A Weakly Supervised Deep Learning Framework for Weed Segmentation and Species Classification in Field Images
TL;DR
A weakly supervised design framework to overcome the missing ground-truth segmentation masks and a very cost-effective, end-to-end optimized solution that bridges the gap between coarse, already available datasets and fine-grained precision required for practical field deployment are provided.
Abstract
The core of precision agriculture lies in the effective management of weed growth, a practice that is crucial for achieving maximum crop yield while using only a required amount of herbicides. Although deep learning has improved automation at a high level, a challenge still remains in the area of high performance segmentation models: pixel-level annotations, which are expensive and rarely available in real-world datasets. Addressing this problem, we proposed a weakly supervised design framework to overcome the missing ground-truth segmentation masks. Instead of relying on manual tracing, our approach is made up of a two-stage pipeline using the bounding box annotations to create pseudo-masks with vegetation indices. This innovative idea allows the DeepLabV3+ model to train and learn to distinguish weed structures from complex soil textures and crop backgrounds. In the further stage, the obtained background-suppressed are fed as input into convolutional neural network architectures for species-level classification. The results show that despite the presence of class imbalance problem and high inter-class similarity, the background suppression strategy provides a very high classification accuracy up-to approximately 88%. To validate that model activations align with plant morphological structures such as leaf margins and venation patterns rather than environmental context, gradient-weighted class activation mapping (Grad-CAM) is applied, with quantitative analysis confirming improved morphological focus in segmentation-preprocessed models. Importantly, bounding box annotations are required only during training, at inference time, the pipeline accepts only a raw RGB field image with no annotation input. Finally, this study provides a very cost-effective, end-to-end optimized solution that bridges the gap between coarse, already available datasets and fine-grained precision required for practical field deployment.