Published Projects
Estimation of canopy height using unmanned ground vehicle with LiDAR for wheat phenotyping
Presented at the 45th Canadian Symposium on Remote Sensing, 10 - 13 June, 2024
The study explores the use of LiDAR data for canopy height estimation, comparing conventional and deep learning methods. The conventional approach, which utilizes percentile and percentage methods for ground and canopy height extraction, indicates that focusing on the upper parts of the canopy yields better predictions of canopy height. In contrast, the deep learning approach demonstrates significant effectiveness in predicting canopy height, with a CNN model showing a high R-squared value of 82.40%, underscoring the potential of advanced preprocessing and machine learning techniques in enhancing the accuracy of height estimation from LiDAR data.
Estimation of leaf nitrogen content with leaf spectrometer in potatoes
Presented at the 13th Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing, 31 October - 02 November, 2023
Nitrogen (N) fertilizer is one of the essential plant nutrients optimized to enhance crop yield and quality in modern agricultural fields. The conventional practice of applying N in a single dose before planting or in a random split dose without having precise knowledge of the amount and timing of crop N requirements has led to excessive use of N fertilizers in Canadian agriculture. Excess N application is a serious environmental issue having a huge impact on greenhouse gas emissions, soil health, and water resources (Xiong, Chen et al. 2015). Optimal N utilization is critical to maximize yield and mitigate the environmental impacts of run-offs from agricultural fields (Yang, Du et al. 2020, Jiang, Zhu et al. 2021). Leaf N content is an important indicator of photosynthetic efficiency (Jung, Song et al. 2021). It could assist agriculturists in better managing N applications at the early vegetation stages, that impact the yield in the later stages of the potato growth cycle. In this study, a commercial potato cultivar, Russet Burbank was grown to generate a prediction model on in-season N fertilizer application based on actual crop N requirements. We estimated the leaf nitrogen content with machine-learning models using spectral data obtained from SpectraVue leaf spectrometer. The treatments include an untreated control with nitrogen applied at five different rates (47.5, 95, 142.5, 190, and 237.5 kg/ha-1). The leaf spectral profiles (360 – 1100 nm) were obtained at the vegetative, tuber initiation, and maturity stages. Right after the spectral profiles were acquired, the leaves were sampled to obtain the leaf nitrogen content using Partial least squares regression models were developed to estimate the leaf N content from the spectral profiles. The developed models were able to estimate leaf N content with an R2 (coefficient of determination) of 78.6%. Increasing the number of observations could minimize the error and further improve the prediction accuracy.
Utilization of hyperspectral imaging to characterize herbicide phytotoxicity in oat and mustard
Presented at the 11th International Conference on Agro-Geoinformatics, 25 - 28 July, 2023
The use of herbicides is one of the predominant methods often deployed to restrict weeds. Over time, weeds mutate and develop resistance against certain herbicides, hence it is essential to screen for herbicide’s efficacy on evolving weed populations. This study aims to utilize proximal hyperspectral imaging to estimate crop injury from herbicide applications in tame oat [Avena sativa; model species for wild oat (Avena fatua)] and oriental mustard [Brassica juncea; model for wild mustard (Sinapis arvensis)]. The treatments included an untreated control along with 8 herbicides at their recommended dose for mustard and an untreated control along with 6 herbicides at their recommended dose for oat. The experiment was conducted at Lethbridge, AB, Canada. The imagery and the visual control rating were obtained at 2 different time points for hyperspectral imaging (HSI). The regression models developed with hyperspectral images were able to estimate crop phytotoxicity with an R2 (determination of coefficient) of 90.93% and 71.80%.
Proximal hyperspectral imaging to classify herbicide-resistant and -susceptible kochia (bassia scoparia)
Presented at the 11th International Conference on Agro-Geoinformatics, 25 - 28 July, 2023
Kochia (Bassia scoparia) is an invasive broadleaf weed species that has been reported to be responsible for up to 90% yield losses in major field crops. Herbicides are primarily being used to control Kochia. However, in recent years the incidence of herbicide resistant kochia has been increasing in North American prairies. There is a pressing need for a technique to recognize herbicide-resistant kochia biotypes for the susceptible ones to promptly suggest an alternative course of action. In this study, we assessed two different herbicides (Glyphosate and Fluroxypyr) with 6 different kochia populations each. Hyperspectral imagery of the treated kochia seedling was obtained on the day of treatment along with 1 and 3 days after treatment. The classification models that were built to identify resistant and susceptible kochia biotypes were able to classify with an accuracy of 75.11% for Glyphosate and 82.14% for Fluroxypyr.
U-Net based image segmentation for the estimation of blast severity in rice
Presented at ASABE Annual International Meeting, 9 - 12 July, 2023
Blast is a fungal disease caused by Magnaporthe oryzae that affects different parts of the rice plant in all development stages. In the late 20th century, this catastrophic widespread disease was responsible for a loss of between 10% to 30% in total production every year. Cultivars produced through conventional, cisgenic, or transgenic breeding are evaluated for resistance against leaf blast in a Uniform Blast Nursery (UNB). UBN simulates a warm and humid environment conducive for the fungus to thrive. Once the seedlings are established, about 21 days after sowing they are inoculated with a spore suspension. The symptoms are assessed based on the lesion size and the percentage of leaf area diseased between 5 to 7 days after inoculation. The disease symptoms are evaluated on a scale from 0 to 9 according to the Standard Evaluation Scale proposed by International Rice Research Institute (IRRI). This study aims to develop a U-Net-based semantic segmentation model that was built to lay out a few novel indices and pixel area parameters that quantify blast disease severity for assessment and validation. A total of 564 proximal images acquired before scoring were used for analysis. The model to segment lesions along with infected and uninfected plant pixels in the images was developed with a Dice coefficient of 0.96. Indices comparing the extracted pixel parameters were able to classify with an accuracy of 87.13%. The model performance predominantly declined due to the misclassification of images with scores 7 to 9 that are classified as susceptible.
UAV multispectral imaging to evaluate plant moisture and desiccant response in lentils (Lens Culinaris)
Presented at the 44th Canadian Symposium on Remote Sensing, 19 - 22 June, 2023
Timely harvest is important with lentils, especially since harvest timing is a compromise between increasing yield from younger pods and losses due to shattering of mature pods. Chemical desiccation has been proven to be very effective for lentils. Since chemical desiccation has an immediate dry-down effect, it is essential to ensure the crops are at maturity. Application of desiccation early would reduce the seed size and yield of the lentil crop. Multispectral imaging has been used as a promising alternative for herbicide tolerance studies. In this study, we used UAV multispectral imaging to estimate plant moisture to determine the optimal window for desiccation. In addition, following application of the desiccant, dry-down efficiency was assessed through visual ratings of the five conventional and two organic desiccants.
Multispectral imaging to estimate herbicide phytotoxicity in oat and mustard
Presented at the 44th Canadian Symposium on Remote Sensing, 19 - 22 June, 2023
Herbicides have become indispensable in modern agricultural production, especially in Western Canada. Extensive herbicide use has led to an increasing number of herbicide-resistant weeds. Hence, it is imperative to constantly screen herbicide efficacy on evolving weed populations. Plants exhibit a range of responses to herbicides with differing modes of action, including chlorosis, necrosis, stunting, epinasty, height reduction, biomass reduction, and mortality. Although these symptoms could be rated individually, they are cumulatively assessed on a 0 – 100 visual rating scale for crop phytotoxicity. Visual assessment for crop phytotoxicity and herbicide efficacy, although subjective, is the accepted standard used by pesticide manufacturers and the Pest Management Regulatory Agency in Canada. However, with the advent of hyperspectral imaging and machine learning, a novel, robust, non-subjective, and cost-effective evaluation system could be developed. In this study, we propose the use of proximal hyperspectral imaging to estimate crop injury/phytotoxicity in tame oat [Avena sativa; model for wild oat (Avena fatua)] and oriental mustard [Brassica juncea; model for wild mustard (Sinapis arvensis)].
Estimation of grain quality parameters in rice for high-throughput screening with near-infrared spectroscopy and deep learning
Published in Cereal Chemistry
Grain quality is a complex trait in rice, compared with other staple crops as it is predominantly consumed as a whole grain. Although considered secondary to yield, to align with consumer preferences, breeders are increasingly interested in quality. At the early stages of a breeding program, grain quality-related traits are often ignored as they are arduous and time-consuming. Near-infrared spectroscopy (NIRS) could be a suitable high-throughput alternative to conventional wet chemistry and image processing-related methods to be adopted for early screening. This study aims to quantify traits essential for rice breeders such as amylose, chalkiness, length, width, and the length/width ratio in rice samples with NIRS. We used conventional algorithms such as principal component analysis (PCA), partial least square regression (PLSR), multilayer perceptron (MLP), support vector classification (SVC), and linear discriminant analysis (LDA) to compare with the proposed convolutional neural network (CNN) for regression and classification.