关键词: 水稻钾素, 胁迫程度识别, ResNet34, 特征融合, 注意力机制

Abstract:

To achieve accurate and fast identification of potassium nutrient stress levels in rice, a rice potassium identification method based on SAE-ResNet34 was constructed with ResNet34 as the core. Taking the late rice ‘Huanghuazhan’ as the research object, six potassium application gradient field trails were set up, while the total fertilization amounts were 0 (K1), 3.78g·m⁻² (K1), 9.45g·m⁻² (K2), 14.17g·m⁻² (K3), 18.90g·m⁻² (K4) and 28.35g·m⁻² (K5), respectively. Based on the high−resolution leaf images data obtained from scanning the fully expanded upper three leaves from tillers during rice tillering and jointing stages. Enhanced super−resolution generative adversarial network (ESRGAN) was incorporated at the data pre−processing stage. Within each residual block, the ReLU activation function was replaced by the Swish activation function, a feature fusion structure based on Addition was designed, and the Efficient local attention (ELA) mechanism module was introduced, with a view to solving the problems of partial feature value loss and low classification accuracy of the network model caused by the reduction of resolution after image resize. The results showed that the rapid recognition method based on SAE−ResNet34 achieved the average recognition accuracy of 82.87% and 84.58% for the six stress levels at the tillering and jointing stages, respectively, on the validation set of the self−constructed rice dataset, which were 7.1 percentage points and 6.7 percentage points higher than that of the original ResNet34 network. The results of confusion matrix showed that the best recognition accuracy for the stress levels at the tillering and jointing stages, were 83.67% for the K3 treatment and 89.11% for the K4 treatment, respectively. Compared with image classification networks such as VGG16, ResNet50 and Swin Transformer, the SAE−ResNet34 network was only slightly behind VGG16 in terms of precision, recall and F1 score, and had the shortest time consumed for 250 rounds of training iterations, with model size of 97.49 MB, which was 7.43MB larger than that of ResNet50 and had the best overall performance. In summary, the identification method based on SAE−ResNet34 network model is able to quickly and accurately identify the degree of potassium nutritional stress during the tillering and jointing stages of rice, which can be used as a scientific reference for nutritional diagnosis and identification of rice and other crops.

Key words: Rice potassium, Stress degree identification, ResNet34, Feature fusion, Attention mechanism