Yang et al. (2026) Leaf thermal infrared imaging and lightweight deep learning enable early detection of water stress in watermelon for precision irrigation

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-04-11
• Authors: Qi Yang, Dongqing Lao, Yufei Wu, Chong Liu, Zipeng Zhang, Zhihao Li, Tianhao Zhao, Paramasivan Balasubramanian, Fayong Li
• DOI: 10.1016/j.agwat.2026.110344

Research Groups

• College of Water Resources and Architectural Engineering, Tarim University, Xinjiang, China
• Key Laboratory of Tarim Oasis Agriculture (Tarim University), Ministry of Education, Alaer, China
• College of Information Engineering, Tarim University, Xinjiang, China
• Department of Chemical & Materials Engineering, University of Auckland, New Zealand
• Department of Biotechnology & Medical Engineering, National Institute of Technology Rourkela, India

Short Summary

This study proposes a thermal-imaging-based deep learning approach to classify watermelon water-stress status for precision irrigation. It systematically evaluates nine deep learning models, identifying EfficientNet-B0 as the most suitable for field deployment due to its optimal balance of high accuracy (0.99) and computational efficiency (0.39 GFLOPs, 8.81 ms inference latency).

Objective

• To construct a thermal image dataset of watermelon leaves capturing early physiological responses to water stress.
• To perform a systematic cross-architecture evaluation of nine representative deep learning models (classical convolutional networks, lightweight networks, and Vision Transformers), assessing recognition accuracy, parameter count, computational complexity, and inference time to guide lightweight deployment on edge devices.
• To develop a method for water stress identification that directly leverages phenotypic features in leaf thermal images, enhancing applicability and robustness under complex field conditions and supporting precision irrigation decisions.

Study Configuration

• Spatial Scale: Field experiment conducted at a modern agricultural training base in southern Xinjiang, China (40°38.43′ N, 81°5.41′ E). The study involved six experimental plots, each 45 square meters (3 meters × 15 meters), with a row spacing of 50 centimeters and a plant spacing of 30 centimeters. The focus was on plant-scale leaf thermal images.
• Temporal Scale: The watermelon growth period was 95 days (sowing on 18 April 2025, harvesting on 20 July 2025). Thermal image data was collected from the vine elongation stage through the maturity stage, specifically on the second day post-irrigation, under clear-sky conditions, at multiple time points (09:00–21:00).

Methodology and Data

• Models used: VGG16, ResNet-34, DenseNet-201, MobileNetV2, EfficientNet-B0, CSPDarknet53, RegNetY, ConvNeXt, and Vision Transformer (ViT).
• Data sources:
  • Leaf thermal images acquired using a low-cost FLIR thermal camera (FLIR ONE Pro for Type-C) mounted on a smartphone.
  • A total of 1200 raw thermal images were collected (651 under normal irrigation, 549 under water-stress conditions).
  • The dataset was augmented to 2168 images using rotation, flipping, and motion blur.
  • Soil moisture sensors (BC202, HydraProbe Lite) were used for continuous monitoring of volumetric water content (VWC) in the crop root zone.
  • Environmental data included daily maximum and minimum air temperatures and precipitation.

Main Results

• EfficientNet-B0 demonstrated the best overall trade-off between classification performance and computational efficiency, achieving an accuracy of 0.9907 and an F1 score of 0.9950, with only 5.3 million parameters, 0.39 GFLOPs, and an inference latency of 8.81 milliseconds.
• The Vision Transformer (ViT) achieved the highest classification accuracy (0.9938) and F1 score (0.9943 for non-stress, 0.9932 for water stress), with perfect precision for non-stressed samples and perfect recall for stressed samples. However, it required significantly more resources (86.6 million parameters, 17.6 GFLOPs).
• McNemar’s test indicated no statistically significant performance difference between EfficientNet-B0 and ViT, DenseNet-201, or CSPDarkNet53 (p-values > 0.05).
• Transfer learning with ImageNet pre-trained weights significantly accelerated model convergence and improved validation accuracy (e.g., EfficientNet-B0 reached ~0.99 accuracy compared to ~0.90 from scratch).
• Water-stressed samples were generally more prone to misclassification across all models, but ViT and EfficientNet-B0 exhibited minimal false negatives.

Contributions

• Provides a systematic and comprehensive comparison of nine diverse deep learning architectures for thermal image-based crop water-stress classification, including classical, lightweight, and modern models.
• Evaluates models not only on recognition accuracy but also on deployment-oriented efficiency metrics (parameter count, computational complexity, inference latency), offering practical guidance for model selection in resource-constrained edge computing environments.
• Develops a robust method for water stress identification that directly leverages phenotypic features in raw leaf thermal images, reducing dependence on absolute temperature measurements, stringent sensor calibration, and environmental stabilization.
• Establishes a thermal image dataset of watermelon leaves specifically designed to capture early physiological responses to water stress.
• Enables early detection of water stress before visible symptoms appear, facilitating proactive and timely irrigation interventions to improve water-use efficiency and mitigate yield losses in arid agricultural systems.

Funding

• Earmarked Fund for XARS (XARS-06)
• National Natural Science Foundation of China (42275014)
• Bingtuan Science and Technology Program (2024YD004)
• Bingtuan Guiding Science and Technology Plan Program (2024ZD098)
• President's foundation of Tarim University (TDZKCX202404)
• Graduate Research Innovation Project of Tarim University (TDGRI2024083)

Citation

@article{Yang2026Leaf,
  author = {Yang, Qi and Lao, Dongqing and Wu, Yufei and Liu, Chong and Zhang, Zipeng and Li, Zhihao and Zhao, Tianhao and Balasubramanian, Paramasivan and Li, Fayong},
  title = {Leaf thermal infrared imaging and lightweight deep learning enable early detection of water stress in watermelon for precision irrigation},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2026.110344},
  url = {https://doi.org/10.1016/j.agwat.2026.110344}
}