基于多模态知识图谱的水稻施肥期判别方法

Rice Fertilization Period Discrimination Method Based on Multi-modal Knowledge Graph

摘要:

目前水稻施肥时间的确定依赖于传统经验与人工巡田观察的综合判断，难以满足当前农业智能化的发展。为此，提出了一种基于多模态知识图谱的水稻施肥期判别方法，综合利用了文本形式的经验信息和图像形式的视觉信息进行施肥期确定。首先构建单模态水稻施肥知识图谱，利用依存句法分析提取返青肥、分蘖肥、穗肥、粒肥4个施肥期对应的跨模态特征短语，结合Chinese CLIP模型得到它们与图像的匹配度以及与施肥期节点的权重后组成新的带有跨模态节点的三元组，完成多模态水稻施肥知识图谱的构建；然后基于多模态知识图谱计算输入信息的综合匹配度，使用大田采集的图像进行交叉验证，综合评估判别方法的准确性和稳定性确定各施肥期的判定阈值，实现对该输入的施肥期判别。以实际采集的各施肥期当日及前、后5 d的600幅图像测试判别方法的准确率，结果表明，基于多模态知识图谱的水稻施肥期判别方法总体准确率达到86.2%，其中粒肥时期判别准确率最高，为90.1%。该施肥期判别方法同时利用文本、图像两种模态的信息，提高了信息利用率，在真实场景下具有判别能力，为水稻施肥期自动确定提供参考。

Abstract:

Currently, the determination of the optimal fertilization time for rice relies heavily on a combination of traditional experience and manual field inspection, which struggles to meet the demands of modern agricultural intelligence. In response, a method for rice fertilization period discrimination was introduced based on a multi-modal knowledge graph, integrating textual experiential information and visual cues for determining the fertilization period. Initially, a single-modal knowledge graph for rice fertilization was constructed. On this basis, cross-modal feature phrases corresponding to the four fertilization periods (re-greening, tillering, heading, and grain-filling) were extracted by using dependency syntax analysis. These phrases were then combined with the Chinese CLIP model to determine their match with images and their respective weights for the fertilization periods, forming new triplets with cross-modal nodes. This led to the creation of a multi-modal rice fertilization knowledge graph. Subsequently, the multi-modal knowledge graph was used to calculate the comprehensive matching degree of input information, and field-collected images were utilized for cross-validation. This process comprehensively evaluated the accuracy and stability of the discrimination method, thereby determining the decision thresholds for each fertilization period. The discrimination methods accuracy was tested by using 600 images captured on the day of each fertilization period and five days before and after. Results showed that the overall accuracy rate of the multi-modal knowledge graph-based rice fertilization period discrimination method was 86.2%, with the highest accuracy rate of 90.1% during the grain-filling period. By utilizing both textual and visual modalities, this method enhanced information utilization and demonstrated discriminative capability in real-world scenarios, offering a reference for the automated determination of rice fertilization periods.