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 eficient 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, ofering a cost-efective alternative to traditional deep learning methods that require extensive computational resources and expertise.
The Food and Agriculture Organization (FAO) forecasts a
0.9% increase in global cereal utilization for 2023/24
compared to the previous year. Wheat, as the most widely
cultivated crop globally, is essential to agriculture, with rising
consumption expected in regions like the European Union,
China, India, the UK, and the US [
Traditional methods for monitoring wheat rust rely on
manual visual inspection, which is time-consuming,
laborintensive, and costly, making it impractical for large-scale
agriculture [
Efective and timely monitoring of yellow rust is essential
for both disease management and sustainable crop
production. Accurate disease mapping facilitates the judicious
application of fungicides and enhances breeding programs
by identifying resistant wheat varieties [
This study introduces a novel approach that integrates
AutoML with context-aware feature engineering for the
detection of stripe rust in wheat. We extract comprehensive
statistical features from UAV-captured images and refine
them using Context-Aware Automated Feature Engineering
(CAAFE), a feature engineering method designed for
tabular datasets [
The application of Unmanned Aerial Vehicles (UAVs) for
plant disease detection has garnered substantial interest,
leading to the development of advanced methodologies that
integrate image processing with Machine Learning (ML)
algorithms. Gu et al. [
In the field of wheat disease detection, Liu et al. [
Broadening the application beyond wheat, research on UAV-based disease detection in rubber trees [21] and citrus plants [22] demonstrated the broad applicability of UAVbased disease detection techniques across diferent agricultural sectors. These studies emphasized the vital role of advanced image processing techniques and ML algorithms in enhancing global food security by enabling efective disease detection in a wide range of crops.
Figure 1 illustrates the architecture of our approach that consists of three main stages, including data preprocessing (Section 3.2), automated feature engineering using CAFE (Section 3.3), and model training and evaluation using AutoML approach (Section 3.4). In the following subsections, we explain the diferent building blocks of our approach.
In this study, we used a publicly available dataset, RustNet
[
Our preprocessing phase involves several key stages: initial image acquisition, conversion to grayscale, resizing, and feature extraction. During feature extraction, we compute essential statistical measures, including mean, standard deviation, variance, correlation, energy, entropy, contrast, skewness, kurtosis, and homogeneity.
Feature engineering is a critical component of machine learning, as it involves transforming raw input data into features that can improve predictive performance [23, 24]. In our approach, we leverage CAAFE, an automated machine learning technique specifically designed for tabular datasets. CAAFE employs an LLM to iteratively generate semantically meaningful features based on a detailed description of the dataset. This process not only generates Python code for creating new features but also provides explanations for the relevance and utility of the generated features.
CAAFE operates iteratively on both the training and validation datasets, and , along with a description of the dataset’s context and features. In each iteration, CAAFE constructs a prompt that includes detailed information about the dataset and the specific feature engineering task, which is then passed to the LLM. Based on this prompt, the LLM generates code to alter or create new features. The generated code is executed on the current datasets ( and ), producing transformed datasets (′ and ′). An ML classifier is subsequently trained on ′ and evaluated on ′. If the classifier’s performance on ′ surpasses its performance on the original , the newly generated feature is retained, and the datasets are updated accordingly. If not, the feature is discarded, and the datasets remain unchanged.
The prompt provided to the LLM includes semantic and descriptive information about the dataset, such as a usergenerated dataset description, feature names, data types, the percentage of missing values, and random sample rows from the dataset. Additionally, a template for the expected format of the generated code and explanations is included, which improves the clarity and quality of the LLM’s output. To further enhance performance, chain-of-thought instructions guide the LLM through a series of intermediate reasoning steps, leading to more efective code generation. By utilizing CAAFE, we integrate domain knowledge into the feature engineering process, all while maintaining interpretability and optimizing predictive performance. This approach ofers a powerful and eficient method for generating high-quality features in complex datasets, marking a promising advancement in machine learning research.
TPOT is an AutoML framework designed for constructing and optimizing machine learning pipelines for both classiifcation and regression tasks. It utilizes tree-based genetic programming [25] to evolve pipelines by treating them as individuals within an evolutionary algorithm. Each pipeline is structured as a tree, with its nodes categorized as either Primitives or Terminals. Primitives represent operators that require input, such as machine learning algorithms needing data and hyperparameter values. Terminals, on the other hand, are constants that provide input to the Primitives. Notably, a Primitive can also act as input for another Primitive, allowing for complex pipeline configurations. The evolutionary process in TPOT operates by applying genetic operations such as mutation and crossover to the pipelines. Mutation involves making small modifications, such as changing a hyperparameter or introducing a new preprocessing step. Crossover, on the other hand, selects two pipelines that share common Primitives and allows them to exchange subtrees or branches. Once these operations are performed, each pipeline is evaluated and assigned a fitness score, which reflects its performance. This fitness score is used in the selection process to determine which pipelines should be retained and evolved further in the next generation, ultimately leading to the creation of highly optimized machine learning pipelines. Generally, these pipeline trees could be arbitrarily large. Nevertheless, extensive machine learning pipelines usually have downsides. Longer pipelines with numerous hyperparameters can be challenging to fine-tune, more prone to overfitting, complicate the understanding of the final model, and demand extended evaluation time, thus slowing down the optimization process. Due to these considerations, a multiobjective optimization technique, NSGAII [26], is employed. It assists in selecting candidates based on the Pareto front, representing the balanced trade-of between pipeline length and performance.
Training and test. For a fair comparison, we adopted the
same train-test split methodology as outlined in the
referenced study, allocating 70% of the RustNet dataset for
training and 30% for testing [
Class disease no_disease Total 380 128 508
Baselines. Given the randomized nature of the
experiments reported in [
CAAFE setting. We leverage the advanced capabilities of OpenAI’s language models, including GPT-3.5, as LLM within the CAAFE framework [27, 28]. The integration of these powerful language models enables CAAFE to generate semantically meaningful features iteratively, enhancing the efectiveness of feature engineering. To ensure robust performance and accuracy, we conduct ten feature engineering iterations using the CAAFE framework. Additionally, in the iterative evaluation of code blocks, we employ TabPFN (Tabular Predictive Functional Network), as proposed by Hollmann et al. [29], to assess the efectiveness of generated features and their impact on model performance.
TPOT setting. To ensure a fair comparison, an equal time budget was allocated for both TPOT and ResNet methodology. Experiments were constrained to a 20-minute time limit. This consistent time allocation ensures parity in computational resources between the methods, enabling a thorough and unbiased evaluation of their respective performances. The input to TPOT is a data matrix after performing the feature engineering step from CAAFE. The hyperparameters for TPOT were configured with a set number of generations, specifically 10, and a population size of 100. The resulting pipeline generated by TPOT, constrained by the specified time budget, is a multi-layer perceptron classifier with a learning rate of 0.01 and regularization parameter of 0.0001. The latter is a penalty term, constraining the size of the weights [30]. The aim of such a strategy is to reduce overfitting and enhance the generalization ability of the NN.
Hardware Resources. We conducted our experiments on a CPU environment. The CPU environment runs on Windows 11 Pro 64-bit (10.0, Build 22621) with 16 core Intel(R) Core(TM) i9-10885H Processor @ 2.40GHz,32 GB DIMM memory, and 1000 GB SSD data storage. All the approaches have been implemented in Python.
Performance metrics. Since the classification problem is being tackled in this study, the performance metrics used are Accuracy, Precision, Recall, and F1-score.
4.2.1. Preprocessing We followed the preprocessing steps described in Section 3.2. Regarding the conversion of the class associated with the image to a numerical equivalent, we adopted for RustNet dataset = 1 and _ = 0.
After the statistical features are extracted from images, the resulting feature set is normalized using min-max normalization, where each feature has a value between 0 and 1. The general formula for min-max normalization is:
A demonstration of CAAFE using the RustNet dataset is illustrated in Figure 2. User inputs are highlighted in blue, MLclassifier-generated data in red, and LLM-generated code is presented with syntax highlighting. The code includes comments 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. The retained generated features from CAAFE after 10 iterations include ’mean_variance_ratio’, calculated as the mean divided by the variance, and ’contrast_energy_ratio’, computed as the contrast divided by the energy. Incorporating the features generated by CAAFE into TPOT improved the accuracy from 93.02% achieved using TPOT alone on the validation dataset to 95.42%, as shown in Table 2.
The results, evaluated using TPOT and ResNet-18, are detailed in Table 2. For TPOT, two variants are considered: the baseline TPOT and TPOT with Context-Aware Automated Feature Engineering (CAAFE), referred to as TPOT (FE). The comparative results demonstrate that both variants of our proposed framework—TPOT and TPOT (FE)—outperform 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 lower accuracy of 85.2% on the same dataset, highlighting the superior performance of our proposed approach. The limited number of epochs achieved within the allocated time budget highlights the substantial computational efort required.
Dataset Model
Accuracy
Precision
TPOT RustNet TPOT (FE)
ResNet-18
This study introduces a novel approach to detect stripe rust in wheat crops, using AutoML and rigorous feature engineering techniques. By extracting a comprehensive set of statistical features from original images and employing ContextAware Automated Feature Engineering, we enhance the discriminative power of the extracted features. Our iterative feature generation process aims to capture subtle patterns and nuances, leading to superior efectiveness compared to state-of-the-art deep learning techniques. The considered wheat rust disease problem has already been tackled in the literature by employing several ML techniques, such as feed forward NNs, KNN, SVM, RF. All of them populated the search space of TPOT, which helped to determine the best one for the considered case. We compared our results against the ones by ResNet-18, a state-of-the-art technique used for the same kind of problem, according to the most recent literature. The experiments were performed on a publicly available dataset retrieved from the relevant literature. Our approach outperformed the above-mentioned state-of-the-art technique, revealing a higher computational efort of the latter in the allotted computing time.
This work has been partially funded by the Estonian Research Council, grant PRG1604, through the funding of SusAn, FACCE ERA-GAS, ICT-AGRI-FOOD and SusCrop ERA-NET, and through the project Increasing the knowledge intensity of Ida-Viru entrepreneurship co-funded by the European Union. ium head blight detection using uav-based spectral and texture features in optimal window size, Remote Sensing 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 technology, Remote Sensing 13 (2021) 123. doi:10.3390/ rs13010123. [20] C. Nguyen, V. Sagan, J. Skobalski, J. Severo, Early detection of wheat yellow rust disease and its impact on terminal 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 diferent spatial resolutions, Drones 7 (2023) 533. doi:10.3390/ drones7080533. [22] S. Ding, J. Jing, S. Dou, M. Zhai, W. Zhang, Citrus canopy spad prediction under bordeaux solution coverage based on texture- and spectralinformation 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 component 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 evolution 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 “get” haiku. https://community. openai.com/t/gpt-3-cantcount-syllables-or-doesnt-get-haiku/18733, 2021. accessed on: 2024-03-1, 2021. [28] OpenAI, openai/openai-cookbook: Examples and guides for using the openai api. https://github.com/openai/openai-cookbook, 2023b. (accessed on 03/1/2023), 2023. [29] N. Hollmann, S. Müller, K. Eggensperger, F. Hutter, Tabpfn: A transformer that solves small tabular 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.