*4.1. Plant Height*

Depending on the type of crop, the accuracy for plant height retrieval varied between 3.4 cm for winter wheat, 7.4 cm for sugar beet, and 12 cm for potato (Figure 3). Previous research efforts on determining plant height via LiDAR have focused on winter wheat, where Jimenez-Berni et al. [2] found an accuracy of 1.7 cm for a ground-based LiDAR system. Madec et al. [7] compared plant height derived from a LiDAR scanner on a ground vehicle with estimates from SfM analysis on RGB-UAV images. The latter method provided an accuracy of 3.47 cm for winter wheat plant height, proving that the methodology presented within this paper proves just as accurate. Homan et al. [17] showed that using a terrestrial laser scanner (TLS), the RMSE was 2.7 cm for wheat, which is comparable to the achieved accuracy by the RiEGL RiCOPTER UAV LiDAR system. This, and previous research, shows that using a UAV-based LiDAR system can be used to accurately assess plant height for winter wheat fields.

However, the accuracy drops for crops with a more complicated canopy structure, such as potato and sugar beet, where it proves to be harder to provide a plant height estimate with high accuracy. This starts with the procedure for field measurements. Winter wheat is an easy-to-measure crop in the field due to its erectophile structure and a homogeneous pattern in height. Potato and sugar beet and their planophile leaf angle distribution prove to be harder to measure within the field, which is further complicated for potatoes since it is grown on ridges. Therefore, it is visually harder to determine the highest point of a specific location. We tried to account for this by averaging field measurements over the three highest points in close proximity, but we cannot ignore the fact that there are uncertainties in the field measurements (Table 3).

In particular, potato proved to be a difficult crop because it grows on ridges, showing the need for a good digital terrain model (DTM) of these ridges. A flight should be performed just after planting the potato to get a good DTM, which can be used to normalize the 3D LIDAR point cloud of the flights further during the growing season. This research started when the potato plants were already developed, resulting in a closed canopy, making it impossible to create a good DTM. This could be one of the reasons for the worse performance of potato height estimation.

As a result, the standard deviations for potato are generally larger than for the other two crops (Table 3). Especially on 25 June when the potato plants were fully grown, it proved more difficult to get accurate measurements.

#### *4.2. Biomass*

The 3DPI algorithm proved to estimate the biomass of winter wheat (NRMSE = 13.9% and R2 = 0.82) and sugar beet (NRMSE = 17.47% and R<sup>2</sup> = 0.68) well, showing slightly higher accuracies compared to the paper of [2], where an NRMSE of 19.82% was achieved for winter wheat. It should be noticed that the early growth stages are not included in this study, which partially explains the differences in NRMSE. For winter wheat, similar results were obtained in [18], where a linear relation was also examined between LiDAR-derived plant height and biomass, reporting an R<sup>2</sup> of 0.88.

Possible improvements in the amount of biomass for winter wheat could be made using the dry weight of the plant material. Especially during the ripening phase, the water content of winter wheat decreases [19], resulting in a mismatch between the 3DPI algorithm and the measured fresh weight biomass. The 3DPI method is developed on plant structure and does not account for the changing water content within the winter wheat. Also, research by [20] showed that LiDAR laser scans proved to be useful in determining leaf water content. This shows that there is an influence on the return signal, which is influenced by the water content that has not been accounted for now. This effect depends on the architecture, size, and density of the crop and, as such, is crop dependent.

The 3DPI algorithm proved to be unsuitable for estimating the biomass of potato. The 3DPI algorithm was developed for winter wheat and not tested on other crops so far. Probably the main reason is the very dense canopy, resulting in very few points in the lower parts of the canopy, which resulted in a small number of returns lower in the canopy. These points underneath the canopy are needed to create a complete height profile relative to the 30 cm high ridges where potatoes are grown on, resulting in a better representation of the above ground biomass. The reason the algorithm works better for sugar beet, which also shows a relatively dense canopy, is that sugar beet has larger inter-row spacing which leads to more hits from larger scan angles.

The biomass prediction could possibly be improved for areas with a low points density by changing the p\_i layer size in the 3DPI algorithm. In this study, a p\_i layer size of 50 mm was chosen for two reasons. Firstly, because our point density was lower than of the experiment done by [2], choosing a p\_i of 10 mm would result in layers with no points in it. This could possibly influence the 3DPI algorithm negatively. Thus, increasing the p\_i layer size resulted in an increase in points per p\_i layer, but how this influences the final 3DPI indicator should be further researched. Secondly, choosing a larger p\_i resulted in a faster computation time, which was useful for the large files we had.

Possible further research could be done using machine learning, where the 3DPI may be one of many predictors to estimate biomass. Using, for example, a PLS regression as mentioned in [21], or using machine learning approaches like [22], where a deep convolutional network was used to

predict biomass using RGB imagery. Or the machine learning approach of [23], where they used a range of spectral indices as predictors to estimate above ground biomass with an NRMSE of 24.95%. Combining these different methodologies with a predictor such as the 3DPI indicator could result in a better prediction of biomass.
