lower yields in high-temperature years than rice applied with organic fertilizers.
Global warming has long become a problem, and the frequency of abnormal weather events is increasing.
In Japan, heavy rain, heavy snow, and high temperatures occurred [
To produce high-quality crops, appropriate farming works are essential. On the other hand, previous
research cited problems with Japanese farming work, such as "new workers cannot share expert farmers’
tacit knowledge" and "it requires large human labor." In other words, the appropriate knowledge transfer
and mechanization of farming works have not progressed. To solve these problems, Smart Agriculture
has been proceeded such as production management systems [
This research used VI for remote monitoring of crops by UAV. VI uses the spectral reflectance of sunlight
on plant leaves [
: : :
Reflectance of band " "
Average of NDRE on date ""
Standard deviation of NDRE on date ""
(1)
(2)
(3)
Monitoring was conducted at private farmer’s paddy fields in Hanamaki City, Iwate Prefecture, Japan,
from 2021 to 2023 at 6 sites in 10 conditions [
The UAV used for monitoring is shown in Figure 3. This UAV is equipped with one RGB sensor for visible light and five monochrome sensors for multispectral. The bands of each monochrome sensor are 450nm (Blue), 560nm (Green), 650nm (Red), 730nm (RE: Red Edge), and 840nm (NIR: Near-Infrared) [20]. However, in this study, only the Red, RE, and NIR bands necessary for calculating NDVI and NDRE were used. To optimize the resolution of the drone images, camera settings were adjusted according to lighting conditions, distance, and drone speed (e.g. flight altitude: 30 meters). The monitoring interval was also set according to the growth stage (i.e. once a week from the panicle formation stage until before harvest). In addition, correcting errors in luminance values and coordinates, a standard reflector was used and 4 ground control points (GCP) were established at the four corners of the site. However, this period was the rainy season so there were missing measurements when drone flight was not possible.
In order to compute VI from aerial photographs taken by UAV, it is necessary to create orthophotos of
each band. Therefore, they were generated using software (Agisoft Metashape, Agisoft LLC, Russia).
Next, calculate NDVI, NDRE, and SNDRE from multiple orthophotos using free and open-source
geographic information system (QGIS [21]). The VI value was calculated for each 3m square mesh.
This was done on all observation days to create VI time-series data for each mesh. That is, there is VI
time-series data for the number of meshes. These meshes were classified into several clusters according
to the pattern of time-series changes in VI using a Python TSC program based on the K-Means++
method to [
As an additional experiment from the previous research [19], a unit acreage sampling was conducted just before harvest. The sample method was to select 5 meshes from each rice field so that the NDVI values at the heading stage were dispersed, and then measure the yield of the 5 plants within each mesh. The selected meshes are shown in Figure 4. By the data of sample yield, the significant diferences in yield between rice fields and which rice fields were most appropriately fertilized were evaluated.
As the results of TSC for Okamizawa, the time-series transitions of each cluster centroid and daily weather data are shown in Table 2. In addition, the cluster distribution map is shown in Table 3. The vertical dashed lines in each figure in Table 2 mean the growth stage transition dates. They were estimated from the growth stage prediction model based on the efective cumulative temperature from the date of transplantation [23, 24].
In 2022, a variety "Hitomebore" was planted which commonly consumed in Japanese households.
As a comparative experiment, the fertilization method was changed for each rice field. In North field,
chemical fertilizer was applied at 0.51g/m² of nitrogen equivalent on July 24th, and in Center and South
ifelds, liquid fertilizer and cow dung were applied on August 1st and 6th [
SNDRE was divided into Cluster-0 (large diference in fluctuation), Cluster-1(upward trend), Cluster2(high values around the heading stage), and Cluster-3(fluctuation range between -0.5 and 0). Cluster-0 is an abnormal value because only a few meshes appear in the corner of the rice field. Cluster-1 was often seen near the ridges of each rice field, so it is thought that the fertilizer washed away by rain and wind was concentrated in these meshes. Cluster-2 appeared in North field and is thought that the efects of chemical fertilizers were evident. Cluster-3 appeared well in Center and South fields. It is thought that liquid fertilizer and cow dung were gradually absorbed into paddy rice over a longer time than chemical fertilizers. From these results, it is thought that the yield in North or Center field is higher, so these two fields were fertilized optimally.
In 2023, "Yumi-azusa" was planted. This variety has a lower taste than Hitomebore, but it is resistant to diseases. So, it is used as commercial rice. The comparative experiment will be roughly the same as in 2022, but detailed work dates were not recorded. Regarding the climate, the daily mean temperature was always higher than normal. In particular, it in August was approximately 4°C higher than normal. Additionally, the weather remained sunny before the heading stage, so there were few precipitations. In NDVI was divided into cluster-0 and cluster-1 (high values around the heading stage), and cluster-2 and cluster-3 (low values around the heading stage). Since high values clusters are often found in Center ifeld, the yield would be higher. NDRE took a higher value compared to 2022. This meant lower stress levels so no growth problems were observed. In SNDRE, cluster-0 was high during the heading stage so it was a good transition that showed the efect of additional fertilizer. This cluster was found in Center ifelds. In any VIs, Center field was expected to be the best fertilization because well-diagnosed clusters are widely distributed in this field. But North and South fields were higher stress levels after the heading stage than Center field. Only Center field might have been less afected by high-temperature years. T m : I 2
V e l b a T
I V D N
The results of the unit acreage sampling are shown in Table 4. Each sample weighs 5 plants. "Average" row shows the average of that column converted into gram per unit area. In 2022, the average yield of Center field was higher and the average yield of South field was lower. This was consistent with the diagnosis from the time-series analysis. In 2023, the Center field had a higher yield than other fields by more than 40g/m². This was in line with the diagnosis from the time-series analysis. In addition, the sample data in 2022 was converted to “g/5plants” units and then a two-way analysis of variance was used to verify the significance of the average yields between years and paddy fields. In the results of this analysis, none of the factors showed any significant diferences at the 5% level (p-value year = 0.488, p-valuefield = 0.345, p-valueyear×field = 0.247). Although there is no significant diference, the yield of North decreased largely in the high-temperature year. It has been reported that organic fertilizers tend to produce higher yields than chemical fertilizers, especially in years with abnormal weather [25, 26]. Chemical fertilizers are convenient because they are quick-acting. However, it is quickly absorbed by crops and weeds so is not as efective at improving soil fertility as organic fertilizers. To accurately measure the diference in efectiveness of diferent types of fertilizers, it seems necessary to improve the unit acreage sampling. Specifically, the sample size was too small (15 or 25 plants per paddy field), so the sampling method and sample size should be improved.
In both 2022 and 2023, the time-series analysis was conducted to diagnose which paddy fields could be properly fertilized. As a result of verifying this through the unit acreage sampling, the yield of well-diagnosed field was the highest in all cases. In other words, by time-series analysis of VIs using TSC, it was possible to evaluate the optimal fertilized fields. Also, the average yield difered by approximately 40g/m² in both years despite applying the same fertilizer to Center and the South fields. This is possibly caused by diferences in soil properties. Center field had good drainage, and South field retains water well.
Using the method of the previous research [18], this paper visualized additional fertilizer efects and high-temperature Influence. By grouping the VI time-series changes of each mesh using TSC, it was possible to visualize the growth and stress conditions, the efect of additional fertilizer, and diagnose which rice fields were fertilized properly. From the unit acreage sampling, all well-diagnosed rice field took the best yield. On the other hand, the yield of chemical fertilizer field decreased significantly in high-temperature years. Thus, under high-temperatures, it was suggested that organic fertilizers which improve soil fertility are more useful than chemical fertilizers which act fast.
To gain further insight, the following three things need to be considered. The first is to analyze images with wavelength bands not used in this study. Using green bands (GNDVI and so on) could visualize the water and nitrogen absorption. Using RGB images could provide ground truth data of soil and crops. The second is an improvement of the TSC algorithm. The K-Means++ method was used in this paper because it supported time-series data in Python. However, better classification can be expected by comparing with hierarchical clustering, DBSCAN, and so on. The third is to supplement missing data. When monitoring data cannot be obtained because the drone could not fly during rainy days, multiple imputation is considered a good approach.
Future works are creating target values for VIs for each growth stage to diagnose well conditions and predict yields. For the former, there is already research. This research makes NDVI target values to achieve the target yield by using a portable NDVI measuring device [27]. It is desirable to apply this to the UAV environment and to formulate target values for other VIs. The latter concerns preliminary evaluation. This research was an ex-post evaluation such as visualization of crop conditions and evaluation of additional fertilizer. However, in the future it is preferable to predict yields with fewer explanatory variables by weather data and growth stage prediction models. Furthermore, a system will need that outputs farm works support in natural language from such prediction models, VIs transitions, cluster maps, and so on. It is expected that UAV monitoring will become more widespread as research continues to meet the needs of the field. [15] T. K. Wardana, Y. Arkeman, K. Priandana, F. Kurniawan, et al., Use of plant health level based on random forest algorithm for agricultural drone target points, Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi) 7 (2023) 637–645. [16] K. Tanaka, A. Kondoh, Mapping of rice growth using low altitude remote sensing by multicopter,
Journal of the Remote Sensing Society of Japan 39 (2019) S1–S17. [17] K. Harashina, K. Yamamoto, M. Maki, Y. Muto, E. Kurashima, Monitoring growth of water-seeded rice in tsunami-stricken paddy fields using uav-mounted multispectral sensor, Journal of the Japanese Society of Irrigation, Drainage and Rural Engineering 87 (2019) 121–126. [18] T. Ito, Lower Costs and Environmental Loading Reduction of Agriculture by Time-series Analysis of Vegetation Indices Using Drones, Master’s thesis, Iwate Prefectural University, Takizawa, 2024. [19] S. Sato, T. Ito, S. Umeki, K. Minamino, Analysis of the impact of high temperatures on paddy rice using drone monitoring and time-series clustering, in: The 86th National Convention of IPSJ, Yokohama, 2024. [20] DJI Co., Ltd., https://www.dji.com/jp/p4-multispectral, 2024. [Accessed 6 April 2024]. [21] QGIS Development Team, https://www.qgis.org/en/site/, 2024. [Accessed 6 April 2024]. [22] C. Shi, B. Wei, S. Wei, W. Wang, H. Liu, J. Liu, A quantitative discriminant method of elbow point for the optimal number of clusters in clustering algorithm, EURASIP journal on wireless communications and networking 2021 (2021) 1–16. [23] E. Kanda, Y. Torigoe, T. Kobayashi, A model to estimate the increase of leaf number on the main culm of the rice plant, Japanese Journal of Crop Science 69 (2000) 540–546. [24] E. Kanda, Y. Torigoe, T. Kobayashi, A simple model to predict the developmental stages of rice panicles using the efective accumulative temperature (modelling, information technology and environment), Japanese Journal of Crop Science 71 (2002) 394–402. [25] Warm region global warming research team, Kyushu/Okinawa branch, NARO, 2007. URL: https: //www.naro.afrc.go.jp/org/karc/seika/kyushu_seika/2007/2007041.html , [Accessed 9 May 2024]. [26] Current status and direction of cultivation technology development compatible with high temperatures, Memoirs of the National Agricultural Research Center for Kyushu Okinawa Region 94 (2010) 71–93. [27] Y. Kaneta, M. Nishida, F. Takakai, T. Sato, New growth diagnosis standards of high-yielding rice and demonstration of high yielding using ndvi by greenseeker handheld crop sensor, Journal of the science of soil and manure 91 (2020) 417–425.