Abstract: Crop drought diagnosis has been an urgent need to be solved with the development of precision irrigation. It is a high demand to improve the accuracy and timeliness of crop water phenotype diagnosis in the field of intelligent irrigation. In this study, an intelligent diagnosis of winter wheat drought was proposed using improved machine vision. Field experiments were carried out to verify. Three treatments were set in the pit test: suitable, moderate drought and severe drought treatment. The early RGB HD images were capture from the winter wheat by digital camera. The sensitive areas of wheat images were segmented by K-means clustering with the improved by HSV color space. The color and texture features were then extracted as principal components. The support vector machine (SVM) that optimized by bat algorithm (BA) was used to classify the data after dimensionality reduction. Compared with the SVM with/without optimization by Genetic Algorithm, the BA-SVM model was more efficient in diagnosis. The K-means clustering combined with HSV color shared the better segmentation than the traditional one. There were the extract seven-dimensional color features of red channel (R), green channel (G), blue channel (B), brightness (V), 2G-R, 2G-B, (G+R+B)/3 and four-dimensional texture features Energy, Homogeneity, Contrast, Correlation. The principal component analysis was used to reduce the dimensionality of 11 dimensional features. There were the first 3-dimensional principal components with a cumulative contribution rate of 97.2%. The BA was used to optimize the penalty factor and kernel parameters of SVM. The recognition accuracy after dimensionality reduction by principal component analysis was superior to other feature combinations. The recognition accuracy was 96.5% and the running time was 31 s, which was 9%, 7.6% and 4.8% higher than those of color feature, texture feature and color + texture feature, respectively. Compared with the genetic algorithm SVM (GA-SVM) and un-optimized SVM, the dimensionality reduction features were improved by 6.5% and 9.3%, and shortened by 7 and 14 s, respectively. Therefore, the intelligent diagnosis of winter wheat drought was constructed using improved algorithms, such as image segmentation, feature extraction, data dimensionality reduction, data recognition and classification. The finding can provide the real-time diagnosis of winter wheat drought for the decision-making on intelligent irrigation.