In order to reduce the overuse of herbicides, an automatic spraying system can be utilized, assisted with machine vision-based techniques to accurately target weeds. In this paper, a fusion-based structure has been proposed for weed detection in visible and thermal images of paddy fields. Due to the lack of publicly available multispectral datasets in this topic, first, a freely accessible dataset was generated including 100 pairs of visible and thermal images of rice and weeds. In this new dataset, the segmented plants were labeled into two groups of rice and weeds. A feature vector including 15 morphological, 12 spectral, 10 textural and 11 new thermal features was extracted from segmented objects. The proposed thermal features were extracted from thermal images, while other types of features were extracted from visible/thermal pairs. Then, a genetic algorithm (GA) was used for optimized feature selection. Next, multiple late and early fusion structures at the decision level were developed and compared for weed detection purposes. Applied fusion structures include: multilayer perceptron neural network (MLP), extreme learning machines (ELM) and extreme learning machines ensembles (ELM-E) artificial neural networks. The best results were obtained by ELM with an accuracy of 98.08% in late fusion structure with GA. The best results for classifying plants in visible images were related to the morphological descriptor and the best results for thermal images belonged to the proposed descriptor.

为了减少除草剂的过度使用，可以使用自动喷洒系统，并辅以基于机器视觉的技术来准确定位杂草。在本文中，提出了一种基于融合的结构，用于稻田可见光和热图像中的杂草检测。由于本主题中缺乏公开可用的多光谱数据集，首先，生成了一个可免费访问的数据集，其中包括 100 对水稻和杂草的可见光和热图像。在这个新数据集中，分割的植物被标记为两组水稻和杂草。从分割的对象中提取了一个特征向量，包括 15 个形态特征、12 个光谱特征、10 个纹理特征和 11 个新的热特征。提出的热特征是从热图像中提取的，而其他类型的特征是从可见/热对中提取的。然后，遗传算法（GA）用于优化特征选择。接下来，为了杂草检测的目的，开发并比较了决策级别的多个晚期和早期融合结构。应用的融合结构包括：多层感知器神经网络（MLP）、极限学习机（ELM）和极限学习机集成（ELM-E）人工神经网络。ELM 在与 GA 的后期融合结构中获得了最好的结果，准确率为 98.08%。可见图像中植物分类的最佳结果与形态描述符有关，热图像的最佳结果属于所提出的描述符。应用的融合结构包括：多层感知器神经网络（MLP）、极限学习机（ELM）和极限学习机集成（ELM-E）人工神经网络。ELM 在与 GA 的后期融合结构中获得了最好的结果，准确率为 98.08%。可见图像中植物分类的最佳结果与形态描述符有关，热图像的最佳结果属于所提出的描述符。应用的融合结构包括：多层感知器神经网络（MLP）、极限学习机（ELM）和极限学习机集成（ELM-E）人工神经网络。ELM 在与 GA 的后期融合结构中获得了最好的结果，准确率为 98.08%。可见图像中植物分类的最佳结果与形态描述符有关，热图像的最佳结果属于所提出的描述符。