1
Development of a Two-Layer Meta-Classifier–Based Drought Stress Detection System for Wheat Crop Sustainability

Ankita Gupta1*, Lakhwinder Kaur2 and Gurmeet Kaur3

1Department of Computer Science and Engineering, CT Group of Institutions, Shahpur Campus, Jalandhar, Punjab, India

2Department of Computer Science and Engineering, Punjabi University, Patiala, Punjab, India

3Department of Electronics and Communication Engineering, Punjabi University, Patiala, Punjab, India

Abstract

Focusing on the Raj 3765 wheat variety’s chlorophyll fluorescence–based canopy level images, this research aims to formulate an ultraprecise, bilayered meta-classifier to curtail economic adversity within wheat farming. Preprocessing of images, utilizing the total variation - L1 (TV-L1) with a primal-dual algorithm, succeeded by contrast stretching (Min-Max), emerged as paramount. This pipeline’s choice hinged on mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) average values. Regarding feature selection and optimization, 23 gray-level co-occurrence matrix (GLCM) features, 9*255 color band features, 10 landmarks, and 20 wheat canopy geometric features proved significant. These features substantially enhanced machine learning algorithms’ efficacy. Most prominent classifiers are included in this research, viz, random forests (RF), naive Bayes (NB), K-nearest neighbors (KNN), and support vector machines (SVM), with RF achieving an accuracy of 93.06%, followed by NB and KNN. A blend of these classifiers showed RF and NB flourishing in the initial layer and KNN in the succeeding layer, exhibiting numerical constancy relative to the difference between training and testing accuracy. Our custom meta-classifier accomplished an apex accuracy of 95.8%. The research concludes with a case study showcasing the potential of such advancements in mitigating farmers’ losses.

Keywords: Wheat, water-stress, machine learning, phenotyping, chlorophyll fluorescence, image processing, meta-learning, sustainability

1.1 Introduction

Wheat and barley are two main crops on which consistent research has been done since the independence of India [1]. Wheat is a nutrient-rich, inexpensive source of energy that is highly consumed by all sections of society throughout the world [2]. Due to an accelerated increase in the population, the country (India) will not be able to feed its population due to lower yields in these crops [3]. The Indian government has been making sincere efforts to overcome this challenge by introducing new high-yielding and stress-tolerant wheat varieties [4]. This can be accomplished by increasing the area under irrigation and technological regimes and developing sustainable agriculture practices [5].

Techniques proved to be most fruitful in differentiating stress-tolerant and stress-susceptible genotypes utilizing physiological tools like canopy morphological analysis [6]. Most of the authors are using the crop water stress index (CWI) as a key metric for the identification of water stress by measuring canopy temperature [5, 7]. But still, there is a need to develop a technique that is purely nonintrusive in a cost-effective manner. It has been observed that an aggregate of 60% of global agriculture uses rain-fed farming practices, which are under threat due to unprecedented rains [8]. This underlines the need for the development of an affordable phenotyping platform for wheat crops under drought and controlled conditions for stress identification. The authors can identify the most dominant factors that are influenced by drought stress nonintrusively by analyzing color spectrum, texture, and statistical features [9–11]. The yield losses can be significantly affected by the water stress caused by the deficiency in soil moisture levels. This leads to an urge to analyze the soil and its characteristics for the development of efficient drought control systems.

There are standard protocols (gravimetric method) for conducting soil sampling, and the analysis of the water content in the soils is done in the laboratory using instrumentation [12, 13]. The method requires a lot of fieldwork and does not exactly give the status of plant conditions in a noninvasive manner. Another approach is the analysis of the subsurface of the soil using multiple instruments and sensors. This method requires field work and the installation of moisture probes along with other peripheral equipment that can be connected to computer networks [14, 15]. The neutron probes use radioactive material to analyze the water content in the soil. Multiple instruments with similar capacity and functionality use principles of resistance, dielectric permittivity, thermal conductivity, and thermalized neutrons [16–18]. Through the soil-water retention or soil-water characteristic curve [19], tensiometers relate the relationship between pressure potential and soil water tension to soil water content [20]. This relationship is unique to each soil or porous medium.

Furthermore, indicators known as “organ level stress indicators” can also be used to identify water stress in wheat [21]. In this study, various parts of the wheat plant, including the canopy, roots, pollens, grains, development stages, hormonal regulation, and leaves, have been analyzed to gain insights into the signals emitted by these components in response to water stress. The objective was to understand how each part of the plant indicates the presence of water stress. This objective can be accomplished by employing high-throughput imaging technologies, such as hyperspectral imaging, thermal imaging, and chlorophyll fluorescence (CHF) imaging. These advanced imaging techniques enable the computation of spectral vegetation indices and xanthophyll indices using satellite imagery and cameras, which capture and quantify the “photosynthetic activity” within a given frame. The utilization of these imaging technologies proves to be highly valuable in assessing and monitoring the impact of water stress on different parts of the wheat plant [22].

1.2 Literature Review

Water stress in plants is a complex phenomenon that has significant implications for their growth and development. As researchers have explored this subject, they have observed a range of changes in plants exposed to dehydration stress, including reductions in the number of leaves, leaf size, leaf area, and leaf longevity [15, 23–26]. These alterations ultimately lead to a decrease in the overall canopy size and shape of the plants. Furthermore, water stress also affects the root-to-shoot ratio and the biomass of the plants, as they struggle to absorb water in different parts of their body under drought conditions [27].

To gain a deeper understanding of water stress in wheat, researchers have conducted experiments using field setups to impose water stress at various stages of crop growth and development [28, 29]. They have focused on parameters such as canopy biomass and root-shoot size to study drought tolerance in wheat crops [30, 31]. It has been observed that during extended dry periods, crops extend their roots deeper into the soil, altering their root-shoot ratio and canopy morphology [31–33].

In addition to the noninvasive techniques mentioned earlier, other technical methods have proven valuable in detecting water stress in plants. Thermal imaging has been widely used to assess changes in plant temperature distribution, providing insights into their water status and stress levels [6, 34]. Hyperspectral imaging, on the other hand, analyzes the interaction of plants with different wavelengths of light, allowing researchers to identify specific spectral signatures associated with water stress [35–38]. These techniques, along with advancements in digital imaging instrumentation and high-throughput phenotyping capabilities of cameras, have transformed precision agriculture. Moreover, the use of 3D imaging and LiDAR has gained prominence, enabling detailed analysis of plant architecture and structure, as well as precise measurement of plant height, canopy structure, and biomass [39].

Traditional plant phenotyping methods, involving manual processing of agricultural samples, are time-consuming and inefficient. However, the advent of data collection tools like cameras and unmanned aerial vehicles (UAVs) has revolutionized this field [10, 15, 40]. CHF imaging using remote sensing is a valuable tool for detecting and assessing drought and water stress conditions in vegetation. It measures the light emitted by chlorophyll during photosynthesis and provides indicators of stress and changes in photosynthetic efficiency. By quantifying fluorescence signals, remote sensing enables the early detection of water stress, estimation of water use efficiency, and large-scale monitoring of vegetation health. This information aids in timely intervention and resource allocation, supporting effective drought management and sustainable agricultural practices.

Machine learning approaches have significantly improved the accuracy of CHF-based systems by leveraging time series data for trend analysis. By automating the process of feature extraction, machine learning pipelines facilitate the development...