Exploring Novel CNN Architectures for Weed Seedling Recognition in Precision Agriculture

Abstract

Precision agriculture (PA) aims to maximize crop yields while minimizing inputs like water, fertilizer, and pesticides. To achieve this, PA relies on advanced technologies such as sensors, drones, and satellite imagery to monitor crops and optimize inputs. However, weeds pose a significant challenge, competing with crops for vital resources and thereby reducing production output. For weeds to be managed and controlled effectively, they must be categorized accurately. Effective weed management requires understanding each weed's characteristics, which can be challenging with traditional methods. In our research, a comprehensive investigation of 14 hybrid ResNets and 2 SqueezeNets architectures to classify multiclass weed seedlings was conducted. They include ResNet(18, 34, 50, and 101), XResNet(18, 34, 50, and 101), XSEResNet(18, 34, 50, and 101), SqueezeNet (1.0, 1.1). The results demonstrate that the ResNet 101 model achieves superior performance with 90% accuracy, surpassing other architectures. It is a deep architecture that can capture more complicated associations in data, which explain its greater performance. Moreover, it was observed that the XSEResNets exhibit a smoother loss curve, which could be attributed to its channel weighting mechanism. This comprehensive analysis establishes ResNet 101 as the most effective pre-trained CNN model within the Fast.ai library for weed seedling classification in PA applications, provided sufficient computational resources are available.