Weed detection in rice fields using aerial images and neural networks

In this paper, we investigate the use of neural networks (NN) to detect weed plants in rice fields based on aerial images. For this purpose, images are taken at 50 meters high with 16.1 megapixels CMOS digital camera mount-ted on an autonomous electrical fixed wind plane. Then, an ortho-mosaic map of the field is created by stitching 250 pictures, as the image is ortho-corrected, the pixel information on the final map is more reliable for the analysis. For the NN training, Gray-Level Co-Occurrence Matrix (GCLM) with Haralicks descriptor are used for texture classification as well as Normalized Difference Index (NDI) for color. As result we have 99% precision for detection of weed on the test data, this indicates that neural networks can have a good performance on the weed detection on rice fields. For weed plants similar in form to rice plants, the level of detection was low, due to images resolution when this are taken at 50 meter high over the ground.