Knowledge-guided multi-task learning with cascade out-of-distribution rejection for rice leaf disease diagnosis

Accurate diagnosis of rice leaf diseases is essential for timely crop protection, but most deep learning classifiers offer little insight into the visual evidence behind their predictions and assign a disease label even to non-target images. This study aimed to develop a diagnostic framework that integrates phytopathological knowledge into deep learning and rejects out-of-distribution (OOD) inputs. We propose the Knowledge-Guided Multi-Task Rice Network (MTRNet), which attaches a disease-diagnosis head and three auxiliary heads for lesion shape, color, and location to a shared ResNet-50 backbone. An expert knowledge matrix automatically generates the attribute labels, providing knowledge-guided supervision without pixel-level annotation. At inference, a two-stage Cascade Inference System (CIS) combines a color-based biological plausibility filter with a consistency check between the predicted disease and its predicted attributes. MTRNet was evaluated on a public dataset of 5,932 images of four rice diseases (Leaf Blast, Brown Spot, Bacterial Leaf Blight, and Tungro) and on 1,000 OOD images, using source-aware five-fold cross-validation that keeps all copies of each source photograph within a single partition. MTRNet achieved an accuracy of 96.32% ± 1.45% and a macro-F1 score of 96.43% ± 1.34%, while providing attribute-level predictions for every diagnosis. The CIS rejected 77.78% ± 0.63% of OOD images while retaining 91.71% of genuine diseased leaves, and a split-sample analysis showed that threshold selection introduced negligible optimism (about 0.2 percentage points). The framework adds only 0.027 M parameters to the backbone, and the plausibility filter requires 0.34 ms per image. These results indicate that combining knowledge-guided multi-task learning with cascade rejection enables accurate, attribute-informed, and computationally efficient rice disease diagnosis with a built-in safeguard against non-target inputs, providing a practical basis for reliable image-based disease screening in agriculture.