AUTHOR=Li Jianian , Gao Long , Wang Xiaocheng , Fang Jiaoli , Su Zeyang , Li Yuecong , Chen Shaomin 

TITLE=Lightweight rice leaf spot segmentation model based on improved DeepLabv3+

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 16 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2025.1635302

DOI=10.3389/fpls.2025.1635302

ISSN=1664-462X

ABSTRACT=IntroductionRice is an important food crop but is susceptible to diseases. However, currently available spot segmentation models have high computational overhead and are difficult to deploy in field environments.MethodsTo address these limitations, a lightweight rice leaf spot segmentation model (MV3L-MSDE-PGFF-CA-DeepLabv3+, MMPC-DeepLabv3+) was developed for three common rice leaf diseases: rice blast, brown spot and bacterial leaf blight. First, the lightweight feature extraction network MobileNetV3_Large (MV3L) was adopted as the backbone of the model. Second, based on Haar wavelet downsampling, a multi-scale detail enhancement (MSDE) module was proposed to improve decision-making ability of the model in transitional regions such as spot gaps, and to improve the sticking and blurring problems at the boundary of spot segmentation. Meanwhile, the PagFm-Ghostconv Feature Fusion (PGFF) module was proposed to significantly reduce the computational overhead of the model. Furthermore, coordinate attention (CA) mechanism was incorporated before the PGFF module to improve robustness of the model in complex environments. A hybrid loss function integrating Focal Loss and Dice Loss was ultimately proposed to mitigate class imbalance between disease and background pixels in rice disease imagery.ResultsValidated on rice disease images captured under natural illumination conditions, the MMCP-DeepLabv3+ model achieved a mean intersection over union (MIoU) of 81.23% and mean pixel accuracy (MPA) of 89.79%, with floating-point operations (Flops) and the number of model parameters (Params) reduced to 9.695 G and 3.556 M, respectively. Compared to the baseline DeepLabv3+, this represents a 1.89% improvement in MIoU, a 0.83% increase in MPA, alongside 93.1% and 91.6% reductions in Flops and Params.DiscussionThe MMPC-DeepLabv3+ model demonstrated superior performance over DeepLabv3+, U-Net, PSPNet, HRNetV2, and SegFormer, achieving an optimal balance between recognition accuracy and computational efficiency, which establishes a novel paradigm for rice lesion segmentation in precision agriculture.