Jacques-André Landry

Classification of hyperspectral data by decision trees and artificial neural networks to identify weed stress and nitrogen status of corn

This study evaluates the potential of decision tree classification algorithms for the classification of hyperspectral data, with the goal of discriminating between different growth scenarios in a cornfield. A comparison was also made between decision tree and artificial neural networks (ANNs) classification accuracies. In the summer of the year 2000, a two-factor field experiment representing different crop conditions was carried out. Corn was grown under four weed management strategies: no weed control, control of grasses, control of broadleaf weeds, and full weed control with nitrogen levels of 60, 120, and 250 N kg/ha. Hyperspectral data using a Compact Airborne Spectrographic Imager were acquired three times during the entire growing season. Decision tree technology was applied to classify different treatments based on the hyperspectral data. Various tree-growing mechanisms were used to improve the accuracy of classification. Misclassification rates of detecting all the combinations of different nitrogen and weed categories were 43, 32, and 40% for hyperspectral data sets obtained at the initial growth, the tasseling and the full maturity stages, respectively. However, satisfactory classification results were obtained when one factor (nitrogen or weed) was considered at a time. In this case, misclassification rates were only 22 and 18% for nitrogen and weeds, respectively, for the data obtained at the tasseling stage. Slightly better results were obtained by following the ANN approach. However, the advantage with the decision tree was the formulation of simple and clear classification rules. The highest accuracy was obtained for the data acquired at tasseling stage. The results indicate the potential of decision tree classification algorithms and ANN usage in the classification of hyperspectral data for crop condition assessment.

ESTIMATION OF CROP BIOPHYSICAL PARAMETERS THROUGH AIRBORNE AND FIELD HYPERSPECTRAL REMOTE SENSING

The potential of airborne hyperspectral remote sensing in crop monitoring and estimation of various biophysical parameters was examined in this study. A field experiment, consisting of four weed control strategies (no weed control, broadleaf control, grass control, and full weed control) as the main plot effect, factorially combined with three nitrogen (N) fertilization rates (60, 120, and 250 N kg ha–1), and replicated four times, was conducted. Hyperspectral data in 72 narrow wavebands (409 to 947 nm) from a Compact Airborne Spectrographic Imager (CASI) sensor were acquired 30 days after planting, at tasseling, and at the fully mature stage. In addition, measurements were made concurrently on various crop physiological parameters: leaf greenness (SPAD readings), leaf area index (LAI), plant height, leaf nitrogen content, leaf chlorophyll content, and associated factors such as soil moisture. Regression models were generated to estimate crop biophysical parameters and yield, in terms of reflectance at one or more wavebands, using the maximum r2 improvement criterion. The models that best represented the data had five wavebands as independent variables. Coefficients of determination (r2) were generally greater than 0.9, when based on the spectral data taken at the tasseling stage. Results were improved when normalized difference vegetation indices (NDVI) were used rather than the five–waveband reflectance values. The wavebands at 701 nm and 839 nm were the most prevalent in the NDVI–based models.

Potential of airborne hyperspectral remote sensing to detect nitrogen deficiency and weed infestation in corn

A compact airborne spectrographic imager (CASI) was used to obtain images over a field that had been set up to study the effects of various nitrogen application rates and weed control on corn (Zea mays). The objective was to determine to what extent the reflectances obtained in the 72 visible and near-infrared (NIR) wavebands (from 409 to 947 nm) might be related to differences associated with combinations of weed control (none, full, grasses only or broadleafs only) and nitrogen application rate (60, 120 or 250 kg/ha). Plots were arranged in split-plot experiment in completely randomized design at the McGill University Research Farm on Macdonald Campus, Ste Anne de Bellevue, Que., Canada. Weeding treatments were assigned to the main-plot units, and nitrogen rates to the sub-plot units. Three flights were made during the growing season. Data were analyzed for each flight and each band separately, then regrouped into series of neighboring bands yielding identical analyses with respect to the significance of the main effects and interactions on reflectance. The results indicate that the reflectance of corn is significantly influenced (α=0.05) at certain wavelengths by the presence of weeds, the nitrogen rates and their interaction. The influence of weeds was most easily observed in the data from the second flight (August 5, 2000), about 9 weeks after planting. The nitrogen effect was detectable in all the three flights. Differences in response due to nitrogen stress were most evident at 498 nm and in the band at 671 nm. In these bands, differences due to nitrogen levels were observed at all growth stages, and the presence of weeds had no interactive effect. Differences in other regions, whether related to nitrogen, weeds or the combination of the two, appeared to be dependent on the growth stage. Furthermore, results comparable to those of the hyperspectral sensor were obtained when a multispectral sensor was simulated, indicating little advantage of using the former.

Development of an Image Processing System and a Fuzzy Algorithm for Site-Specific Herbicide Applications

In precision farming, image analysis techniques can aid farmers in the site-specific application of herbicides, and thus lower the risk of soil and water pollution by reducing the amount of chemicals applied. Using weed maps built with image analysis techniques, farmers can learn about the weed distribution within the crop. In this study, a digital camera was used to take a series of grid-based images covering the soil between rows of corn in a field in southwestern Quebec in May of 1999. Weed coverage was determined from each image using a “greenness method” in which the red, green, and blue intensities of each pixel were compared. Weed coverage and weed patchiness were estimated based on the percent of greenness area in the images. This information was used to create a weed map. Using weed coverage and weed patchiness as inputs, a fuzzy logic model was developed for use in determining site-specific herbicide application rates. A herbicide application map was then created for further evaluation of herbicide application strategy. Simulations indicated that significant amounts of herbicide could be saved using this approach.

Development of a herbicide application map using artificial neural networks and fuzzy logic

Abstract The primary objective in this project was to develop a precision herbicide-spraying system in a corn field. Ultimately, such a system would involve real-time image collection and processing, weed identification, mapping of weed density, and sprayer control using a digital camera. A proposed image processing method involving artificial neural networks was evaluated for image recognition accuracy, computer time and memory requirements. The greenness method, based on a pixel-by-pixel comparison of red-green-blue intensity values, was successfully developed. The recognition of weeds in the field was then simplified by taking images between the corn rows. The images were processed by the greenness method to obtain percent greenness in each image. This information was used to create weed coverage and weed patchiness maps. Based on these maps, herbicide application rates were determined for each spot in the field. This was done by using the weed coverage and weed patchiness maps as inputs to a simulated fuzzy logic controller, and integrating the output of the controller over the field area corresponding to the input images. Simulations using different fuzzy rules and membership functions indicated that the precision spraying has potential for reducing water pollution from herbicides needed for weed control in a corn field.

Hyperspectral Image Classification to Detect Weed Infestations and Nitrogen Status in Corn

The potential of hyperspectral aerial imagery for the detection of weed infestation and nitrogen fertilization level in a corn (Zea mays L.) crop was evaluated. A Compact Airborne Spectrographic Imager (CASI) was used to acquire hyperspectral data over a field experiment laid out at the Lods Agronomy Research Centre of Macdonald Campus, McGill University, Quebec, Canada. Corn was grown under four weed management strategies (no weed control, control of grasses, control of broadleaf weeds, and full weed control) factorally combined with nitrogen fertilization rates of 60, 120, and 250 N kg/ha. The aerial image was acquired at the tasseling stage, which was 66 days after planting. For the classification of remote sensing imagery, various widely used supervised classification algorithms (maximum likelihood, minimum distance, Mahalanobis distance, parallelepiped, and binary coding) and more sophisticated classification approaches (spectral angle mapper and linear spectral unmixing) were investigated. It was difficult to distinguish the combined effect of both weed and nitrogen treatments simultaneously. However, higher classification accuracies were obtained when only one factor, either weed or nitrogen treatment, was considered. With different classifiers, depending on the factors considered for the classification, accuracies ranged from 65.84% to 99.46%. No single classifier was found useful for all the conditions.

Relaxed genetic programming

A study on the performance of solutions generated by Genetic Programming (GP) when the training set is relaxed (in order to allow for a wider definition of the desired solution) is presented. This performance is assessed through 2 important features of a solution: its generalization error and its bloat, a common problem of GP individuals. We show how a small degree of relaxation improves the generalization error of the best solutions; we also show how the variation of this parameter affects the bloat of the solutions generated.

A Genetic-Programming-Based Method for Hyperspectral Data Information Extraction: Agricultural Applications

A new method, called genetic programming-spectral vegetation index (GP-SVI), for the extraction of information from hyperspectral data is presented. This method is introduced in the context of precision farming. GP-SVI derives a regression model describing a specific crop biophysical variable from hyperspectral images (verified with in situ observations). GP-SVI performed better than other methods [multiple regression, tree-based modeling, and genetic algorithm-partial least squares (GA-PLS)] on the task of correlating canopy nitrogen content in a cornfield with pixel reflectance. It is also shown that the band selection performed by GP-SVI is comparable with the selection performed by GA-PLS, a method that is specifically designed to deal with hyperspectral data.

DEVELOPMENT OF NEURAL NETWORKS FOR WEED RECOGNITION IN CORN FIELDS

The main objective of this project was to develop a weed recognition system based on artificial neural networks to assist in the precision application of herbicides in corn fields. Digital images were collected in May 1998 using a commercially available digital camera. The intensities of the three primary colors (red, green, and blue) were compared for each pixel of the images. The three intensities of a pixel remained unchanged when, in the pixel, the green intensity was greater than each of the other two; otherwise, the three intensities of the pixel were set to zero. Background objects, except plants, were thus removed from the images. The resulting pixel intensities of the modified images were used as the inputs for Learning Vector Quantization (LVQ) artificial neural networks (ANNs). ANNs were trained to distinguish corn from weeds, as well as to differentiate between weed species. The success rate for a single ANN in distinguishing a given weed species from corn was as high as 90%, and as high as 80% in distinguishing any of four weed species from corn. Better success rates might be obtainable with more elaborate schemes for data input and/or structural improvements such as cascading. The image–processing time for the ANNs was as short as 0.48 s per image, thus making it useful for real–time data processing and application of herbicides. The development of such ANNs for weed recognition could be useful in precision farming to guide site–specific herbicide application and ultimately reduce the total amount of herbicide applied as well as lowering the risk of pollution.

Automated annotation of corals in natural scene images using multiple texture representations

Current coral reef health monitoring programs rely on biodiversity data obtained through the acquisition and annotation of underwater photographs. Manual annotation of these photographs is a necessary step, but has become problematic due to the high volume of images and the high cost of human resources. While automated and reliable multi-spectral annotation methods exist, coral reef images are often limited to visible light, which makes automation difficult. Much of the previous work has focused on popular texture recognition methods, but the results remain unsatisfactory when compared to human performance for the same task. In this work, we present an improved automatic method for coral image annotation that yields consistent accuracy improvements over existing methods. Our method builds on previous work by combining multiple feature representations. We demonstrate that the aggregation of multiple methods outperforms any single method. Furthermore, our proposed system requires virtually no parameter tuning, and supports rejection for improved results. Firstly, the complex texture diversity of corals is handled by combining multiple feature representations: local binary patterns, hue and opponent angle histograms, textons, and deep convolutional activation feature. Secondly, these multiple representations are aggregated using a score-level fusion of multiple support vector machines. Thirdly, rejection can optionally be applied to enhance classification results, and allows efficient semi-supervised image annotation in collaboration with human experts.