Objective This study aims to evaluate the drought degree of field cotton by studying leaf curling, and to provide basis for precise irrigation and intelligent decision-making of cotton field.

Methods Cotton field moisture gradient test was designed to collect cotton RGB image data in the critical growth period. CottonNet, a detection model based on ShuffleNetV2, was constructed to evaluate the drought degree of cotton in the field, aiming at realizing rapid and accurate drought monitoring under the complex field environment. In order to improve the feature extraction ability of cotton leaves, the SimAM attention mechanism was introduced into the baseline network. Replace Conv5 convolutional blocks in the network with Ghost modules to reduce the number of model parameters and model size to fit resource-constrained devices.

Results The accuracy of CottonNet model in identifying cotton drought degree was 98.02%, which was better than traditional models such as ResNet18. Meanwhile, the number of parameters in Cottonnet model was only 2.13 M, which was 14% less than the baseline model, and the model size was only 4.3 MB, which was 12% less than the baseline model.

Conclusion CottonNet model can improve the recognition accuracy of cotton drought degree, and is more suitable for evaluating cotton drought degree in field.