Abstract: Aiming at the challenges facing traditional machine learning algorithms such as cumbersome modeling steps for hyperspectral data and conventional convolutional neural networksnot being very expressive in details on hyperspectral images, a network structure based on multi-scale feature fusion was designed. By up-sampling and pooling layer parameter optimization, feature layers of different depths in 1D-CNN were fused to obtain richer discriminative features for hyperspectral. The network training used unique thermal coding for labeling training, which solved the problem of the classifier having difficulty in processing attribute data and alsoplayed a role in expanding the features to a certain extent. The results showed that the accuracy of classifying regions of interest using multi-scale feature fusion 1D-CNN in feature classification experiments was improved by 63.99% and 5% compared to SVM and conventional 1D-CNN networks. In the deficiency recognition experiments, the recognition accuracies of nitrogen deficiency, phosphorus deficiency, and potassium deficiency, as well as normal potato leaves, were above 99%, and were improved by 1.7%, 6.82%, 2.99%, and 24.8%,respectively, using the proposed algorithm compared with SVM. Compared with conventional 1D-CNN, the accuracy of recognition of normal leaves, potassium-deficient leaves, and phosphorus-deficient leaves was improved by 0.03%, 0.17%, and 0.76%, respectively.The fusion of hyperspectral information features at multiple scales and the combination of 1D-CNN for feature extraction can improve the accuracy of feature classification of hyperspectral images as well as the accuracy of deficiency identification of potato plants.