Maryam Bafandkar, MSc

As a product owner, I was involved with research marketing, working on most sections of the Canvas Business Model, finding user requirements and designing the product strategies, features, and sustainable roadmaps. Developed AI features for SEO purposes.

As a member of the product and development team, After 6 months of meeting with clinicians and collecting data, we developed a practical application with 96% accuracy which is used in the Montreal Hospital.

As a Co-founder of the company, I was involved in market research, defining the product roadmap, and assigning tasks to the development team. I led the team and managed project finances.

Automatically distinguishing different types of plant images is a challenging problem relevant to both Botany and Computer Science disciplines. Plant identification at the species level is a computer vision task called fine-grained categorization, which focuses on differentiating between hard-to-distinguish object classes. This classification problem is complicated and challenging because of the lack of annotated data, inter-species similarity, the large-scale features in appearance, and a large number of plant species. A plant classification system capable of addressing the complexity of this computer vision problem has important implications for society at large, not only in public computer science education but also in numerous agricultural activities such as the automatic detection of cash crops and non-crop plants (called weeds).

Worked with NRC, JRH hospital and the health community to combat COVID-19 in Canada by using my background knowledge in web designing and developing and utilizing Data science and explainable approaches to chest X-ray images.
A practical COVID-19 prediction and interpretation have been developed to assist the clinician in predicting the severity of the COVID-19 virus for each patient.

The goal of this project was to visual and crops and weeds images to choose the right images and delete the wrong ones. The impact of project increased the accuracy on the test data by 26%.
After data visualization and clearing process, we split the data into training, validation and test dataset.

Image feature extraction CNN Architectures on VGG, Resnet, InceptionNet, and XceptionNet. A Gold mine dataset for computer vision is the ImageNet dataset. It consists of about 14 M hand-labelled annotated images which contain over 22,000 day-to-day categories. Every year ImageNet competition is hosted in which the smaller version of this dataset (with 1000 categories) is used with an aim to accurately classify the images. Many winning solutions of the ImageNet Challenge have used state of the art convolutional neural network architectures to beat the best possible accuracy thresholds. In this kernel, I have discussed popular architectures such as VGG16, 19, ResNet, AlexNet etc. In the end, I have explained how to generate image features using pre-trained models and use them in machine learning models.

Worked on a million number of labelled plant images. The following steps have been applied:

• Re-labeling images were done to decrease the number of classes, hence, simpler
  classification problem;
• Filtering plants by age criteria to remove possible noise from the dataset;
• Grouping was done to generate a hierarchical structure of labels.

In a hierarchical CNN classifier, the upper nodes classify the input images into subgroups
[109]. Then, deeper levels classify deeper discrimination, for example, for fruit or human face
classification problem. The upper nodes classify fruit or human face images into subgrouping
like yellow-coloured objects together or human faces together [109]. Then, deeper nodes
classify bigger differences, such as "lemon" v/s "orange" fruits or "old man," v/s "kid boy,"
human faces. Study [109] proves that hierarchical CNN models perform at par or even
better than standard CNNs. In tree-based or hierarchical CNN structures, initial layers of
a CNN from the top of structure learn very general features [87] that have been exploited
for transfer learning [110, 86]. [108] used the hierarchical CNN’s structure to deal with
distinguishing the similar classes. They designed multi-layer hierarchical CNNs where an
abstract higher-level network initially determines which subnetwork a sample should be
directed to. Lower level networks are designed to find discriminating features amongst
similar classes in the subnetwork. Each sub-network is called a class assignment classifier.

To formulate a definition of the concept of digital agriculture, we consider the stages of
agricultural development. Agriculture 1.0 is the first level of agriculture that was based on the
use of manual labour (early 20th century). Agriculture 2.0 is called the "Green Revolution"
(the late 1950s), when fertilizers, pesticides began to be actively used [55]. Fertilizer is a
natural or artificial substance containing the chemical elements that improve the growth and
productiveness of plants [66]. Moreover, a pesticide is any substance used to kill, repel,
or control certain forms of plant or animal life that are considered to be pests. Pesticides include herbicides for destroying weeds and other unwanted vegetation, insecticides for
controlling a wide variety of insects, fungicides used to prevent the growth of moulds and
mildew, disinfectants for preventing the spread of bacteria, and compounds used to control
mice and rats [65].
Moreover, pesticide contamination moves away from the target plants, resulting in
environmental pollution. Such chemical residues impact human health through environmental
and food contamination [104]. It is generally accepted that pesticides play an important role
in agricultural development because they can reduce the losses of farm products and improve
the affordable yield and quality of food [3, 96]. Fertilizers and pesticides are a necessary evil
for industrial agriculture [76]. They have adverse effects on soil, plants, environment, and
human health [11]. Nevertheless, high efficiency in the agricultural sector is not currently
possible without herbicides and pesticides [14]. Also, there is a high demand to produce
more agricultural foods and products to meet the growing population. It is vital to have
precise agriculture with less waste and sustainable outcomes [45].
After agriculture 2.0, using fertilizers and pesticides, came agriculture 3.0 called precision
agriculture (the 1990s and 2000s), and agriculture 4.0 called digital agriculture (early 2010s)
[55]. They utilize digital technologies and technical means in agricultural production making
it possible to bring exact measurements to a new level when information on all agricultural
processes and operations exist in digital form, and, at the same time, the transfer, processing
and analysis of data are automated.
The development of technological advances has been growing year by year [88]. Using
digital agriculture to classify plants is one of the trends [50]. Crop and weeds can largely
vary within the same field. Identifying weeds and removing them from the field contributes
significantly to the final yield. Digital agriculture allows scanning large areas of a plant and
distinguishes between weeds and crops. In recent years, research into digital agriculture
using machine learning methods has become an active area of study. [47] believed weeding
is an effective way to increase crop yields. Their focus is on improving weed and crop
recognition accuracy on weed dataset [59]. They found CNN [33] are favourable for multiclass
crops and weeds recognition with limited labelled data in agricultural recognition tasks.
Read more on my Article on the Link below.

Using a public dataset of 4,234 plant images from the Aarhus University Signal Processing group in collaboration with the University of Southern Denmark, that consists of descriptions under a controlled condition concerning camera radiance and stabilization. The project aims to solve the problem of plant seedling classification problems utilizing the transfer leaning technique by exploring three famous Convolutional Neural Network models, namely AlexNet, VGG16, and Xception CNN architectures to pick the right model and provide solutions to increase its performance and accuracy. Thisproject is compared with the ISI paper worked on the same dataset but with different approaches, published in 2019. The trained model achieved 97% accuracy during testing. The tuned CNN candidate model can significantly deal with overfitting problems to improve the performance of the model.

• Run Faculty.Bio Business as a Co-Founder and CEO.
• Lead Webdeveloper, Marketing, Sale people
• Manage and organize Faculty.Bio project which is a big online platform.
• Work on collaboration and partnership of project
• Work on business and marketing strategies

• Assessed business needs to align business initiatives with information technology and AI solutions.
• Provide support to the business units and training/documentation on these technology solutions.
• Leading different big-size projects at the same period of time
• Assessed the business plan of projects based on the feasibility, cost and their values for Canadian customers.