Context-Aware AutoML for Accurate Wheat Disease Detection Muhammad Uzair1 , Radwa ElShawi1,* and Stefania Tomasiello1,2 1 Institute of Computer Science, University of Tartu, Estonia 2 Department of Industrial Engineering, University of Salerno, Fisciano, Italy Abstract Timely detection and management of crop diseases are crucial for food security and agricultural productivity. Traditional methods, which rely on manual inspection, are often slow and prone to human error. With the rise of diseases like stripe rust in wheat, there is a growing need for efficient automated detection methods. This paper proposes a novel classification strategy that leverages Automated Machine Learning (AutoML) in combination with advanced feature engineering techniques. We develop a scalable framework that detects stripe rust by extracting comprehensive statistical features from images, distinguishing disease symptoms from healthy crops. To enhance feature quality, we employ Context-Aware Automated Feature Engineering, which iteratively generates meaningful features to capture subtle patterns in the data. Our method achieves 95.35% accuracy on the RustNet dataset, significantly outperforming the state-of-the-art ResNet-18 model, which achieved 85.2% accuracy. These findings highlight the potential of AutoML and automated feature engineering to revolutionize disease detection in agriculture, offering a cost-effective alternative to traditional deep learning methods that require extensive computational resources and expertise. Keywords AutoML, disease detection, feature engineering, large language models 1. Introduction have led to a growing interest in automating the ML process. This has spurred the development of Automated Machine The Food and Agriculture Organization (FAO) forecasts a Learning (AutoML) techniques [9, 10], which simplify the 0.9% increase in global cereal utilization for 2023/24 com- creation of ML pipelines by automating stages such as data pared to the previous year. Wheat, as the most widely culti- preprocessing, feature engineering, model selection, and vated crop globally, is essential to agriculture, with rising optimization. By reducing the need for manual interven- consumption expected in regions like the European Union, tion, AutoML streamlines the development of effective ML China, India, the UK, and the US [1]. However, wheat faces models, making advanced disease detection more accessible significant threats from diseases and pests, causing sub- and efficient. stantial annual losses, roughly one-fifth of global yield [2]. This study introduces a novel approach that integrates Among these, wheat stripe rust, caused by Puccinia stri- AutoML with context-aware feature engineering for the de- iformis f.sp.tritici, is particularly devastating, leading to se- tection of stripe rust in wheat. We extract comprehensive vere yield losses [3]. This disease has become increasingly statistical features from UAV-captured images and refine prevalent worldwide, posing serious risks to food security them using Context-Aware Automated Feature Engineering and agricultural sustainability. (CAAFE), a feature engineering method designed for tabu- Traditional methods for monitoring wheat rust rely on lar datasets [11]. CAAFE leverages a large language model manual visual inspection, which is time-consuming, labor- (LLM) to iteratively generate additional semantically mean- intensive, and costly, making it impractical for large-scale ingful features based on the dataset description, enhancing agriculture [4]. Recent advancements in imaging technolo- the discriminatory power of the features. These refined gies, especially the use of Unmanned Aerial Vehicles (UAVs), features are then processed using the Tree-Based Pipeline offer a promising alternative for automated crop disease Optimization Tool (TPOT) [12], an AutoML framework that detection. UAVs can capture high-resolution images of large automates the selection, optimization, and construction of fields, enabling more efficient and accurate disease monitor- classification models. Our proposed framework was rigor- ing [5, 6]. This technology, combined with advanced image ously evaluated on the publicly available RustNet dataset [6], processing techniques, holds great potential for timely and achieving a remarkable accuracy of 95.35%. This represents precise identification of disease outbreaks. a substantial improvement over the state-of-the-art ResNet- Effective and timely monitoring of yellow rust is essential 18 model, which attained an accuracy of only 85.2% [6]. for both disease management and sustainable crop produc- tion. Accurate disease mapping facilitates the judicious application of fungicides and enhances breeding programs 2. Related Works by identifying resistant wheat varieties [6]. Machine learn- ing (ML) techniques play a crucial role in achieving high The application of Unmanned Aerial Vehicles (UAVs) for precision in disease detection, focusing on extracting rele- plant disease detection has garnered substantial interest, vant features from images and utilizing classifiers such as leading to the development of advanced methodologies that Neural Networks, Random Forest, Support Vector Machines, integrate image processing with Machine Learning (ML) and K-Nearest Neighbors [7, 8]. However, the complex- algorithms. Gu et al. [13] introduced a method for detect- ity and manual effort required to develop these ML models ing and quantifying the severity of narrow brown leaf spot, a common disease affecting rice crops. The methodology Published in the Proceedings of the Workshops of the EDBT/ICDT 2025 began with the extraction of color features and vegetation Joint Conference (March 25-28, 2025), Barcelona, Spain indices from UAV-acquired images. Pearson’s correlation * Corresponding author: Radwa ElShawi (radwa.elshawi@ut.ee). analysis was then employed to identify the four most sig- $ muhammad.uzair@ut.ee (M. Uzair); radwa.elshawi@ut.ee nificant features, which were subsequently used as inputs (R. ElShawi); stefania.tomasiello@ut.ee (S. Tomasiello) for support vector regression, achieving a high degree of  1234-5678-9012 (R. ElShawi); 0000-0001-8208-8285 (S. Tomasiello) Β© 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License accuracy in disease severity estimation. Attribution 4.0 International (CC BY 4.0). CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings In the field of wheat disease detection, Liu et al. [14] focused on identifying powdery mildew using UAV im- agery. They meticulously extracted textural features such as contrast, correlation, and variance, and applied Partial Least Squares Regression (PLSR) for comprehensive anal- ysis, yielding a nuanced understanding and quantification of the disease’s impact. Additionally, a study on monitor- ing wheat scab using UAV remote sensing [15] emphasized the value of texture features derived from multiple spec- tral bands. When combined with vegetation indices, these features provided extensive data for disease monitoring, with Support Vector Regression (SVR) demonstrating effec- tiveness in predictive analysis. Zhang et al. [16] utilized a combination of spectral and textural features to detect Fusarium Head Blight in wheat crops, employing Logis- tic Regression to highlight the critical role of feature-rich datasets in accurate disease classification and monitoring. Subsequent studies [17, 18] advanced this approach by inte- grating spectral, textural, and color features with various classification models, including Support Vector Machines Figure 1: Flowchart of the proposed framework (SVM) and Neural Networks (NNs). These studies high- lighted the significance of feature extraction techniques and the adaptability of ML algorithms in managing the complex datasets derived from UAV imagery. Furthermore, research with regular irrigation. Lemhi 66 cultivar in borders was on wheat yellow rust detection illustrated the interaction highly susceptible to stripe rust, with three inoculated bor- between traditional ML algorithms and Deep Learning (DL) ders and one non-infected border. Images were collected techniques [19, 6, 20]. While ML methods such as SVR, NNs, only from the borders in Field 2. and Random Forests demonstrated significant effectiveness, DL models have shown promising potential for enhancing 3.2. Data preprocessing accuracy and efficiency in disease detection tasks. Broadening the application beyond wheat, research on Our preprocessing phase involves several key stages: ini- UAV-based disease detection in rubber trees [21] and citrus tial image acquisition, conversion to grayscale, resizing, plants [22] demonstrated the broad applicability of UAV- and feature extraction. During feature extraction, we com- based disease detection techniques across different agricul- pute essential statistical measures, including mean, standard tural sectors. These studies emphasized the vital role of deviation, variance, correlation, energy, entropy, contrast, advanced image processing techniques and ML algorithms skewness, kurtosis, and homogeneity. in enhancing global food security by enabling effective dis- ease detection in a wide range of crops. 3.3. Context-Aware Automated Feature Engineering 3. Methodology Feature engineering is a critical component of machine learning, as it involves transforming raw input data into fea- Figure 1 illustrates the architecture of our approach that tures that can improve predictive performance [23, 24]. In consists of three main stages, including data preprocessing our approach, we leverage CAAFE, an automated machine (Section 3.2), automated feature engineering using CAFE learning technique specifically designed for tabular datasets. (Section 3.3), and model training and evaluation using Au- CAAFE employs an LLM to iteratively generate semanti- toML approach (Section 3.4). In the following subsections, cally meaningful features based on a detailed description of we explain the different building blocks of our approach. the dataset. This process not only generates Python code for creating new features but also provides explanations for 3.1. Dataset the relevance and utility of the generated features. CAAFE operates iteratively on both the training and val- In this study, we used a publicly available dataset, RustNet idation datasets, π·π‘‘π‘Ÿπ‘Žπ‘–π‘› and π·π‘£π‘Žπ‘™π‘–π‘‘ , along with a descrip- [6]. RustNet comprises 508 images categorized into two tion of the dataset’s context and features. In each itera- classes: disease and no disease. Among these, there are 281 tion, CAAFE constructs a prompt that includes detailed images depicting instances of disease and 227 images with- information about the dataset and the specific feature en- out any disease. RustNet is based on data collected from gineering task, which is then passed to the LLM. Based two experimental wheat fields were imaged in Pullman, WA, on this prompt, the LLM generates code to alter or create the US, in 2021. Field 1, located at Palouse Conservation new features. The generated code is executed on the cur- Field Station, comprised two winter wheat trials: one for rent datasets (π·π‘‘π‘Ÿπ‘Žπ‘–π‘› and π·π‘£π‘Žπ‘™π‘–π‘‘ ), producing transformed testing fungicides on ’PS 279’ variety and another for as- datasets (𝐷′ π‘‘π‘Ÿπ‘Žπ‘–π‘› and 𝐷′ π‘£π‘Žπ‘™π‘–π‘‘). An ML classifier is subse- sessing stripe rust resistance in 23 winter wheat cultivars. quently trained on 𝐷′ π‘‘π‘Ÿπ‘Žπ‘–π‘› and evaluated on 𝐷′ π‘£π‘Žπ‘™π‘–π‘‘. If Both trials had randomized designs with four replications, the classifier’s performance on 𝐷′ π‘£π‘Žπ‘™π‘–π‘‘ surpasses its perfor- planted on November 1, 2020. Urediniospores of P. stri- mance on the original π·π‘£π‘Žπ‘™π‘–π‘‘, the newly generated feature iformis were inoculated twice to induce disease. Field 2, at is retained, and the datasets are updated accordingly. If not, Spillman Agronomy Farm, housed spring wheat nurseries the feature is discarded, and the datasets remain unchanged. The prompt provided to the LLM includes semantic and Class Train Test Total descriptive information about the dataset, such as a user- disease 208 73 281 generated dataset description, feature names, data types, no_disease 172 55 227 the percentage of missing values, and random sample rows Total 380 128 508 from the dataset. Additionally, a template for the expected format of the generated code and explanations is included, Table 1 which improves the clarity and quality of the LLM’s out- Number of images in train and test split for the RustNet dataset put. To further enhance performance, chain-of-thought instructions guide the LLM through a series of intermediate reasoning steps, leading to more effective code generation. Baselines. Given the randomized nature of the experi- By utilizing CAAFE, we integrate domain knowledge into ments reported in [6], we conducted new experiments using the feature engineering process, all while maintaining in- the same computational setup as described in their study. terpretability and optimizing predictive performance. This Specifically, we employed ResNet-18, following the origi- approach offers a powerful and efficient method for gener- nal architecture and hyperparameters outlined in [6], and ating high-quality features in complex datasets, marking a initialized the model with pre-trained weights. promising advancement in machine learning research. CAAFE setting. We leverage the advanced capabilities of OpenAI’s language models, including GPT-3.5, as LLM 3.4. AutoML approach within the CAAFE framework [27, 28]. The integration of these powerful language models enables CAAFE to generate TPOT is an AutoML framework designed for constructing semantically meaningful features iteratively, enhancing the and optimizing machine learning pipelines for both classi- effectiveness of feature engineering. To ensure robust per- fication and regression tasks. It utilizes tree-based genetic formance and accuracy, we conduct ten feature engineering programming [25] to evolve pipelines by treating them as iterations using the CAAFE framework. Additionally, in individuals within an evolutionary algorithm. Each pipeline the iterative evaluation of code blocks, we employ TabPFN is structured as a tree, with its nodes categorized as either (Tabular Predictive Functional Network), as proposed by Primitives or Terminals. Primitives represent operators that Hollmann et al. [29], to assess the effectiveness of generated require input, such as machine learning algorithms needing features and their impact on model performance. data and hyperparameter values. Terminals, on the other TPOT setting. To ensure a fair comparison, an equal time hand, are constants that provide input to the Primitives. No- budget was allocated for both TPOT and ResNet methodol- tably, a Primitive can also act as input for another Primitive, ogy. Experiments were constrained to a 20-minute time limit. allowing for complex pipeline configurations. The evolu- This consistent time allocation ensures parity in computa- tionary process in TPOT operates by applying genetic oper- tional resources between the methods, enabling a thorough ations such as mutation and crossover to the pipelines. Mu- and unbiased evaluation of their respective performances. tation involves making small modifications, such as chang- The input to TPOT is a data matrix after performing the fea- ing a hyperparameter or introducing a new preprocessing ture engineering step from CAAFE. The hyperparameters step. Crossover, on the other hand, selects two pipelines for TPOT were configured with a set number of genera- that share common Primitives and allows them to exchange tions, specifically 10, and a population size of 100. The subtrees or branches. Once these operations are performed, resulting pipeline generated by TPOT, constrained by the each pipeline is evaluated and assigned a fitness score, which specified time budget, is a multi-layer perceptron classifier reflects its performance. This fitness score is used in the with a learning rate of 0.01 and regularization parameter of selection process to determine which pipelines should be 0.0001. The latter is a penalty term, constraining the size retained and evolved further in the next generation, ulti- of the weights [30]. The aim of such a strategy is to reduce mately leading to the creation of highly optimized machine overfitting and enhance the generalization ability of the NN. learning pipelines. Generally, these pipeline trees could be Hardware Resources. We conducted our experiments arbitrarily large. Nevertheless, extensive machine learning on a CPU environment. The CPU environment runs on Win- pipelines usually have downsides. Longer pipelines with dows 11 Pro 64-bit (10.0, Build 22621) with 16 core Intel(R) numerous hyperparameters can be challenging to fine-tune, Core(TM) i9-10885H Processor @ 2.40GHz,32 GB DIMM more prone to overfitting, complicate the understanding of memory, and 1000 GB SSD data storage. All the approaches the final model, and demand extended evaluation time, thus have been implemented in Python. slowing down the optimization process. Due to these con- Performance metrics. Since the classification problem siderations, a multiobjective optimization technique, NSGA- is being tackled in this study, the performance metrics used II [26], is employed. It assists in selecting candidates based are Accuracy, Precision, Recall, and F1-score. on the Pareto front, representing the balanced trade-off be- tween pipeline length and performance. 4.2. Results 4.2.1. Preprocessing 4. Experimental Evaluation We followed the preprocessing steps described in Section 3.2. 4.1. Experimental setup Regarding the conversion of the class associated with the image to a numerical equivalent, we adopted for RustNet Training and test. For a fair comparison, we adopted the dataset π‘‘π‘–π‘ π‘’π‘Žπ‘ π‘’ = 1 and π‘›π‘œ_π‘‘π‘–π‘ π‘’π‘Žπ‘ π‘’ = 0. same train-test split methodology as outlined in the refer- After the statistical features are extracted from images, enced study, allocating 70% of the RustNet dataset for train- the resulting feature set is normalized using min-max nor- ing and 30% for testing [6]. Detailed information regarding malization, where each feature has a value between 0 and 1. these splits is provided in Table 1. The general formula for min-max normalization is: Figure 2: Exemplary run of CAAFE on the RustNet image dataset. User-generated input is shown in blue, ML-classifier generated data in red, and LLM-generated code with syntax highlighting. The generated code contains a comment per generated/deleted feature following a template (Feature name, description of usefulness, features used in the generated code, and sample values). CAAFE improved the ACC on the validation dataset from 0.946 to 0.953 over 10 iterations, but only those improving ACC are shown. π‘₯𝑖 βˆ’ π‘šπ‘–π‘›(𝑋) π‘₯′𝑖 = π‘šπ‘Žπ‘₯(𝑋) βˆ’ π‘šπ‘–π‘›(𝑋) where π‘₯′𝑖 is the normalized value, π‘₯𝑖 ∈ 𝑋, 𝑖 = 1, 2, . . . , 𝑛 is the original value. 4.3. Feature Engineering A demonstration of CAAFE using the RustNet dataset is il- lustrated in Figure 2. User inputs are highlighted in blue, ML- classifier-generated data in red, and LLM-generated code is presented with syntax highlighting. The code includes com- Figure 3: ResNet-18 Accuracy for RustNet dataset ments for each generated feature, adhering to a predefined template in CAAFE’s prompt. This template comprises the feature name, its utility description, the features utilized in the generated code, and sample values for these features. 4.4. AutoML The retained generated features from CAAFE after 10 itera- The results, evaluated using TPOT and ResNet-18, are de- tions include ’mean_variance_ratio’, calculated as the mean tailed in Table 2. For TPOT, two variants are considered: the divided by the variance, and ’contrast_energy_ratio’, com- baseline TPOT and TPOT with Context-Aware Automated puted as the contrast divided by the energy. Incorporating Feature Engineering (CAAFE), referred to as TPOT (FE). The the features generated by CAAFE into TPOT improved the comparative results demonstrate that both variants of our accuracy from 93.02% achieved using TPOT alone on the proposed frameworkβ€”TPOT and TPOT (FE)β€”outperform validation dataset to 95.42%, as shown in Table 2. the baseline ResNet-18 model. TPOT achieved an accuracy of 93.02%, which was further enhanced to 95.35% with the integration of CAAFE. In contrast, ResNet-18 achieved a pathogen, Food Security 12 (2020). doi:10.1007/ lower accuracy of 85.2% on the same dataset, highlighting s12571-020-01016-z. the superior performance of our proposed approach. The [4] J. su, C. Liu, X. Hu, X. Xu, L. Guo, W.-H. Chen, Spatio- limited number of epochs achieved within the allocated temporal monitoring of wheat yellow rust using uav time budget highlights the substantial computational effort multispectral imagery, Computers and Electronics in required. Agriculture (2019). doi:10.1016/j.compag.2019. 105035. Dataset Model Accuracy Precision Recall F1-Score [5] D. Basurto-Lozada, A. Hillier, D. Medina, D. Pulido, TPOT 93.02 92.99 92.90 92.80 S. Karaman, J. Salas, Dynamics of soil surface temper- RustNet TPOT (FE) 95.35 95.79 94.85 95.22 ature with unmanned aerial systems, Pattern Recog- nition Letters 138 (2020). doi:10.1016/j.patrec. ResNet-18 85.20 86.13 86.54 86.15 2020.07.003. Table 2 [6] Z. Tang, M. Wang, S. Michael, K.-H. Dammer, X. Li, Performance of TPOT and ResNet-18 on RustNet dataset R. Brueggeman, S. Sankaran, A. Carter, M. Pumphrey, Y. Hu, X. Chen, Z. Zhang, Affordable high through- put field detection of wheat stripe rust using deep learning with semi-automated image labeling, 2022. 5. Conclusion doi:10.20944/preprints202204.0177.v1. [7] U. Shafi, R. Mumtaz, Z. Shafaq, S. Zaidi, Z. Mah- This study introduces a novel approach to detect stripe rust mood, S. Zaidi, Wheat rust disease detection tech- in wheat crops, using AutoML and rigorous feature engineer- niques: a technical perspective, Journal of Plant ing techniques. By extracting a comprehensive set of statis- Diseases and Protection 129 (2022). doi:10.1007/ tical features from original images and employing Context- s41348-022-00575-x. Aware Automated Feature Engineering, we enhance the dis- [8] T. HayΔ±t, H. Erbay, F. VarΓ§Δ±n, F. HayΔ±t, N. Akci, The criminative power of the extracted features. Our iterative classification of wheat yellow rust disease based on a feature generation process aims to capture subtle patterns combination of textural and deep features, Multimedia and nuances, leading to superior effectiveness compared Tools and Applications 82 (2023) 1–19. doi:10.1007/ to state-of-the-art deep learning techniques. The consid- s11042-023-15199-y. ered wheat rust disease problem has already been tackled [9] R. Elshawi, S. Sakr, Automated machine learning: in the literature by employing several ML techniques, such Techniques and frameworks, in: R.-D. Kutsche, as feed forward NNs, KNN, SVM, RF. All of them populated E. ZimΓ‘nyi (Eds.), Big Data Management and Ana- the search space of TPOT, which helped to determine the lytics, Springer International Publishing, Cham, 2020, best one for the considered case. We compared our results pp. 40–69. against the ones by ResNet-18, a state-of-the-art technique [10] H. Eldeeb, M. Maher, O. Matsuk, A. Aldallal, used for the same kind of problem, according to the most R. El Shawi, S. Sakr, Automlbench: A comprehensive recent literature. The experiments were performed on a experimental evaluation of automated machine learn- publicly available dataset retrieved from the relevant liter- ing frameworks, 2022. doi:10.2139/ssrn.4516282. ature. Our approach outperformed the above-mentioned [11] N. Hollmann, S. MΓΌller, F. Hutter, Large language mod- state-of-the-art technique, revealing a higher computational els for automated data science: Introducing CAAFE effort of the latter in the allotted computing time. for context-aware automated feature engineering, in: Thirty-seventh Conference on Neural Information Pro- Acknowledgments cessing Systems, 2023. URL: https://openreview.net/ forum?id=9WSxQZ9mG7. This work has been partially funded by the Estonian Re- [12] R. Olson, J. Moore, Tpot: A tree-based pipeline opti- search Council, grant PRG1604, through the funding of mization tool for automating machine learning, 2019. SusAn, FACCE ERA-GAS, ICT-AGRI-FOOD and SusCrop doi:10.1007/978-3-030-05318-5_8. ERA-NET, and through the project Increasing the knowl- [13] C. Gu, T. Cheng, N. Cai, W. Li, G. Zhang, X.-G. Zhou, edge intensity of Ida-Viru entrepreneurship co-funded by the D. Zhang, Assessing narrow brown leaf spot sever- European Union. ity and fungicide efficacy in rice using low altitude uav imaging, Ecological Informatics 77 (2023) 102208. doi:10.1016/j.ecoinf.2023.102208. References [14] Y. Liu, L. An, N. Wang, W. Tang, M. Liu, G. Liu, H. Sun, M. Li, Y. Ma, Leaf area index estimation under wheat [1] FAO, Fao cereal supply and demand brief, 2023, powdery mildew stress by integrating uav-based spec- www.fao.org/worldfoodsituation/csdb/en/, https:// tral, textural and structural features, Computers www.fao.org/worldfoodsituation/csdb/en/, 2023. Ac- and Electronics in Agriculture 213 (2023) 108169. cessed: 2024-11-22. URL: https://www.sciencedirect.com/science/article/ [2] S. Savary, L. Willocquet, S. Pethybridge, P. Esker, pii/S0168169923005574. doi:https://doi.org/10. N. McRoberts, A. Nelson, The global burden of 1016/j.compag.2023.108169. pathogens and pests on major food crops, Na- [15] W. Zhu, Z. Feng, S. Dai, P. Zhang, X. Wei, Using ture Ecology & Evolution 3 (2019) 1. doi:10.1038/ uav multispectral remote sensing with appropriate s41559-018-0793-y. spatial resolution and machine learning to monitor [3] X. Chen, Pathogens which threaten food secu- wheat scab, Agriculture 12 (2022) 1785. doi:10.3390/ rity: Puccinia striiformis, the wheat stripe rust agriculture12111785. [16] Y. Xiao, Y. Dong, W. Huang, L. Liu, H. Ma, Wheat fusar- ium head blight detection using uav-based spectral and texture features in optimal window size, Remote Sens- ing 13 (2021). URL: https://www.mdpi.com/2072-4292/ 13/13/2437. doi:10.3390/rs13132437. [17] H. Zhang, L. Huang, W. Huang, Y. Dong, S. Weng, J. Zhao, H. Ma, L. Liu, Detection of wheat fusarium head blight using uav-based spectral and image feature fusion, Frontiers in Plant Science 13 (2022) 1004427. doi:10.3389/fpls.2022.1004427. [18] L. Liu, Y. Dong, W. Huang, X. Du, H. Ma, Monitoring wheat fusarium head blight using unmanned aerial vehicle hyperspectral imagery, Remote Sensing 12 (2020) 3811. doi:10.3390/rs12223811. [19] A. Guo, W. Huang, Y. Dong, H. Ye, H. Ma, B. Liu, W. Wu, Y. Ren, C. Ruan, Y. Geng, Wheat yellow rust detection using uav-based hyperspectral technol- ogy, Remote Sensing 13 (2021) 123. doi:10.3390/ rs13010123. [20] C. Nguyen, V. Sagan, J. Skobalski, J. Severo, Early detec- tion of wheat yellow rust disease and its impact on ter- minal yield with multi-spectral uav-imagery, Remote Sensing 15 (2023) 3301. doi:10.3390/rs15133301. [21] T. Zeng, J. Fang, C. Yin, Y. Li, W. Fu, H. Zhang, J. Wang, X. Zhang, Recognition of rubber tree powdery mildew based on uav remote sensing with different spatial resolutions, Drones 7 (2023) 533. doi:10.3390/ drones7080533. [22] S. Ding, J. Jing, S. Dou, M. Zhai, W. Zhang, Cit- rus canopy spad prediction under bordeaux so- lution coverage based on texture- and spectral- information fusion, Agriculture 13 (2023). URL: https: //www.mdpi.com/2077-0472/13/9/1701. doi:10.3390/ agriculture13091701. [23] S. Wold, K. Esbensen, P. Geladi, Principal com- ponent analysis, Chemometrics and Intelligent Laboratory Systems 2 (1987) 37–52. doi:10.1016/ 0169-7439(87)80084-9. [24] H. Eldeeb, R. El Shawi, Empowering machine learning with scalable feature engineering and interpretable automl, IEEE Transactions on Artificial Intelligence PP (2024) 1–16. doi:10.1109/TAI.2024.3400752. [25] W. Banzhaf, P. Nordin, R. Keller, F. Francone, Genetic programming: An introduction on the automatic evo- lution of computer programs and its applications, 1998. [26] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective genetic algorithm: Nsga-ii, Evolutionary Computation, IEEE Transactions on 6 (2002) 182 – 197. doi:10.1109/4235.996017. [27] OpenAI, Gpt-3 can’t count syllables - or doesn’t β€œget” haiku. https://community. openai.com/t/gpt-3-cant- count-syllables-or-doesnt-get-haiku/18733, 2021. ac- cessed on: 2024-03-1, 2021. [28] OpenAI, openai/openai-cookbook: Exam- ples and guides for using the openai api. https://github.com/openai/openai-cookbook, 2023b. (accessed on 03/1/2023), 2023. [29] N. Hollmann, S. MΓΌller, K. Eggensperger, F. Hut- ter, Tabpfn: A transformer that solves small tab- ular classification problems in a second, 2022. URL: https://arxiv.org/abs/2207.01848. doi:10.48550/ ARXIV.2207.01848. [30] Mlpclassifier documentation, https://scikit-learn.org/ stable/modules/generated/sklearn.neural_network. MLPClassifier.html, 2024. Accessed: 2024-11-22.