In order to increase the production of any agricultural system, activities such as crop monitoring for assessing growth, stresses, pests, fertiliser, water, nutrient condition and irrigation are all required [1
]. In addition to this, post-harvesting handling, such as tree pruning, has also been shown to be beneficial for enhancing yields [3
]. Pruning includes cutting and trimming of branches, and as such it affects the structural attributes of tree crops. Pruning of fruit trees promotes new growth [4
], makes manual fruit-picking easier, and increases light interception, which is important for fruit quality [5
]. Tree pruning has also been shown to have implications for crop harvest and nutrition, pest and disease control, soil protection and irrigation strategies [7
]. Increasing flowering, fruit colour, soluble solids concentrations and flower bud formation, and decreasing titratable acid content are other benefits linked to pruning of fruit trees [3
However, tree pruning is a costly practice, especially if done using manual labour, which is usually the case for small orchards [11
]. Often, tree crown reduction goals are set to optimise pruning [12
], but the assessment to determine whether these goals have been achieved is generally based on manual measurement or empirical models, which are time-consuming and potentially inconsistent [4
]. Hence, there is a need for more efficient and consistent tree crop pruning monitoring strategies that can be applied in a consistent manner at the orchard level.
Remote sensing is ideally suited for monitoring tasks and has the capability of providing multi-temporal information on tree structure, and changes in these, over time [13
]. However, as many orchards are relatively small (1–50 ha) [8
], the use of high spatial resolution satellite and airborne imagery quickly becomes cost-prohibitive [14
]. The rapid development of Unmanned Aerial Vehicles (UAVs) and miniaturised sensors in the last decade is now offering an alternative to more traditional satellite and airborne-based remote sensing [15
]. This is largely due to the fact that UAVs are light-weight, low-cost, suitable for autonomous data collection, and highly deployable, allowing remotely sensed imagery to be collected at any time for smaller areas (<1 km2
), even in cloudy conditions [4
]. However, limiting factors include high wind speeds, rain and spatial coverage, when operating UAVs.
The mapping of tree structural parameters such as tree height and crown size provides key indicators for plant growth, biomass, yield, as well as for assessing pruning practices [4
]. As high spatial resolution imagery is required for assessing the structure of individual tree crowns, UAV imagery is ideally suited for this task. UAV imagery has been used in many different agricultural settings [14
], but only to a limited extent for tree crops. For instance, measurements of plant height is a common UAV application because of the ability to produce photogrammetrically derived Digital Surface Models (DSM) from Structure-from-Motion of overlapping photos with different view angles of the same feature [4
]. Plant height can be used to model biomass, which is crucial information for predicting crop yield [20
Most UAV-based tree crop mapping applications have focused on olive trees [4
]. These studies, which all achieved high correlations between field and image derived structural parameters, focused on deriving chlorophyll and leaf area index using a six-band multi-spectral Tetracam [23
], and map tree height, crown diameter, volume and area using RGB and multi-spectral imagery [4
]. Jimenez-Brenes et al. [4
] used UAV-based RGB imagery to map tree position, projected crown area, height and volume of olive trees before, after and one year after pruning. Tree crown structure was assessed for trees subjected to three different kinds of pruning techniques, i.e., mechanical, adapted and traditional. It was found that trees subjected to more aggressive pruning experienced much more subsequent vegetative development for the three studied pruning techniques.
In forestry applications, local maxima identification techniques have been used for identification and delineation of individual tree crowns [26
], and these techniques have also been applied successfully by [18
] using UAV imagery for assessing tree height and crown diameter. Recently, segmentation approaches and geographic object-based image analysis (GEOBIA) of high spatial resolution imagery have become the preferred means for delineating individual tree crowns, due to the additional information available in the classification/delineation process in terms of shape, context, class-related and multi-scale information [29
]. Because of the suitability of GEOBIA for information derivation from high spatial resolution imagery [34
], several UAV-based studies have recently started to incorporate GEOBIA into their image processing workflow [4
]. Diaz-Varela et al. [24
] used an object-based supervised classification and the Classification and Regression Tree (CART) algorithm for delineating olive trees. Torres-Sanchez et al. [25
] developed a simple object-based mapping approach based on thresholding olive tree crown DSM values in relation to neighbouring ground for tree crown delineation. The object-based mapping approach developed by [4
] was based on that in [25
]. This new approach relied heavily on the generated DSM for identifying the tree crown boundaries. However, as photogrammetrically point cloud generated DSMs often do not align perfectly with tree crown edges, as shown in this research, incorrect measurements of crown area and volume may be obtained if these edges are not adjusted based on spectral information. In addition, Jimenez-Brenes et al. [4
] reported that only 80% (512) of the trees within the orchard were correctly photo-reconstructed on the three image dates, which highlights the need to include spectral information as well in the object-based tree crown delineation process rather than heavily relying on the generated DSM.
There is scant literature on the use of UAVs for mapping the influence of pruning on tree crop structural development and change [4
]. Existing research has only focused on olive trees, which are spectrally and structurally different from other tree crops such as lychee, citrus, mango, macadamia, and avocado [4
]. Hence, existing methods for mapping structure and pruning effects based on olive trees may not be feasible for other tree crops. To expand upon this lack, this research paper explores a novel and innovative approach to assess changes in tree structure, i.e., tree crown perimeter, width, height, area and Plant Projective Cover (PPC), using multi-spectral UAV derived imagery collected before and after pruning. To do this, we focus our study on the analysis of a commercial lychee orchard in eastern Australia. As no UAV-based mapping of lychee tree structure and pruning was identified in the literature, and as lychee trees are spectrally and structurally different to olive trees, this research provides new insight into the mapping of tree crops.
An object-based tree crown delineation approach is introduced, representing an additional novelty that addresses limitations of other UAV-based studies e.g., [4
], which relied heavily on the use of a DSM for tree crown delineation. This research shows that these existing approaches are not feasible for lychee tree crown delineation and a novel approach incorporating spectral and context information is introduced. The approach also addresses the issue outlined by [25
] of mapping trees with overlapping crowns. Also, additional tree crop structural parameters are mapped, i.e., crown perimeter and plant projective cover compared with [4
]. Given the lack of any systematic evaluation of how UAV-based data acquisition configurations, including varying flying heights, affect image derived information extraction of tree structure, a secondary objective was to assess any variations in the results as a function of various flying heights (30 m/4.1 cm Ground Sampling Distance (GSD), 50 m/6.5 cm GSD and 70 m/8.8 cm GSD).
Characterizing the impacts of pruning on tree structural parameters is required to inform and enhance the management of orchards and improve crop productivity. We present an innovative and novel approach that exploits multi-spectral UAV imagery to measure tree structural differences pre- and post-pruning, and apply this to a small commercial lychee orchard. The developed GEOBIA approach was found to be particularly useful for delineating individual tree crowns and deriving object shape and spectral and textural information for correlation with field-based measurements of tree structure. It was also found that existing GEOBIA approaches designed for tree crown delineation of olive trees were not feasible for delineation of lychee tree crops and required additional spectral and context information. Mapping tree structure and pruning effects may hence require specialized approaches to be applied for different tree crops. The multi-spectral imagery was found to accurately assess pre- and post-pruning tree crown structure, including tree crown perimeter, area, width, height and PPC. Tree crown perimeter was most accurately mapped at a flying height of 70 m, while tree crown width measurements were similar at all three flying heights. Tree height was most accurately mapped at a 30 m flying height, as larger flying heights affected the accuracy of the derived DSM and DTM. Imagery collected at 70 m height produced slightly higher correlation with field measured PPC for most predictor variables.
These results highlight that despite the compromise in accuracy of tree height estimates (0.1936 m RMSE as opposed to 0.3568 m), a flying height of 70 m may be the best choice for assessing pre- and post-pruning tree structural differences to gain efficiency in terms of flight duration, area coverage, and image processing time, without losing a significant amount of information. As an additional benefit, the proposed UAV-based approach is likely to reduce costs (compared with manual assessment) and increase consistency compared to traditional field-based estimates. Future research should focus on collecting and analysing similar data for other orchard sites and for trees grown under different conditions, e.g., different tree ages, tree varieties, climatic conditions, and pruning strategies, to test if the developed approach can be applied more generally and the results remain consistent with broader application.