Target leaf spot is one of the main fungous diseases of cucumber. Under suitable conditions, especially under the conditions of the large temperature difference between day and night or saturated humidity, the disease develops rapidly, leads to the reduction of cucumber yield and brings economic losses. The cucumber target leaf spot segmentation can provide an effective basis for the identification and diagnosis of cucumber disease, which has great significance. In this study, a cucumber spectral image was taken as the research object, and U-net deep learning network was utilized to construct the semantic segmentation model for cucumber target leaf spot segmentation. Firstly, the regions with more prominent lesions in the visible spectrum images were selected for training and testing. We captured 135 regions out of 40 images as samples, and each region was 200×200 pixel. The Image labeler tool of Matlab was used to label the samples to mark the affected area and the healthy area. Then, the U-net network was constructed, which contains 46 layers and 48 connections. The cucumber target leaf spots’ feature extraction is completed by convolution layer, ReLU layer and max-pooling. The upsampling is completed by deep connection layer, up convolution layer and up-ReLU. The copy and crop operations and feature fusion are completed by skip connection. The U-net was used for training to get the semantic segmentation model. From 135 samples, 96 were randomly selected as training samples and the remaining 39 as test samples. Set the iterations 240, L2 regularization coefficient 0.000 1, initial learning rate 0.05, momentum parameter 0.9, gradient threshold 0.05, and then utilize the samples for training and testing. After 10 repeated training and testing, the results showed that the average execution time of the semantic segmentation model based on U-net and visible spectrum images was 46.4 s. The average memory occupation was 6 665.8 MB, and it shows that the model has a high execution efficiency. The pixel accuracy of the model was 96.23% ~ 97.98%, mean pixel accuracy was 97.28%~97.87%, mean intersection over union was 86.10%~91.59%, frequency weighted intersection over union was 93.33%~96.19%. It shows that the model has good stability and strong generalization ability. This research used less training samples to obtain a segmentation model with high accuracy, which provides a reference for small sample machine learning and provides a method basis for other vegetable disease spot segmentation, disease identification and diagnosis.