ABSTRACT
The application of computer vision in agriculture has already contributed immensely to restructuring the existing field practices starting from the sowing to the harvesting. Among the different plant parts, the economic part, the yield, has the highest importance and becomes the ultimate goal for the farming community. It depends on many genetic and environmental factors, so this curiosity about knowing the yield brought several precise pre-harvest prediction methods using different ways. Out of those techniques, non-invasive yield prediction techniques using computer vision have been proved to be the most efficient and trusted platform. This study developed a novel methodology, called SlypNet, using advanced deep learning networks, i.e., Mask R-CNN and U-Net, which can extract various plant morphological features like spike and spikelet from the visual image of the wheat plant and provide a high-throughput yield estimate with great precision. Mask R-CNN outperformed previous networks in spike detection by its precise detection performance with a mean average precision (mAP) of 97.57%, a F1 score of 0.67, and an MCC of 0.91 by overcoming several natural field constraints like overlapping and background interference, variable resolution, and high bushiness of plants. The spikelet detection module's accuracy and consistency were tested with about 99% validation accuracy of the model and the least error, i.e., a mean square error of 1.3 from a set of typical and complex views of wheat spikes. Spikelet yield cumulatively showed the probable production capability of each plant. Our method presents an integrated deep learning platform of spikelet-based yield prediction comprising spike and spikelet detection, leading to higher precision over the existing methods.
ABSTRACT
In recent years, deep learning techniques have shown impressive performance in the field of identification of diseases of crops using digital images. In this work, a deep learning approach for identification of in-field diseased images of maize crop has been proposed. The images were captured from experimental fields of ICAR-IIMR, Ludhiana, India, targeted to three important diseases viz. Maydis Leaf Blight, Turcicum Leaf Blight and Banded Leaf and Sheath Blight in a non-destructive manner with varied backgrounds using digital cameras and smartphones. In order to solve the problem of class imbalance, artificial images were generated by rotation enhancement and brightness enhancement methods. In this study, three different architectures based on the framework of 'Inception-v3' network were trained with the collected diseased images of maize using baseline training approach. The best-performed model achieved an overall classification accuracy of 95.99% with average recall of 95.96% on the separate test dataset. Furthermore, we compared the performance of the best-performing model with some pre-trained state-of-the-art models and presented the comparative results in this manuscript. The results reported that best-performing model performed quite better than the pre-trained models. This demonstrates the applicability of baseline training approach of the proposed model for better feature extraction and learning. Overall performance analysis suggested that the best-performed model is efficient in recognizing diseases of maize from in-field images even with varied backgrounds.
