Semi-supervised learning for detection of sedges in sod farms

Abstract

Deep learning-based methods for weed detection and precise herbicide application are promising for reducing herbicide input and weed control costs. However, training the neural network to recognize weeds requires annotating a large number of training images, which is time-consuming and labor-intensive. In addition, in turf sod farms, sods need to be periodically harvested, leading to varying turfgrass and bare soil areas, which increase the complexity of weed detection. To solve this problem, this research explored semi-supervised learning (SSL) methods to train image classification neural networks. The experiments were conducted to compare the training results using different SSL strategies with 100 and 200 labeled images at three image sizes of 240 × 240, 360 × 360, or 480 × 480 pixels. The training dataset images mainly contained purple nutsedge (Cyperus rotundus L.) and green kyllinga (Kyllinga brevifolia) at the pre-flowering or seedhead stage. The F1 score, precision, and recall were used to evaluate the performance of the trained neural networks. The results showed that the network based on the FixMatch SSL strategy trained with the input images of 240 × 240 pixels exhibited the highest F1 score, reaching 98.1% when trained with 100 labeled images and 98.2% when trained with 200 labeled images. To summarize, these results suggest that SSL achieved a great training performance with a small number of annotations. FixMatch SSL was the most effective neural network training strategy evaluated. For the weed detection task, it was observed that neural networks trained using an input image size of 240 × 240 pixels exhibited superior performance compared to the networks trained with other image sizes. In addition, employing the SSL method with only 200 labeled images enhanced the performance of the neural network, surpassing that of fully supervised learning (FSL) approaches.