基于多特征融合的水稻生长期识别模型

Multi-feature Fusion-based Model for Recognizing Different Growth Stages of Rice

摘要:

我国是全球最大的水稻种植国,水稻不同生长期的精准识别对于实现稻田智能化管理至关重要。以水稻生长的始穗期、齐穗期、乳熟期和黄熟期为识别目标,提出了一种基于多特征融合的水稻生长期识别模型。以无人机获取的高分辨率遥感影像作为数据源,分别采用Swin Transformer模型和ResNet18模型提取了影像的多尺度特征和深层语义特征;设计了一种改进的特征融合模块,包括自适应特征融合(AFF)模块和全局上下文建模(GC)模块。AFF模块通过学习特征之间的权重关系,实现不同特征的自适应融合。GC模块引入全局上下文信息,增强模 型对全局特征的感知能力,提高模型对复杂场景的适应性;采用Lion 优化器和动态学习率策略,加速模型的收敛速度。同时,引入Dropout层以防止过拟合,提高模型的泛化能力。对比实验表明,本文提出的模型在水稻抽穗期和成熟期识别任务上取得了97. 14%的准确率,相较于MobileNetV2、EfficientNet和Swin Transformer 分别提高了3. 29、 2. 53、1. 09 个百分点。在保证高准确率的同时,该模型的参数量和计算量相对较少,具有较好的轻量化特性,适用于实际应用场景。

Abstract:

China is the largest rice-growing country in the world, and accurate recognition of different growth stages of rice is essential for achieving intelligent rice field management. Targeting the classification of rice growth stages, including initial heading stage, full heading stage, milk ripening stage, and yellow ripening stage, a recognition model for rice growth stages was proposed based on multi- feature fusion. The specific methodology included multi-feature extraction module, utilizing high- resolution remote sensing images obtained from drones as the data source, multi-scale features and deep semantic features of the images were extracted by using the Swin Transformer model and ResNet18 model. An improved feature fusion module was designed, which included an adaptive feature fusion ( AFF) module and a global context modeling ( GC) module. The AFF module achieved adaptive fusion of different features by learning the weight relationships between features. The GC module incorporated global context information to enhance the model’s ability to perceive global features, thereby improving adaptability in complex scenarios. Model optimization, utilizing the Lion optimizer and a dynamic learning rate strategy to accelerate the model??s convergence speed. Additionally, a Dropout layer was introduced to prevent overfitting and improve the model??s generalization ability. Comparative experiments demonstrated that the proposed model achieved a 97. 14% accuracy rate in recognizing rice at the heading and maturity stages, which was 3. 29, 2. 53, and 1. 09 percentage points higher than that of other mainstream models such as MobileNetV2, EfficientNet and Swin Transformer, respectively. Notably, while maintaining high accuracy, the model had relatively fewer parameters and lower computational costs, showcasing lightweight characteristics suitable for practical applications.