Application of deep learning based on thermal images to identify the water stress in cotton under film-mulched drip irrigation

Integrated with deep learning algorithms, machine vision techniques have emerged as a robust method for the swift and non-destructive assessment of crop moisture status across extensive agricultural landscapes. Within the agricultural sector, where water serves as a crucial resource, especially for cash crops such as cotton, precise monitoring of water status in fields is paramount. This study presents a novel approach for the evaluation of thermal imagery of cotton crops to predict their water status. The research focused on the application of a methodology that harnesses deep learning architectures in tandem with extant local irrigation practices. A considerable dataset, comprising approximately 5200 images, was collected under conditions of submembrane drip irrigation conditions. Cross-validation was employed to evaluate five deep learning models: VGG16, ResNet-18, MobilenetV3, DenseNet-201, and CSPdarknet53 for training and testing purposes. Experimental outcomes indicated that the MobilenetV3 model outperformed other deep learning architectures. It demonstrated exceptional capability in identifying cotton water stress classes, achieving an F1 value of 0.9990, with an average processing time of 44.85 ms. These results underscore the effectiveness of utilizing deep learning algorithms for accurately assessing water stress in cotton under mulched drip irrigation. This establishes a foundational basis for the development of a precise, cost-effective, and real-time monitoring system for the management of water in cotton fields.