2.1.

Experimental Site

The trial paddy is located in the Niigata Agricultural Research Institute in Niigata Prefecture, known as the rice capital of Japan. Under normal farming practices, paddy fields are plowed and filled with water in April, and the rice-growing season starts in May and ends in late September. The rice was planted in two different paddy fields in two different years; otherwise, growing conditions, including rice variety, transplanting density, row direction, fertilizer treatments, and water supply were kept the same. The mean weather conditions were presumed to be unchanged between the two growing seasons. The test area was $2\mathrm{m}\times 15\mathrm{m}$ and was divided into five plots. Koshihikari, a popular rice variety in Japan, was planted with three different planting densities and two orientations (Table 1). The planting densities were classified as sparse, moderate, and dense, corresponding to 11.2, 15.1, and $21.2\text{plants}/{\mathrm{m}}^2$, respectively, and one orientation was parallel to the movement of the rice-planting machine while the other was perpendicular to it [Fig. 1(b)]. Twenty day-old seedlings were transplanted in mid-May to the water-filled paddy field, which was treated with $3\mathrm{g}\mathrm{N}/{\mathrm{m}}^2$ of ground fertilizer.

Table 1

Planting density and geometry.

Fig. 1

Schematic of the field experiment setup for measuring the 3-D laser point cloud. (a) Laser scanning plane and (b) planting geometry of the five plots (P1 to P5). Arrows indicate row orientation. Row orientation is not considered in plot 3.

2.2.

Laser Scanner Instrument and Data Acquisition

The instrument was set up to obtain a 3-D point cloud of the rice plant from above. The system included a line laser scanner (SICK LMS 200) and a slide rail along which the scanner could be moved.²² Both were hung from a rack at a height of 3 m from the ground surface, and the central scan angle of the laser scanner corresponded to the nadir. The scanning plane was set vertically downward and perpendicular to the row direction in plot 1 (Fig. 1). The field of view was 50 deg, with a scan angle from $-25\mathrm{deg}$ to 25 deg; the angular resolution was 0.25 deg; and the laser beam wavelength was 905 nm. The laser footprint diameter from the 3 m installation height ranged from 2.5 to 3 cm, and the overlap in laser footprints within the scanning line was more than 60%. Each scanning line had 200 scanning spots. To obtain the 3-D point cloud data, the laser scanner was moved along the slide rail for a distance of $\sim 1\mathrm{m}$ at a speed of $2\mathrm{cm}/\mathrm{s}$ under motorized power. This speed was chosen as the maximum velocity of the motor device. Figure 1 illustrates a schematic of the instrument on the rail in the test area. The setup described here, in which the scanner is mounted upside down above the plant canopy, was previously used by Lumme et al. with a 3-D laser scanner for ear recognition and plant height determination.²³

The measurements were carried out during the vegetative stage of growth, in which rice plant height rapidly increases. Sixteen observations were carried out in total between June and July of each year, and in only the first field observation of each year was the paddy field filled with water. Additional observations were made after the harvest in September, to obtain measurements from the bare soil, while a total station was also used to measure the elevation of the four corners of the target area in each plot. The data from both measurement methods were used to compute the ground position, thereby providing reference data with which the results of the proposed method could be validated. In addition to the acquisition of 3-D laser point clouds, rice plant height, the number of stems, and vegetation coverage were measured manually several times during the study period.

3.1.

Manual Measurement Results

According to the manual measurements made in this study, rice plant height ($H$) increased linearly over time [Fig. 2(a)]. $H$ did not differ between the 2 years, probably because the same cultivar and fertilizer treatments were used. Moreover, rice plants in moderate density plots had higher $H$ than in other plots, with the maximum $H$ occurring in the moderate density plot with parallel row orientation. The largest detected difference between plots was 7 cm in 2013 and 4 cm in 2014. The standard deviation of average $H$ was $\sim 3\mathrm{cm}$ for the 2 years. Therefore, the effects of planting density and row orientation on $H$ are not thought to be significant in this study. Biswas et al.²⁴ also stated that plant height did not differ significantly with differing planting geometry. However, the number of stems and vegetation coverage showed clear differences between the two growing seasons [Figs. 2(b) and 2(c)], with both parameters showing higher values in 2014. For example, 55 days after transplanting, the number of stems and vegetation coverage in 2014 were both 1.4 times larger than at the same time in 2013, despite rice plant height being the same in both years.

Fig. 2

Mean values of manually measured rice growth parameters. (a) Rice plant height, (b) number of stems, and (c) vegetation coverage.

3.2.

Targeted Data Extraction

The range data contained in the laser point clouds were converted to vertical distances ($D$), which represent the distances between the scanning points and the installation height of the laser scanner [Eq. (1)]. Here, $D_i$ represents the vertical distance of the laser scanning points, ${\theta }_i$ is the inclination angle at the $i$’th position in the scanning line, and $r_i$ is the range from the sensor to the target at the $i$’th scan angle [Fig. 1(a)].

To remove the effects of incident angle, small target areas ($90\mathrm{cm}\times 60\mathrm{cm}$) were selected inside each test plot within $\pm 8\mathrm{deg}$ inclination angles [Fig. 3(a)]. The 3-D laser point clouds corresponding to the target areas were then extracted. Histograms plotting the percentage of laser pulses against vertical distance for plot 3 are shown in Fig. 4. In this, changes in the histogram shape indicate rice growth with time.

Fig. 3

Plot 3 target area in 2014. (a) Location of target area, (b) vertical distance relative to scan angle of scanline A–A in plot 3, (c) range image, and (d) intensity image.

Fig. 4

Histograms of 3-D point cloud data from plot 3, collected in 2014. Red dashed lines show the ground position after harvesting the crop. Blue dashed line shows the water surface position in the first observation. In other field observations, the paddy was not filled with water.

To determine the position of the ground surface, the average vertical distance to scanning points ($D_G$) in bare ground data was used. Comparison of this with the total station measurement showed good agreement, as expected, and the difference of 1 mm in $D_G$ between the two methods is deemed insignificant. Thus, it was assumed that the collection of scanning points with a vertical distance approximately equal to $D_G$ represented a scanning of the ground surface. On this basis, the number of laser pulses reaching the ground surface was seen to decrease as rice growth progressed. However, paddy fields are typically filled with water, and in such situations it is difficult for the laser pulse to reach the ground surface because some laser wavelengths tend to be fully or partially absorbed by water. A simple test performed in 2013 showed that pulses from a SICK LMS 200 laser scanner could reach the ground surface beneath 4 cm of clear water, which is the standard water depth in the study area [Figs. 5(a) and 5(b)]. However, during the first field observation on June 18, 2014, there were many bubbles on the water surface and unclear water conditions prevailed [Fig. 5(c)]. Subsequently, it was found that with a vegetation coverage of 20%, more than 50% of laser pulses were absorbed by the water and did not reach the ground surface. Thus, the histogram peak in Fig. 4(a) indicates the position of the water surface and the red dashed line indicates the ground surface position directly beneath it. Fortunately, from the second observation in 2014 onward, the ground surface was not covered by water, so the laser pulse was able to reach the ground surface [Fig. 5(d)]. The histogram peaks closest to the horizontal axis indicate the position of the ground surface [Figs. 4(b)–4(g)]. It should be noted that the ground surface was typically uneven, with mounds around the rice plant trunks; this accounts for the small group of scanning points with smaller $D_G$ vertical distance values (Fig. 4).

Fig. 5

The likelihood of SICK LMS 200 laser pulses reaching a ground surface covered by water. Images (a) and (b) are an infrared image and an intensity image collected by SICK LMS 200 on June 21, 2013, respectively. For this, potted rice plants were put into clear water, and the tops of the pots at a water depth of around 4 cm could be recognized from the intensity image. In contrast, (c) the paddy field was covered by turbid water with many bubbles on the water surface on June 18, 2014, and (d) the paddy field was not covered by water on June 24, 2014.

3.3.

Estimation of Rice Plant Height

Problems associated with the detection of the ground surface have already been discussed. It is possible to determine the ground surface shortly before transplanting or after harvesting the crop; however, in this study, the laser pulse easily reaches the rice canopy, and therefore we choose a new reference point at the top of the rice plant to determine $H$. The proposed method utilizes percentile analysis, which has been commonly used for analyzing plant heights obtained with a laser scanner.^25,26 Therefore, a procedure using percentile ranks and vertical distance was performed, whereby the closest percentiles to the top ($p^{\mathrm{t}}$) and bottom ($p^{\mathrm{b}}$) of the rice plant were determined. On the basis of this, the position of the top ($D^{\mathrm{t}}$) and the bottom of the rice plant ($D^{\mathrm{b}}$) were then located (Fig. 6).

Fig. 6

Determination of the top and bottom location, where $p^{\mathrm{t}}$ and $p^{\mathrm{b}}$ are the closest percentiles to the top and bottom of the rice plant, respectively; $D^{\mathrm{t}}$ and $D^{\mathrm{b}}$ are the vertical distances from the scanning point to the installation height of the laser scanner at $p^{\mathrm{t}}$ and $p^{\mathrm{b}}$.

First, the top of the rice plant ($D^{\mathrm{t}}$) was identified. During rice crop monitoring, numerous factors such as wind, insects, or small airborne particles affect laser scanning data. Therefore, we assume here that $<1\%$ of laser scanning points were affected by these factors. A search procedure was performed to identify $D^{\mathrm{t}}$. The vertical distances at difference percentile ranks between 0 and 100 were computed from laser point clouds, with an equal interval of 1.0. The first percentile closest to the top of the rice plant was confirmed, and the vertical distance corresponding to the first percentile ($D^{\mathrm{t}}$) was considered to be located at the top of the rice plant. This did not differ significantly between the various plots, and the standard deviation in $D^{\mathrm{t}}$ was $\sim 4\mathrm{cm}$ over the 2 years. Thus, it appears that the effect of planting geometry on vertical distance at the first percentile was not considerable in this study.

In the next step, the position of the bottom of the plant ($D^{\mathrm{b}}$) was identified. Three percentile ranks (95th, 80th, and 70th) were chosen as the closest percentiles to the bottom of the rice plant. Figure 7 shows that the vertical distance at the 95th percentile rank is the most similar to $D_G$, whereas the vertical distance at the 70th percentile rank shows the greatest difference. For the 80th percentile rank, the vertical distance approximates $D_G$ in the first 40 or 50 days but diverges from $D_G$ as time progresses (Fig. 7).

Fig. 7

Vertical distance at various percentile ranks. (a–c) Results from 2013 and (d–f) from 2014. Columns show results from (a and d) densely planted plots, (b and e) moderate density plots, and (c and f) sparse plots. Red dashed lines show the position of the ground surface after harvesting the crop.

Finally, relative vertical distances ($rD$) were calculated, as the distance between the top and bottom of the rice plant (Fig. 6). Figure 8 illustrates a correlation in all cases when $rD$ is plotted against $H$. The regression lines show a high coefficient of determination ($r^2$); however, these results depend on the input data and the chosen bottom position. The strongest correlations were identified with $p^{\mathrm{b}}=95$, with $r^2>0.86$ [Figs. 8(c), 8(f), and 8(i)], and the value of $r^2$ decreases with smaller $p^{\mathrm{b}}$. In 2013, data from the last four observations from plot 4 differ significantly from the general trend, leading to the low $r^2$ of 0.48 when $p^{\mathrm{b}}=70$ [Fig. 8(a)].

Fig. 8

Plots of $rD$ against measured rice plant height. The rows show the results from (a–c) 2013, (d–f) 2014, and (g–i) a combination of the 2 years. The columns show the results with the reference position computed at (a, d, g) 70th, (b, e, h) 80th, and (c, f, i) 95th percentile ranks. Red dash-dotted lines show the regression lines.

It should be noted that narrow rice leaf blades, which can be smaller in width than the laser footprint size, lead to difficulties in detecting both the rice leaf tip and the individual rice leaves with the laser scanner. Therefore, $rD$ was always smaller than $H$, such that all regression lines have a slope greater than 1.0. The slopes differs between the two years; as cultivation conditions were kept the same, this could be attributed to a difference in the number of stems. A larger number of stems appears to correspond to steeper slopes, as in 2014 all regression line slopes are steeper than those of 2013. Hence, the number of stems should be considered carefully in future studies.

To evaluate the new method detailed here, estimated rice plant heights were compared with measured $H$. The smallest root mean square error ($\mathrm{RMSE}\approx 4\mathrm{cm}$) was achieved when the calculated plant bottom position was closest to the ground surface, when $p^{\mathrm{b}}=95$. For the 80th and 70th percentiles, the RMSE was 8 to 9 cm when data from both years were combined. All other positions of the plant bottom with $p^{\mathrm{b}}<70$ determined above the ground surface gave acceptable results ($\mathrm{RMSE}\le 10\mathrm{cm}$), although it should be noted that plant bottom positions were limited to above the 50th percentile. These results illustrate the capability of this new method to estimate plant height from 3-D laser point cloud data. Furthermore, we identified a constant bias between $H$ and $rD$ with $p^{\mathrm{b}}=95$ [Figs. 8(c), 8(f), and 8(g)]. By forcing the slopes of the regression lines of these cases to 1.0, $H$ was found to be 16.0 cm greater than $rD$.