LAI estimation across California vineyards using sUAS multi-seasonal multi-spectral, thermal, and elevation information and machine learning

Abstract

In agriculture, leaf area index (LAI) is an important variable that describes occurring biomass and relates to the distribution of energy fluxes and evapotranspiration components. Current LAI estimation methods at subfield scale are limited not only by the characteristics of the spatial data (pixel size and spectral information) but also by the empiricity of developed models, mostly based on vegetation indices, which do not necessarily scale spatiality (among different varieties or planting characteristics) or temporally (need for different LAI models for different phenological stages). Widely used machine learning (ML) algorithms and high-resolution small unmanned aerial system (sUAS) information provide an opportunity for spatial and temporal LAI estimation addressing the spatial and temporal limitations. In this study, considering both accuracy and efficiency, a point-cloud-based feature-extraction approach (Full Approach) and a raster-based feature-extraction approach (Fast Approach) using sUAS information were developed based on multiple growing seasons (2014–2019) to extract and generate vine-scale information for LAI estimation in commercial vineyards across California. Three known ML algorithms, Random Forest (RF), eXtreme Gradient Boosting (XGB), and Relevance Vector Machine (RVM), were considered, along with hybrid ML schemes based on those three algorithms, coupled with different feature-extraction approaches. Results showed that the hybrid ML technique using RF and RVM and the Fast Approach with 9 input variables, called RVM-RF_Fast model, performs better than others in a visual and statistical assessments of the generated LAI being also computationally efficient. Furthermore, using the generated LAI products in the quantification of energy balance using the two-source energy balance Priestley-Taylor version (TSEB-PT) model and EC tower data, the results indicated excellent estimation of net radiation (Rn) and latent heat flux (LE), good estimation of surface heat flux (G), and poor estimation of sensible heat flux (H). Additionally, TSEB-PT sensitivity analysis performed by regenerating LAI maps based on the generated LAI map (from − 15% of the original LAI map to + 15% with a 5% gap) showed that LAI uncertainty had a major impact on G, followed by evapotranspiration partitioning (T/ET), H, LE, and Rn. When considering the annual growth cycle of grapevines, the impact of LAI uncertainty on the T/ET in the veraison stage was larger than in the fruit set stage.

Appendices

Appendix

Variables from the full feature extraction approach mentioned in Feature extraction approaches

See Table 4

Table 4 Variables considered in this study for the Full approach gained from the spectral, thermal, DSM, and point cloud data

Important variables supporting each ML algorithm mentioned in LAI model via the fast approach and Model selection for LAI estimation

See Tables 5, 6 and 7

Table 5 Identified important variables for each ML algorithm gained from the full feature extraction approach

Table 6 Identified important variables for each ML algorithm gained from the fast feature extraction approach

Table 7 Metrics for the evaluation of the RVM-RF_Full and RVM-RF_Fast models in training and test datasets

The ANOVA and Tukey tests mentioned in Model selection for LAI estimation

The ANOVA and Tukey tests are used again to check whether the mean of different ML models, considering the input variables from both the Full and Fast approach, is statistically the same as the mean from the ground LAI measurements. Table 8 shows the results from both the ANOVA test and the Tukey test. The most important question at this point is whether the null hypothesis (the same as previously explained) is accepted when comparing the mean value of the ground measurements with the mean value estimated by the hybrid RVM-RF algorithm for both situations (input variables obtained from the Full approach and the Fast approach).

When evaluating the eight experiments based on the inputs from the Full approach, Table 5 reports different results compared with that shown in Fig. 7. Only three experiments, “RVM-XGB_Full-Cover,” “RVM-RF_Full,” and “RVM-XGB_Full-Weight,” report the same mean value as the ground measurement. Evaluating the eight experiments based on the inputs from the Fast approach,

Table 6 shows a similar signal to Fig. 8. Except for the experiment called “RVM-XGB_Fast-Gain,” all of the combination models can provide the same mean estimation as the ground LAI measurements. Six experiments accepted the null hypothesis, which is more than when the Full approach is considered as the feature-extraction approach.

See Table 8

Table 8 ANOVA and Tukey test for the ML algorithms considering a few variables as the input (color figure online)

During the evaluation of each experiment, the most important point is verified: the null hypothesis is accepted for the hybrid RVM-RF algorithm for both situations. The difference shown in this table suggests that the ANOVA and Tukey tests are necessary as a confirmation tool to double-check the candidate ML algorithm.

Relevant locations mentioned in Model selection for LAI estimation for both RVM-RF_Full and RVM-RF_Fast

Considering the characteristic of the RVM algorithm providing RVs, which have non-zero weights, from the training dataset (Tipping 2001; Fletcher 2010; Torres-Rua et al. 2012; Khader and McKee 2014), this section explores RV information based on the hybrid ML algorithm (RVM-RF). Since this structure provides accurate LAI estimation when considering the inputs from the Full approach and efficient LAI estimation when considering the Fast approach, the same process is explored under both situations: RVM-RF_Full and RVM-RF_Fast.

RVM-RF_Full model

Figure 12 shows the geographic distribution of the training dataset (red circles) and the RVs (green circles). The green circles obtained from the RVM algorithm illustrate where the ground measurements are more important. For SLM, the RVs concentrate on the right side of the field, and one RV is located on the bare soil area. RIP 760 shows a pair of parallel lines and more RVs from the south line. RIP 720 recognizes only two positions on the northeast corner as important locations, and one of them provides the RV three times (another is two times). BAR recognizes around seven out of 49 positions as important locations, and most of them come from the south block. Taking more ground LAI measurements at the green-circle sites shown in Fig. 12 (a large variability is shown at these locations) would be recommended in the future if the combination of the RVM-RF_Full model is adopted.

See Fig. 12

Fig. 12

The geographic distribution of the RVs from the RVM-RF_Full model

RVM-RF_Fast model

Similar to Figs. 12 and 13 shows the geographic distribution of the training dataset (red circles) and the RVs (green circles). One difference compared to Fig. 12 is that no position provides RVs three times. SLM identifies more RV locations. RIP 720 and RIP 760 each provide six RV positions. BAR identifies five positions corresponding to five RVs.

See Fig. 13

Fig. 13

The geographic distribution of the RVs from the RVM-RF_Fast model