Back to search

Agricultural University of Athens

Yield prediction in processing tomato crop, through precision agriculture practices

Abstract

dc:description

This research is designed to make a valuable contribution to the field of precision agriculture, by exploring the potential of state-of-the-art technologies and techniques for yield prediction in processing tomato crop. The core aim was to develop and evaluate a robust methodology incorporating cutting-edge technologies, remote sensing data, and sophisticated analytical techniques like machine learning and statistical analysis. The primary objective is to improve the accuracy and dependability of yield predictions at local and regional levels. This was achieved through a progressive approach implemented yearly, utilizing non-invasive methods to track the crop's biological cycle and refine predictive yield models. Over the course of this study, a progressive methodology was implemented to gather data and refine methodologies. It commenced with a systematic literature review focusing on yield predictions within precision agriculture, to offer an extensive overview of the latest advancements in this domain. Simultaneously, pilot activities were conducted over three years. Ten pilot fields were chosen to integrate proximal, aerial, and satellite measurements with yield assessments in the initial two years, primarily exploring the correlation between crop yield and NDVI (Normalized Difference Vegetation Index), a widely used indicator. This phase aimed to uncover similarities between satellite technology, UAS (Unmanned Aerial System), and proximal sensors concerning crop yield evaluation. Conducting an extensive comparison of satellite, airborne, and proximal technologies, their individual strengths, and limitations were emphasized within the precision agriculture context. During the second year, alongside the detailed investigation of specific fields using satellite, UAS, and proximal sensors at the field scale, an expanded study involved 108 fields at the regional scale, incorporating satellite data analysis. This phase aimed to assess NDVI and four additional vegetation indices (VIs) in predicting crop yield. Time series data comprising five VIs at a regional scale were deployed to explore the relationship between these indices and the critical phenological stages of the crop. Furthermore, machine learning techniques were applied to VI data collected through satellite images at different growth stages to evaluate their predictive performance for yield in industrial tomato fields. Specifically, AutoML algorithms and statistical analysis were utilized to gauge the correlation between yield and VIs retrieved from satellite datasets. The transformed data was utilized to train and test both statistical and machine learning algorithms, encompassing linear and nonlinear regression models, along with ensemble methods based on decision trees. The analysis encompassed various regression models such as ordinary least squares (OLS), Theil-Sen, and Huber, as well as tree-based methods, support vector machines (SVM), and automatic relevance determination (ARD). This broader scope aimed to deepen the understanding of crop yield estimation based on the findings from the initial systematic review. During the third year, a more intricate approach was adopted, focusing on the evaluation of spectral bands derived from satellite imagery. Each band's individual performance in predicting crop yield was assessed, allowing for a comprehensive evaluation of their unique contributions to overall yield estimation accuracy. This granular analysis provided deeper insights into the importance and impact of each spectral band in refining the precision of crop yield predictions. The performance of both statistical and machine learning models was assessed to gain profound insights into the most efficient growth stages and VIs for precise yield prediction. The findings of this study indicated substantial similarity between UAS and Sentinel-2 data, especially in the later stages of the crop's phenological cycle, implying a heightened agreement as the crop matured. The study identified a moderate correlation between proximal sensor data and UAS/satellite datasets, with discrepancies attributed to variances in measurement wavelengths, and the specific focus area of each platform. Regarding the relationship of the VI dynamics in relation to the crop's phenological cycle, the lowest mean values for all VIs were observed shortly after transplanting, marked by limited canopy cover and exposed soil between rows. Notably, as the season progressed, canopy cover percentage gradually increased, particularly at the midpoint of the season when processing tomato crop reached their peak vigor just before reallocating sugars to their fruits. Regarding yield prediction, the proximal sensor showed higher explanatory power during the initial canopy growth phase, while UAS and Sentinel platforms improved their performance as canopy coverage expanded. The satellite platform demonstrated superior performance during the crucial flowering stage, emphasizing its effectiveness in elucidating yield variability. The study employed basic statistics, notably the Pearson correlation coefficient, highlighting the VIs' optimal performance during specific growth stages. Additionally, machine learning algorithms were integrated to enhance yield prediction accuracy rigorously evaluated through regression analysis and a 5-fold cross-validation procedure. The research identified NDVI, RVI and SAVI as the most effective VIs for yield predictions, achieving high R² values and low RMSEs, particularly 90 days after transplanting. Ensembles comprising two regressors emerged as the optimal choice for enhanced predictive accuracy. Remarkably, band combinations of Band 4, Band 8, and Band 12 stood out. The Red Edge/NIR bands displayed notable performance, especially within the 80 to 90 days post-transplanting window, exhibiting the strongest correlation with yield. Overall, the findings suggest that UAS and satellite sensors demonstrate greater accuracy in predicting crop yield towards the end of the season and offer increased precision during the later stages of development. In contrast, proximal sensors showcase correlations at earlier stages of crop growth. Combining data from multiple sensors and growth stages can enhance prediction accuracy. Using VI and spectral band data in conjunction with machine learning techniques may represent a more effective and economically efficient method for predicting tomato yield.

Degree

thesis:*
Grantor dc:publisher
Agricultural University of Athens
Year dc:date
2023

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Darra, Nikoleta
  • Δάρρα, Νικολέτα

Subjects

dc:subject × 18

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/55505

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

Darra, Nikoleta; Δάρρα, Νικολέτα. Yield prediction in processing tomato crop, through precision agriculture practices. Agricultural University of Athens, 2023. http://hdl.handle.net/10442/hedi/55505