Abstract:Straw coverage on the soill surface after tillage is a key parameter for evaluating the quality of straw return to the field. Existing methods struggle to effectively identify individual rice straw stalks. To address this, a semantic segmentation method that integrated prior embedding with multi-scale feature fusion was proposed to achieve accurate recognition of individual rice straw stalks. A fixed-threshold segmentation method based on color distance was employed to preprocess the original image and generate a prior map, providing prior information input for subsequent recognition tasks while suppressing background interference. An enhanced U - Net model, MRCF - DA - CDPE, was designed. It employed parallel multi-scale convolutions to capture information features across scales--from fragmented straw fragments to large straw clusters-while utilizing channel and spatial attention to select key features and suppress disturbances like soil bright spots. Continuous distance information from preprocessing was embedded into the network as an additional input channel, providing physical guidance for segmentation. Image preprocessing strategies demonstrated that this approach enhanced the average accuracy of each base model by 2 ~ 4 percentage points, with U - Net achieving the highest recognition accuracy. Validation tests of the improved model demonstrated that the MRCF - DA - CDPE model achieved 86.93% average intersection-over-union ratio (I0U), 94.89% average precision, and 0. 850 2 Kappa coefficient, representing improvements of 2. 72, 2.98 percentage points and 0. 056 7 respectively over the baseline U - Net model. This method achieved precise recognition of individual straw stalks, providing technical support for straw return quality inspection and tillage effectiveness evaluation.
