Back to search

Agricultural University of Athens

Extracting and assessing plant features spatio-temporally by means of 3D sensing in apple trees

Abstract

dc:description

Knowledge of tree structural characteristics is gaining importance in advanced fruit growing, especially when implementing precision horticulture techniques. Accurate extraction and processing of tree structure is critical for precision horticulture, whereas spatio-temporal crop monitoring aim to optimize farm management practices such as pruning, thinning, selective harvest, chemical and fertilizer spraying. Growers have traditionally obtained this information by using manual sampling techniques, which tend to be laborious, spatially sparse, expensive, inaccurate and prone to subjective biases. Three-dimensional (3D) sensing, offers new possibilities for a more efficient agricultural practice with the use of highly advanced technological instruments and methodologies. Essential reasons contributing to this possibility include the increasing computational processing power of computers, the decrease in cost and size of electronics, the increase of efficiency in laser emitting diodes. Therefore, the demand for advanced knowledge of individual crops, requires robust and accurate data processing techniques needed by the grower to support precision agriculture protocols.In the present study, precision horticulture techniques were applied in apple orchards. The aim of this study was to extract 3D structural information of apple trees spatially and to associate this information with soil properties, microclimate, tree physiology, yield and fruit quality. To meet this end, a light detection and ranging (LiDAR) system was mounted on a commercial tractor to scan tree rows of the orchard.The first contribution of this thesis is the development of a tree segmentation methodology to detect individual trees in 3D domain, based on stem position in high density orchards. The approach focuses on trees trained as slender spindled, which presents unique challenges for segmentation algorithms due to their intertwined nature. Initially, the accuracy of the LiDAR system assessed using a metal-box with known dimensions in field conditions. Tree stem position of each tree was estimated based on a bivariate point density histogram. Each tree was segmented using a cylindrical boundary projected around the estimated stem positions. Therefore, geometrical parameters of the segmented tree point clouds were estimated and compared with manual measurements. The assessment of the LiDAR system considering the metal box exhibited enhanced error due to points with enhanceddistance between laser aperture and the object. The coefficient of determinations were achieved between the measurements and estimated data from the segmented point clouds in terms of tree height, stem diameter and canopy volume.The second contribution considers the implementation of structural parameters in a water balance model for observing their interaction with microclimate and soil properties over growing season. In particular, leaf area (LA) and height of trees, estimated LiDAR system, with daily weather data and soil variables are utilized to estimate the daily tree water needs different areas of soil electrical conductivity (ECa). The results revealed that tree geometry interacts with ECa. Furthermore, the utilization of 3D plant and microclimate informationincreased the spatial resolution of the water balance model, and its accuracy. Furthermore, the estimated LA by the LiDAR presented high correlations with ECa at different areas. In parallel, a correlation was observed between the ECa and total available water content in the root zone. Whereas, the implementation of LA information from different ECa areas pointed to the reasonability of spatially resolved water balance, allowing to estimate daily tree water needs.Moreover, the study emphasized in the spatio-temporal estimation of tree geometry in 3D using the LiDAR system at different growth stages. Firstly, the spatial impact of ECa and orchard landscape in the growth of tree volume was analyzed, indicating high variation over the season and its significant dependence with the soil properties. The LiDAR system was used to monitor the whole field repeatedly over the growth season. The estimated canopy volume related negatively with ECa areas (low, mid, and high) of the orchard, indicating that tree growth was approximately 30% higher in low ECa compared to mid and high ECa areas. These results suggest that the canopy growth varies spatially, whereas the terrestrial laser scanning offers a feasible method for implementation in decision support systems, in precision horticulture.Additionally, LiDAR-based estimation of LA was achieved using LA of defoliated trees with the corresponding points per tree. In parallel, linearity and backscattered reflectance of LiDAR, which suggested in fruit detection methodology, were utilized to segment the woody parts from points per tree (PPT) at different growth stages, improving the relationship with LA. The spatial responses of LA were investigated with the ECa and yield over the season. Whereas, a k-nearest neighbor model, was built to predict the fruit fresh mass (FM) and soluble solids content by means of soil ECa and LA. These results confirm spatio-temporal variation of canopy growth.Furthermore, the results suggest the value of LA data in decision support systems aimed at optimizing orchard management practices, particularly addressing the impact of orchard management on fruit quality.Estimation of LA was improved when points of woody parts were removed from points per tree, providing higher adjusted coefficient of determination at 55 days after bud break. ECa and LA were positively correlated with fruit size. Motivated by these correlations, k-nearest neighbor models were built to predict fruit quality at harvest. Validation was performed by applying the model of one year on data of the subsequent year. Based on geoposition and ECa, classification accuracy in the test set validation was 39.7 %, whereas classification with geoposition and LA resulted in increased accuracy by 63.3 %. Results confirm spatio-temporal variation of canopy growth as was expected.Tree structural parameters can impact on fruit size, number and quality parameters. Therefore, in the final part of the study, fruit detection algorithms using the LiDAR system were proposed to estimate fruit diameter at different growth stages. Trees were scanned pre and after defoliation at each individual date. Fruit segmentation was based on threshold values of LiDAR calibrated backscattered reflectance and the geometrical features of linearity and curvature. The high complexity of tree structure highlights the need to set up an assessment pipeline considering fruits from defoliated trees as reference points. Therefore, evaluation of fruit number and position founded on foliated trees is performed by comparison against the reference labels on defoliated. The results show the high efficacy of LiDAR to localize and estimate fruit size in 3D domains, overpassing imaging systems limitations susceptible to highly variable external factors that lead to significant appearance variations. An analysis of LiDAR calibrated backscattered reflectance of tree elements was performed, exhibiting discrete values for leaves, woody parts and apple fruits, respectively. The diameter of apple fruits estimated from the foliated trees was related to the reference values based on the perimeter of the fruits, revealing high adjusted coefficient of determination at DAFB120. When comparing the results obtained on foliated and defoliated tree’s data, the estimated number of fruit’s on foliated trees at 120 days after full bloom 94.8% of the ground truth values. The algorithm resulted in maximum values of 88.2% precision, 91.0% recall, and 89.5 F1 score at 120 days after full bloom.The components of this thesis have been designed to address to future machine vision challenges under the local variable environmental conditions identified in high density apple orchards. Regarding tree segmentation, theaccuracy of the proposed methodology should be evaluated at different perennial trees and planting systems. Future research should further investigate the implementation of 3D temporal plant information (e.g. LA, volume) from high density orchards in variable rate applications such as irrigation, fertilisation. The inner relationship of fruit quality and LA should be observed at different cultivars and climates to improve decision making in orchards. Estimating fruit number over the growing season can optimise crop load and selective harvesting. In the future, the detection accuracy of the proposed methodology should be assessed under different apple cultivars and LiDAR sensors potential and architecture.

Degree

thesis:*
Grantor dc:publisher
Agricultural University of Athens
Year dc:date
2021

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Tsoulias, Nikolaos
  • Τσούλιας, Νικόλαος

Subjects

dc:subject × 14

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/51617

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Tsoulias, Nikolaos; Τσούλιας, Νικόλαος. Extracting and assessing plant features spatio-temporally by means of 3D sensing in apple trees. Agricultural University of Athens, 2021. http://hdl.handle.net/10442/hedi/51617