2.1.

Study Area

The surveys were conducted in the area of the city of Jiansanjiang (N 47°15′21″, E 132°37′43″) in Heilongjiang Province in the far northeast of China (Fig. 1). The province with a continental monsoon climate is an important basis for agricultural products in China.⁵² Situated in the east of Heilongjiang, the Sanjiang Plain covers almost $100,000{\mathrm{km}}^2$, which is about $\frac{1}{4}$ of the provinces’ total area. Cold and dry winters and short but warm, humid summers are characteristic for the middle temperate and humid climate of the Sanjiang Plain, which is marked by the East Asian summer monsoon.^53,54 The field campaigns were carried out at two sites: (1) a field experiment at the Keyansuo experiment station (Fig. 1) where various treatments for the cultivation of rice were applied and (2) a farmer’s field (Fig. 1) with a conventional management.

Fig. 1

Location of the study sites in China (modified from Ref. 27).

2.2.

TLS Measurements

The chosen survey period of late June to July captures the key vegetative stage of the rice plants, when the stem elongation process takes place. Remarkable differences in plant development occur due to the increase of tillers and plant height during this stage. For the monitoring approach, three campaigns were carried out on both fields, which were each time conducted on two consecutive days. The campaign dates are given in Table 2.

Table 2

Dates of the scan campaigns and corresponding phenological stages.

For all field campaigns, the terrestrial laser scanner Riegl VZ-1000,⁵⁶ provided by Five Star Electronic Technologies located in Beijing, was used. The scanner operates with the time-of-flight technique, where the time between transmitting and receiving a pulsed laser signal is measured. The time is used for calculating the distance between the sensor and target. Parallel scan lines are achieved with a rotating multifacet polygon mirror and the rotation of the scanners head itself, which implies a wide field of view, up to 100 deg in vertical and 360 deg in horizontal direction. The infrared laser beam has a high precision of 5 mm and an accuracy of 8 mm. Apart from a measurement rate of up to $\mathrm{122,000}\text{points}/\mathrm{s}$, long range distance measurements of up to 1400 m are possible. In addition, the system is capable of an online full-waveform analysis and according to echo digitization.

Additionally, a digital camera, Nikon D700, was mounted on the laser scanner. From the recorded RGB-photos, the point clouds gained from the laser scanner can be colorized, resulting in three-dimensional RGB point clouds and the corresponding surfaces can be textured. The camera was connected to the scanner for adjusting the camera settings and ensuring an accurate alignment between the devices. During the acquisition, the whole system was remotely controlled with the RiSCAN Pro Software on a notebook, linked via an LAN connection.

During the campaigns at the field experiment, the scanner was fixed on a tripod which raised the sensor up to 1.5 m above ground. Where possible, a small trailer behind a tractor was used for achieving a greater height of about 3 m (Fig. 2). The study area was scanned from nine scan positions for capturing all fields of the Keyansuo experiment station and minimizing shadowing effects. Although the data from all positions were used for the analysis, four of them were of major importance, as they were located closely to the investigated N field experiment. Two positions were accomplished without the trailer at the north edge and two positions with the trailer at the south edge of the field. The whole setting is shown in Fig. 3.

Fig. 2

Overview of the investigated field experiment from scan position six (Fig. 3). On the right side the scanner with the tripod mounted on the small trailer can be seen (taken: July 4, 2011).

Fig. 3

Experimental design and scan positions of the field experiment. Number in the plot represents rice variety ($1=\text{Kongyu}$ 131; $2=\text{Longjing}$ 21); treatment (1 to 9 in Table 1); and repetition (1 to 3).

On the farmer’s field (Fig. 4), the scanner was also mounted on the tripod. Accordingly, the sensor height was about 1.5 m above ground. Due to a limited access on the small dikes between the plots, it was impossible to use a trailer or to reach any lifted position. The field was scanned from seven scan positions. For this study, the whole field is divided into the overall field and two intensively investigated units (W and E in Fig. 4). In order to get a high resolution for the latter ones, four scan positions were placed at their corners. Twelve thin, long bamboo sticks per unit were stuck in the ground, placed in an equally spaced grid, which can be detected in the point clouds and located in the field to ensure the spatial linkage to other plant parameter measurements. An additional grid with 28 measurement points represented by bamboo sticks was placed in the overall field.

Fig. 4

Experimental design and scan positions of the farmer’s field.

Common tie points in all scans of each site are required to enable the merging of all scan positions in the postprocessing. Therefore, high-reflective cylinders, which can be easily detected by the laser scanner, were fixed on ranging poles built upon the dikes between the fields.⁵¹ The reflectors had to be detected from all scan positions for computing the spatial relationship between all positions of the scanners and the cylinders. In the first TLS campaigns, the position of each pole was marked in the fields. All scans of each date from a respective field can be merged together by re-establishing the ranging poles for the other campaigns.

2.3.

Manual Measurements

During the whole vegetation period, manual measurements were performed at both sites for monitoring the development of the rice plants. The phenological stage of the plants and, more precisely, the steps in the plant development are defined here by the BBCH-scale.⁵⁷ The abbreviation BBCH was derived from Biologische Bundesanstalt (German Federal Biological Research Centre for Agriculture and Forestry), Bundessortenamt (German Federal Office of Plant Varieties), and Chemical Industry, who funded the development of the scale. For both sites, the BBCH-values at the campaign dates are given in Table 2. The similar values confirm the comparable phenological development of the rice plants on the field experiment and the farmer’s field.

Corresponding to each TLS campaign, plant heights were manually measured. On the field experiment, 8 to 10 hills per plot were regarded. Each hill consists of 4 to 6 rice plants.⁵⁵ In both intensively investigated units of the farmer’s field, the heights of four hills around each bamboo stick were measured.

As mentioned above, destructive biomass sampling was performed several times during the vegetation period at the field experiment. Samples were taken from both varieties for the respective three repetitions of treatment 1 to 5 ($n=30$). Due to the small plot size, it was not feasible to take additional samples corresponding to the TLS campaigns. As the dates of sampling differ from the TLS campaign dates, the biomass values were linearly interpolated.

On the farmer’s field, the four hills around each bamboo stick in the two intensively investigated units were destructively taken after the TLS measurements for measuring the biomass ($n=24$). After each campaign, the grid of bamboo sticks was moved for having an undisturbed area around the bamboo sticks for the following campaign. Furthermore, in the overall field destructive samplings were taken around the mentioned 28 bamboo sticks on the 26th of June, as an additional independent validation dataset (Fig. 4). For all samplings, the cleaned above ground biomass was dried in a compartment dryer and weighed after dehydrating. The average dry biomass $\mathrm{per}{\mathrm{m}}^2$ was calculated, considering the number of hills $\mathrm{per}{\mathrm{m}}^2$, which was counted in the field corresponding to each sampling.

2.4.

TLS Data Processing

The general workflow for the postprocessing of the TLS data is shown in Fig. 5. It consists of the (1) registration and merging of all point clouds, (2) filtering and extraction of the area of interest (AOI), (3) spatial, and (4) statistical analyses, considering the manual measurements.

Fig. 5

General overview of the workflow for the postprocessing of the terrestrial laser scanning data.

The first steps (1 and 2) were carried out directly in Riegl’s software RiSCAN Pro, which was already used for the data acquisition. First of all, the scans from the respective three campaigns were imported into one RiSCAN Pro project. The registration of the scan positions was executed with an indirect registration method, based on the above mentioned high-reflective cylinders acting as tie points. With an automatic algorithm, corresponding tie points between the scan positions can be found. After the registration, the datasets still showed alignment errors, due to nonoptimal reflector positioning, imprecise re-establishing of the ranging poles, or instabilities during the measurements. A further adjustment was applied to minimize these errors. RiSCAN Pro offers the Multistation Adjustment, which is based on the iterative closest point algorithm.⁵⁸ The position and orientation of each scan position were modified in multiple iterations for getting the best fitting result for all of them.

The point clouds still contained noise, caused by reflections on water in the field or on small particles in the air. Thus, a further filtering based on the reflectance, measured for each point during the data acquisition, was performed. Points under a certain reflectance value, regarded as noise, were removed. As the reflectance value depends on the distance from the sensor to the field as well as other factors, the critical value was slightly different for each scan.

Subsequently, all point clouds of each respective date were merged to one dataset and the AOI was manually extracted. For an easier orientation and the distinction between fields and dikes, the point clouds were previously colorized from the recorded pictures. The AOI was further separated for each date and plot to have a common spatial base. A filtering scheme was used for selecting the maximum points and determining the crop surface. Finally, those filtered point clouds were exported as American Standard Code for Information Interchange (ASCII) files for spatial and statistical analyses.

2.5.

Spatial Analysis

ArcGIS Desktop 10 by Esri was used for constructing the CSMs and following spatial analyses. The exported ASCII files were converted to vector point data and interpolated with the inverse distance weighting (IDW) algorithm for receiving a raster representing a digital surface model with a consistent spatial resolution of 1 cm. IDW is a deterministic, exact interpolation method, and retains a measured value at its sample location.⁵⁹ Hence, the method is suitable for preserving the accuracy of measurements with a high density, like the TLS point clouds.

A common reference surface is required for the calculation of the plant heights. Usually a high-accurate digital elevation model (DEM) is used; achievable from scanning the AOI without any vegetation.⁵¹ Since it was not possible to obtain such data in this study, another method was applied: the lowest parts in the point clouds from the first date, accordingly containing the least dense vegetation, were manually selected for interpolating a DEM surface representing the real ground. As it can be seen in Fig. 6, the rice plants were small enough for clearly identifying points on the ground and the water height in the irrigated field was $<4\mathrm{cm}$ at this stage. Hence, enough ground points at the edges and around the hills remained for interpolating a DEM.

Fig. 6

Corner of the field experiment showing the least dense vegetation (taken: June 21, 2011).

Finally, the CSMs, introduced in Ref. 51 for plant growth monitoring at field level, were established for each date. The application of CSMs is presented in Ref. 60. A CSM represents the crop surface with high spatial resolution at one campaign date, gained from the merged and filtered point cloud. As shown in Fig. 7 (Ref. 61), CSMs are used for determining the actual plant height for a given growing stage. Therefore, the DEM from the first acquisition date, representing the ground, is subtracted from the CSM, representing the crop surface. The result is the plant height above ground with the same spatial resolution as the CSMs, which is visualized in maps of plant height. By subtracting a CSM of an earlier date from a CSM of a later date, the plant growth between the dates can be spatially measured, e.g., CSM_2 minus CSM_1 in Fig. 7. The spatial patterns of the plant growth are visualized in maps of plant growth. In the following, plant growth is always defined as the spatio-temporal difference in height.

Fig. 7

General concept of crop surface models (CSMs) (Ref. 61).

2.6.

Statistical Analysis

The statistical analyses were performed in Microsoft Excel 2010. For a better visualization, diagrams were plotted in OriginPro 8.5 by OriginLab. The plant height values, calculated pixel-wise for the CSMs, were averaged for each plot of the field experiment and each circular buffer area with a radius of 1 m around the bamboo sticks for the farmer’s field, respectively. The plots of the field experiment were previously clipped with an inner buffer of 60 cm for preventing border effects. Additionally, the manually measured heights were averaged for each plot ($n=54$) or area around the bamboo sticks ($n=24$). These values were compared with the mean plant heights derived from the CSMs of the same respective spatial feature for evaluating and validating the laser scanning results.

2.7.

Biomass Regression Model

As mentioned before, the problem of the nondestructive estimation of crop biomass on field level is not solved yet, while indirect approaches successfully used plant height as predictor. In order to investigate the correlation between plant height and biomass of rice plants, a regression model was derived from the results of the field experiment. The transferability of the model to the farmer’s conventionally managed field was validated by comparing the simulated and measured biomass. As mentioned, different rice varieties were cultivated on both test sites. The two rice varieties from the field experiment were combined in the regression model to ensure an adequate number of measured values ($n=90$) and attain a reasonable mean value for the transfer to the farmer’s field. The combination of the different treatments covers the influence of the varying amount of used fertilizer. Previously, three test models were established for testing the general concept. For each test model, the regression equation from two repetitions of the field experiment were used for simulating the biomass of the third repetition.

The workflow can be structured in five steps:

3.1.

Spatial Analysis

After the described data processing of the captured TLS point clouds, the CSMs for each date and both sites were generated and the plant heights were calculated pixel-wise. Thus, the following spatial and temporal patterns and variations within one CSM and between different CSMs can be obtained. As an example, Fig. 8 shows 12 maps of plant height derived from the CSMs. For all three TLS campaign dates, the maps of two repetitions of the same fertilizer treatment for Kongyu 131 (plots 162 and 163) and Longjing 21 (plots 261 and 262) are shown. All field experiment plots and the whole farmer’s field are represented in the way it is shown in Fig. 8.

Fig. 8

CSM-derived maps of plant height for four selected plots of the field experiment (left: Kongyu 131; right: Longjing 21, marked in Fig. 3).

The linear structure of the rice plant rows within the plots is detectable at the first campaign but disappears later due to plant development. Regarding the field experiment, slight differences between the rice varieties can be identified. The latter difference is captured by examining the mean plant height per plot, which shows higher values for Longjing 21. The averaged difference between the varieties increases over the time (4, 5, and 10 cm). In addition, the plant growth is observable, which is determined as height difference between consecutive CSMs and visualized as maps of plant growth. In Fig. 9, maps of Kongyu 131 (plot 133) and Longjing 21 (plot 232) for both time intervals are shown as an example. In both intervals, the growth patterns are almost homogeneous within the plots for both varieties. According to the increasing height difference between the varieties over the time, the growth values are higher for Longjing 21.

Fig. 9

Maps of plant growth for two selected plots field of the field experiment, derived from the difference between two consecutive CSMs (left: Kongyu 131; right: Longjing 21, marked in Fig. 3).

The mean plant heights calculated from the CSMs were validated against the averaged manually measured plant heights for each plot or area around the bamboo sticks for verifying the results. Figure 10 shows the difference between these values for the first campaign on the field experiment. The variance is quite small. About 40% of the plots show a difference of $<2\mathrm{cm}$, further 45% differ by 2 to 5 cm, and just 15% show a higher error, reaching the maximum at about 10 cm. The mean difference between all CSM-derived and manually measured plant heights is about 3 cm for the plots of the field experiment and about 9 cm for the buffer areas around the bamboo sticks of the intensively investigated units of the farmer’s field, each with a standard deviation of about 5 cm.

Fig. 10

Difference between the averaged manually measured plant heights and the CSM-derived mean plant heights for each plot for the first campaign of the field experiment.

3.2.

Statistical Analysis

The CSM-derived and the manually measured plant heights, averaged for each plot or buffer area, were used for executing correlation and regression analyses. Common statistical values are shown in Table 3. For each campaign and both sites, the mean heights are quite similar. The differences between the mean CSM-derived and manually measured plant heights are about 3 cm for the field experiment and 9 cm for the farmer’s field. The standard deviation within each campaign is about 5 cm for both sites. All minimum values are lower for the CSM-derived mean plant heights, whereas the maximum values are more similar. All values and the resulting regression lines for both fields are shown in Fig. 11. The correlation coefficients are very high for each field (both $R^2=0.91$).

Table 3

Mean CSM-derived and manually measured plant heights for both fields.

Fig. 11

Regression of the mean CSM-derived and manually measured plant heights for the field experiment ($n=162$) and the intensively investigated units on the farmer’s field ($n=72$).

3.3.

Biomass Regression Model

As mentioned earlier, for destructive biomass sampling on the field experiment, only treatments 1 to 5 were considered, with the main different levels of N fertilization. Hence, the number of samples and accordingly the averaged plant height values differ from the comparison of the height measurement methods (Table 3). In the intensively investigated units of the farmer’s field, the biomass was taken around all bamboo sticks. For each campaign, the mean value, standard deviation, minimum, and maximum were calculated for the plant height and dry biomass (Table 4). The mean plant heights of each campaign are similar for the field experiment and the farmer’s field, with a difference of $<5\mathrm{cm}$. In contrast, the averaged dry biomass values of the field experiment are 20% to 30% lower than the values of the farmer’s field at the second and third campaign.

Table 4

Mean CSM-derived plant heights and biomass values.

The general concept of the biomass regression model was validated with three test models. Therefore, the regression equation achieved from two repetitions of the field experiment was used for calculating the biomass of the respective third repetition. The mean deviations of the simulated values from the actual measured values are 3%, 16%, and 19%.

Considering now both fields, the relationship between the mean plant height and dry biomass is visualized in a scatter plot (Fig. 12). The lower biomass values of the field experiment are also visible, but the linear correlation is similar for both sites. The regression equation from the field experiment ($y=11.06x-224.18$) was used for deriving the biomass regression model. Following, the biomass on the intensively investigated units of the farmer’s field was estimated with the model, based on the CSM-derived plant heights. Figure 13 shows the simulated biomass, with once the standard deviation calculated for each campaign, and the actual measured values. The reliability of the established model is supported by the strong correlation between simulated and measured values ($R^2=0.90$). The mean difference between the values is $90\mathrm{g}/{\mathrm{m}}^2$ (about $\frac{1}{4}$ of the mean measured dry biomass), with a standard deviation of $80\mathrm{g}/{\mathrm{m}}^2$.

Fig. 12

Regression of the mean CSM-derived plant height and dry biomass for the field experiment ($n=90$) and the intensively investigated units on the farmer’s field ($n=72$).

Fig. 13

Theoretical biomass simulated with regression model and the measured values for the intensively investigated units on the farmer’s field (each: $n=72$).

The regression model was validated and the transferability to any point in time within the observation period was checked with the biomass measurements on the overall farmer’s field. As the increase in plant height over the time is almost linear in the observation period, a linear function achieved from all CSM-derived plant height values was used for interpolating the plant heights for the 26th of June. The theoretical biomass was estimated with the regression model and compared to the measured values. Table 5 gives the basic statistics for the simulated and measured biomass values. The mean difference between both values is $15\mathrm{g}/{\mathrm{m}}^2$ (about 20% of the mean measured dry biomass), with a standard deviation of $36\mathrm{g}/{\mathrm{m}}^2$.

Table 5

Biomass values for the overall farmer’s field.