Crop and Weed Segmentation Using ResNet-Unet Architecture

The effectiveness of a ResNet-Unet (RU) architecture for semantic segmentation of soil and weeds from sorghum plants in high-resolution UAV photos is investigated. The model makes use of Unet's encoder-decoder structure for precise segmentation and ResNet's deep residual learning for efficient feature extraction. To increase the resilience of the model, data augmentation methods such random rotations, flips, and horizontal and vertical adjustments were used, along with contrast enhancement methods like Gamma Correction and Contrast Limited Adaptive Histogram Equalization (CLAHE). Utilizing the ResNet-Unet framework, we were able to attain a maximum SØrensen-Dice Coefficient (DS) of 0.929, accompanied by a standard deviation of 0.0041. Classification measures show impressive accuracy: among background, sorghum, and weed classes, precision, recall, and F1-score range from 93.01\% to 99.68\%. The efficiency of RU architecture for precision agriculture applications is demonstrated by this study, which also provides a viable option for targeted weed management and automated weed detection.