|
|
:
|
To facilitate weed identification across various regions of the country by deep learning, artificial intelligence, and image processing technologies, a comprehensive project entitled “Smart Agriculture Using the Internet of Things” was implemented in 2021 across twelve provinces, including those in Tehran, Razavi Khorasan, Semnan, West Azerbaijan, Ardabil, Khuzestan, Fars, Sistan, Golestan, Gilan, Hamedan, and Kermanshah agricultural research centers. The project was conducted in five major phases: (1) identification and documentation of dominant weed species in agricultural fields; (2) image data collection of different weed growth stages; (3) database development; (4) algorithm design and programming; and (5) system testing and validation. Image processing and selection for the final dataset involved 10 procedure steps: image acquisition, quality enhancement, restoration, color image processing, wavelet and multi-resolution process, compression, morphological processing, segmentation, representation and description, and, ultimately, recognition. Experimental results demonstrated that the MobileNetV3 model, despite its relatively simple architecture, achieved a 92% accuracy under controlled conditions and 88% accuracy under field conditions. These findings indicate that in agricultural settings, lightweight and optimized neural network models often outperform more complex architectures, mainly because they can be effectively deployed on cost-efficient hardware platforms. The study found that transfer learning played a key role in the model's success. By starting with pre-trained weights and fine-tuning the network with locally collected data on Iranian weeds, this approach significantly reduced the data requirements and training time, while ensuring high performance and adaptability to regional agricultural conditions.
|