*4.3. General Models*

The accuracy of the prediction of plant height using the general prediction model is 10.1 cm, which is lower than those for the crop-specific prediction models of sugar beet (7.4 cm) and winter wheat (3.4 cm).

The general model for biomass prediction displays a strong influence of sugar beet on the total prediction model (Figure 7). The general model has the best predictive power at k = 3, showing this influence. Sugar beet had the highest predictive power for k = 3, with the other two crops at k = 1 (Figure 5). The reason for this over-representation of sugar beet is due to larger sample size of 15 samples compared to nine for winter wheat and the better performance of the sugar beet biomass model compared to that of potato. Combining these two explains why at k = 3, the model has the highest predictive power. When combining data from different crops, extra attention should be paid to include an equal sample size of the multiple crops. Therefore, using a crop-specific model to estimate crop height from UAV-LiDAR data proves to increase the prediction accuracy. However, depending on the accuracy requirements, a general model could provide plant height estimations with an accuracy of 10.1 cm and a biomass estimation with an accuracy of 17.07%. The results of this research show that crop-specific biomass models have a higher retrieval accuracy; however, further research is required to evaluate if general models can be relevant for mixed cropping systems.

Again, the trade-off is visible between a general biomass model, compared to the specialized models for each individual crop. Furthermore, the model now has to be fitted through three kinds of crops, which are completely different from each other. It was also indicated that different K-values were needed to get a good fit of the specialized biomass models. A certain combination of 3DPI indicator and measured biomass for potato does not necessarily correspond to the same combination of the 3DPI indicator and measured biomass for winter wheat. Therefore, limiting this general applicability of the model, the 3DPI appears to be crop sensitive. Possible other factors, such a height statistics derived from the LiDAR 3D point cloud as analyzed in [21] or the spectral vegetation indices of [23], could be helpful to make the 3DPI method less crop sensitive, which is something for further research.

#### *4.4. Influence of Flight Altitude and Speed on Biomass and Plant Height Estimation*

Plant height estimation from 3D points can be accurately achieved when the data is acquired at high speed and relatively high altitude. The LiDAR-derived plant height from DAY3-HA-HS, which was acquired at 92.68 m.a.g.l. and a speed of 7.39 m/s, still shows the same patterns as with DAY2-LA-MS and DAY2-LA-LS, but the general plant height is lower (Figure 9). The LiDAR-derived plant height underestimates the plant height compared to DAY2-LA-MS on average with 4.03 cm. This underestimation is still smaller than the sugar beet plant height model error of 7.2 cm, showing that acquiring plant height could be done with these settings. Therefore, speeding up data collection and covering larger areas is feasible.

Applying the generic model for sugar beet on the 3D points clouds from the flights DAY2-LA-MS, DAY2-LA-LS, DAY3-HA-HS, and DAY3-LA-LS showed that between DAY2-LA-MS, DAY2-LA-LS and DAY3-LA-LS, there is no real difference in LiDAR-derived plant height. This indicates that there is no real benefit in slowing down to 2.53 m/s or decreasing flying height. The different flight specifications do result in different point densities and point distributions. However, this results mostly in more points underneath the canopy, which has only a minor influence on the top of canopy points. To increase accuracy, a DTM could be created before plants start to grow, decreasing the dependence on getting enough returns underneath the canopy.

For biomass estimation, the under the canopy points are important, where for DAY3-HA-HS, a large under estimation is made averaging 858.8 g/m2 compared to the biomass prediction for DAY2-LA-MS. There are not enough points to represent the full height of the plant. The non-normalised accuracy of the winter wheat biomass model was 415.8 g/m2. Showing the inaccuracy of data acquired at 92.68 m.a.g.l. and 7.39 m/s compared to that of DAY2-LA-MS, acquired at 24.16 m.a.g.l. and 4.17 m/s.

Changing flight characteristics shows some areas with patches where the biomass for wheat is underestimated (Figure 10). These patches appear to result from areas that are not covered with any flight lines and almost no perpendicular flight lines. It appears to be crop-specific as these patches do not appear for sugar beet and potato. A possible reason why it affects winter wheat more is that laser pulses from the LiDAR UAV penetrate more easily due to its erectophile leaf structure compared to the planophile leaf structure of potato and sugar beet. Sofonia et al. [24] showed that for a LiDAR-based application, a cross-flight pattern worked best. A better cross pattern and closer flight lines could, therefore, possibly improve biomass estimation in general because laser pulses are acquired from multiple directions increasing the chance of hitting parts of the plant underneath the canopy and remove these unwanted patches. For an accurate biomass prediction, a homogeneous point density is needed.

To allow a good estimation of biomass and crop height over the growing season, it is necessary to keep the flight patterns consistent for all flights. Our results show that biomass and crop height estimation errors increase if the point density and point distribution vary from the circumstances used to calibrate the models. To determine the optimal flight pattern, altitude, and speed, more intensive experiments should be done.

#### *4.5. Outlook*

Our results show the possibilities for UAV LiDAR to estimate plant phenotypes, such as biomass and plant height, accurately and with high throughput. This fills the gap between slower but highly accurate tractor-based LiDAR systems and high-throughput but less-detailed manned airborne systems as indicated by [25]. Therefore, providing a solution to situations where a quick analysis of the field is required, tractor-based solutions are not suitable. Furthermore, this study shows that crop trait monitoring can be done throughout the season, using the same model trained on data from the whole season. Moreover, this research showed that under uncontrolled conditions, relevant biomass and plant height estimations can be made, which is marked as a bottleneck in the review paper of [26]. When using UAV-LiDAR for high-throughput estimation of plant height and biomass [27], time should be invested in creating a detailed and structured flight plan. A flight plan would ideally consist of a cross-flight pattern and it will cover the whole field, where flights lines are created alongside the boundaries.

This research furthermore showed that there are limitations to the biomass estimation for certain crops and that models developed for a specific crop cannot directly be used for other crops, and generic models should be used with care. For a fully operational approach, an effort should be made towards combining LiDAR with hyperspectral data, as mentioned by [22,23,26], so models can be trained for a range of crops. This will increase both the accuracy and general applicability of high-throughput biomass estimation models. Also, alternative methods for canopy height determination from UAV-based 3D point cloud datasets have recently been published [28].

#### **5. Conclusions**

Retrieval of plant height and biomass using UAV-LiDAR proved to be possible for sugar beet and winter wheat. While for potato both plant height and biomass estimation proved to be hard due to the complex canopy structure and the ridges on which potatoes are grown. For plant height, the data acquisition can be performed relatively fast and at high altitudes increasing opportunities for high-throughput approaches. However, for accurate biomass estimates, the flight conditions (altitude, speed, location of flight lines) should be kept constant. The higher acquisition speed, compared to, for

example, tractor-based LiDAR systems, means that UAV-LiDAR can be used to assess large areas and can provide data quickly. Creating a reliable general model to predict biomass for different crops results in a lower accuracy, especially when crops with a dense canopy like potato are included. To increase the predictive performance of both LiDAR-derived plant height and biomass, a clear DTM should be created before germination of the plants. This accurate DTM could then be used to accurately perform the height normalization of the 3D LiDAR point cloud.

**Author Contributions:** Conceptualization, H.B., L.K., J.t.H.; methodology, H.B., L.K. and J.t.H.; software, H.B., J.t.H.; validation, H.B., L.K., J.t.H.; formal analysis, J.t.H.; investigation, H.B., J.t.H.; resources, L.K.; writing—original draft preparation, J.t.H.; writing—review and editing, H.B., L.K., J.t.H.; visualization, J.t.H.; supervision, H.B., L.K.; project administration, L.K.; funding acquisition, L.K. All authors have read and agreed to the published version of the manuscript.

**Funding:** The access to the RiCOPTER was made possible by the Shared Research Facilities of Wageningen University & Research. This work was supported by the SPECTORS project (143081), which is funded by the European cooperation program INTERREG Deutschland-Nederland.

**Acknowledgments:** We would like to acknowledge all the people that assisted with this research, where without them, the field campaign would have taken much longer. Therefore, special thanks to the following persons without whom this research could not be performed. Berry Onderstal for assisting with the biomass sampling. Marcello Novani, Kim Calders, Tessa Rozemuller, Tim Jak, Sophie Stuhler and Hannah Stuhler for assisting in the field campaigns. And Matthijs ten Harkel for correction and editing of the paper manuscript.

**Conflicts of Interest:** The authors declare no conflict of interest.
