Agricultural University of Athens
Investigating the application of spectral imaging and AI in precision horticulture (agriculture)
Abstract
dc:descriptionSpectral imaging and Artificial Intelligence in precision horticulture are commonly used for a variety of applications ranging from disease detection to quality estimation. However, most of the available solutions require deep understanding of software engineering and they mostly focus on disease detection and post-harvest applications. This study aimed to (i)develop Artificial Intelligence models utilizing spectral data that can identify different fertilisation levels, (ii)develop Artificial Intelligence models utilizing spectral data capable of identifying plant water deficit, (iii)compare the performance of traditional machine learning algorithms with novel user-friendly Auto Machine Learning (AutoML) techniques and(iv)evaluate the feasibility of developing a generalisation-capable AI model utilizing spectral data. Towards that end, a progressive methodology was implemented to gather data and develop the required methodologies. During the first year spectral data from broccoli plants that were submitted to different fertilization schemes were collected, while during the second year spectral data were collected from broccoli plants that were submitted to different irrigation schemes. Besides spectral data during both years, dry matter measurements were conducted not only for broccoli but also for apple, leek and mushroom. Finally, during the third year, all AI methodologies were developed, and AI experiments were conducted. Throughout these three years, this study evaluated and compared traditional Artificial Intelligence approaches with AutoML systems towards water/ acclimation and nutrient deficiency stress identification using spectral imaging. For both types of stress, AutoML was compared to a traditional machine learning approach (Partial Least Squares – Discriminant Analysis) used for classification of spectral data. On both occasions, data were captured with the use of the IMEC snapscan Visible Near Infrared hyperspectral camera (400-900nm). Moreover, the study aimed to investigate generalisation capabilities of spectral imaging and how each step of the “traditional” pre-processing pipeline followed for spectral data modelling affects its generalisation capabilities and performance. The pipeline, followed by both stress experiments and tested for its generalisation capabilities, consisted of the following steps: Outlier removal→Data smoothing→Data Scaling→Feature selection→Feature Extraction→Modelling. Techniques used for various steps across the pipeline included Savitzky Golay smoothing, Standard and Min Max scaling, f and mutual info regression for feature selection, umap, autoencoeder and PCA feature extraction and various machine learning models ranging from linear to quadratic models and reaching the complexity of neural networks. For identifying nutrient stress, the AutoML system achieved results that were superior to those achieved by the Partial Least Squares – Discriminant Analysis (PLS-DA) algorithm. Namely, an accuracy of 0.72 was achieved when using the CIELAB colour space and 0.94 when combining the CIELAB colour space with the hyperspectral data. When using the hyperspectral data standalone, the results improved (accuracy 1.00), this performance was achieved using all 150 bands, however, it is worth mentioning that the same performance was maintained even when using the single statistically most important wavelength (874 nm). On the other hand, for the identification of water/acclimation stressed plants, both the Automated Machine Learning system and the PLSDA algorithm achieved an accuracy of 1.00 across all stress levels. Finally, hyperspectral imaging has proven capable of generalizing across different fruits and vegetables, achieving an (RMSEP) = 0.0137 using the Partial Least Squares Regression algorithm on a 10x5-fold cross-validation protocol. Overall, the results suggest that Automated Machine Learning can achieve and even outperform traditional spectral imaging machine learning approaches for detecting water/ acclimation and nutrient deficiency stress. Moreover, the use of the CIELAB colour space for training the models failed to match the performance of using the spectral data, while combining the two did not lead to a performance increase compared to just using the spectral data. The evaluated techniques used for preprocessing affected the two regression algorithms, Automatic Relevance Determination Regression (ARD) and Partial Least Squares Regression (PLSR) in a different way, with the best performance achieved when the complete pipeline was used. Furthermore, feature selection appeared to be the preprocessing technique that had the most negative impact on the linear regression performance when used standalone. However, its use to fit a quadratic transformation of the features was found to be a good compromise. Overall, the pipeline using either ARD algorithm or PLSR algorithm showed strong generalization and performance in the Visible Near Infrared wavelength based dry matter content estimation across diverse crops. To conclude, the use of Spectral imaging with AutoML solutions may provide a user friendly and cost-effective method for detecting plant stress, while at the same time, spectral imaging model generalisation can be achieved provided that a universal data acquisition protocol is followed, with promising results even without following complex data preprocessing pipelines. Finally, it should be noted that the present study did not examine the lag factor. It is assumed that with the prevalence of water or nutrient scarcity, the change in spectral data will not be automatic.
Degree
thesis:*- Grantor dc:publisher
- Agricultural University of Athens
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Μalounas, Ioannis
- Μαλούνας, Ιωάννης
Subjects
dc:subject × 32- Γεωργία ακρίβειας
- Τεχνητή νοημοσύνη
- Φασματοσκοπία
- Αυτόματη μηχανική μάθηση
- Τηλεπισκόπηση
- Λίπανση
- Άρδευση
- Αβιοτικό στρες
- Προσαρμοστικό στρες
- Γενίκευση
- Περιεκτικότητα σε ξηρή ύλη
- Πολυωνυμική παλινδρόμηση
- Εξαγωγή χαρακτηριστικών
- Precision agriculture
- Artificial intelligence
- Spectral imaging
- Remote sensing
- AutoML
- Fertilisation
- Irrigation
- Abiotic stress
- Acclimation stress
- Generalisation
- Dry Matter Content
- Polynomial Regression
- Feature extraction
- Άλλες Γεωπονικές Επιστήμες
- Γεωπονικές Επιστήμες και Κτηνιατρική
- Γεωπονική μηχανική
- Other Agricultural Sciences
- Agricultural and Veterinary Sciences
- Agricultural Engineering
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58619