Appearance quality identification and environmental factors tracing of Lyophyllum decastes for precise environment control using knowledge graph

Abstract

In the factory production of Lyophyllum decastes, inappropriate cultivation environments can lead to appearance quality issues, which in turn affect both yield and quality. However, the appearance characteristics of Lyophyllum decastes influenced by environmental factors share similarities, and the environmental factors that cause appearance quality problems exhibit coupling and complexity. Therefore, the identification of appearance characteristics and tracing of environmental factors present significant challenges. To address this issue, this paper proposes a multimodal learning network, DCRes-GAT, which integrates an improved Residual Neural Network (DCResNet) and a Graph Attention Network (GAT) to accurately identify the features of Lyophyllum decastes, while simultaneously tracing environmental factors and providing control recommendations. First, a knowledge graph based on the prior knowledge of quality and environmental factors is constructed, mapping this information to a point space and extracting key features. Next, DCResNet is employed to extract optical features from Lyophyllum decastes images. In addition, the receptive field is expanded through dilated convolutions, while pixel-level details are preserved, and a Convolutional Block Attention Module (CBAM) is incorporated to identify subtle visual differences. Finally, a dot product operation fuses point-space features with visual features, achieving accurate identification of characteristics and providing suggestions. Experimental results demonstrate that the DCRes-GAT model performs excellently, with a feature identification accuracy of 99.45%, and can precisely diagnose key environmental factors that cause appearance quality problems, achieving a diagnostic accuracy of 99.84%. This provides a basis for the precise control of the cultivation environment of Lyophyllum decastes.