Aleksandar (Alex) Vakanski

 

Crop Health Assessment in Precision Agriculture

Precision agriculture (PA), or precision farming, is a concept for site-specific crop management based on observation and measurement of the crop variability in the field. The objective of PA is to reduce the use of herbicides, pesticides, fertilizers, and other chemical substances, and to increase crop yield. This objective can be achieved by using advanced technologies, where based on obtained knowledge growers can make informed decisions regarding the amount of herbicide or fertilizer to be applied to different areas in the field. Despite the confirmed economic and environmental benefits of PA, professional reports and scientific surveys indicate low rates of adoption. Among the many barriers towards increased adoption of PA, the need for new tools for agricultural data analytics that are ‘useful’ and ‘easy to use’ has been acknowledged as an essential factor.

The project proposes an approach for precision agriculture related to the assessment of crop health from aerial images collected with Unmanned Aerial Systems (UAS), popularly known as ‘drones’. The research focuses on the implementation of deep artificial neural networks (NN) for processing crop images. The ultimate goal is the design of NNs to distinguish healthy from diseased plants in aerial images, based on a set of learned features.

Although there is a large body of work on image-processing for crop health assessment based on the conventional machine vision approaches, only a few recent works have applied deep NN approaches for such tasks to natural field images. One challenge is the lack of large-scale labeled crop image datasets, as the performance of NN is reliant on the amount and quality of available data; furthermore, for satisfactory performance the field images are preferred to be crop-specific, and even region-specific. The proposed research addresses the problem of lack of agricultural data by providing an open dataset of multispectral aerial images of potato crop with labeled regions of healthy and stressed plants.

Advancing the data analytics aspects of PA also requires novel designs of deep NN architectures that are able to process natural aerial images of crop fields obtained by remote sensing. Rather than relying on models for classification of objects in images, a shift toward deep models that perform more challenging image processing tasks is needed, such as object detection. The goal in object detection is to: identify whether certain objects of interest (e.g., stressed plants) are present in an image, find the coordinates of the location of all present objects of interest in the image, and determine the category (i.e., disease or stress type) for all found objects. The project addresses two aspects of crop health evaluation: (a) spatial recognition of crop stress, and (b) temporal correlation of crop stress in images taken at different times during the growing season.

 

Publications

1. S. Butte, A. Vakanski, K. Duellman, H. Wang, and A. Mirkouei, "Potato crop stress identification in aerial images using deep learning-based object detection," Agronomy Journal, vol. 113, no. 5, pp. 3991–4002, Sep. 2021. [Bibtex] [Wiley Online Library