Automatic Spinach counting from UAV imagery using machine vision.
Summary
In this study common machine vision image segmentation algorithms, such as the Excess Green Index and Otsu’s method along with deep learning and convolutional neural networks will be used in order to create a fully automatic method of counting the number of plants. This is achieved by segmenting and binarizing an ortho-mosaic of a spinach field. The result is a binary image, where all the true pixels represent pixels that belong to a spinach plant. By training a neural network to recognize individual spinach plants and classify them as such, the number of pixels per individual spinach plant can be automatically calculated.
Afterward by diving the total amount of pixels by the average amount of pixels per plant the number of spinach plants can be calculated. Tests on smaller images where the plants could be counted by hand showed that the algorithm is capable of automatically counting the number of plants with an accuracy of 90%. The study also tested this approach on an ortho-mosaic of a smaller resolution and it still performs as expected. With the biggest error being 9.6% meaning that the algorithm is capable of counting plants from ortho-mosaics with different spatial resolutions.