Multispectral Imaging, Image-processing and Classification for Agriculture

The prevention and control of weed infestation is a matter of importance in the agricultural domain. American farmers alone annually spend $4 billion to protect $16 billion in crops. This work moves in this direction and aims to classify weeds from crops and then spray the weeds in real-time. Another aim is to find the correlation coefficient between the nitrogen treatment levels of the crops and the statistical values computed. In this work, multispectral reflectance images are used in conjunction with classification techniques to detect, classify and spray weeds on a real-time basis in the field. Multispectral images were grabbed real-time in the field. This multispectral reflectance information was then used to develop an algorithm that would classify the plant matter from the background. The four images at the four different wavelengths ; 546nm, 690nm, 750nm and 803nm; were superimposed and the largest pixel was selected at every point of the resultant image. This image was then divided by the image grabbed at the wavelength (690nm) corresponding to the red portion of the spectrum. The resultant image was then thresholded, eroded and dilated to finally get an image consisting of only plant matter. A competitive neural network is then used to classify weeds from crops. To increase the speed of the process, we also use a priori information and adopt an algorithm that does not use neural networks but rather statistical methods. In the field, corn plants are planted in rows at a distance of 30 inches apart from each other. The algorithm does the weed-crop classification only along the rows. In between the rows, anything green is considered to be weed and is sprayed. Pipelining techniques like loop unrolling and loop parallelism have been incorporated in the algorithm. This enables the tractor, on which the classification system is mounted, to run at a maximum speed of 40 mph. Also, images in Field 81 of the Purdue Agricultural Research Park were imaged in conjunction with latitude and longitude values provided by a GPS system. Statistical parameters like mean, standard deviation, reflection index, normalized reflection index, normalized difference vegetation index and greenness ratio are calculated. These values are then correlated with the nitrogen treatment levels of the corn plants. The data of the corn plants are also taken by a SPAD meter by Litton Task Inc. The statistical values that they calculate are then correlated with the ones calculated by our system. The best correlation is found to be in the near infra-red region of wavelength 803nm.