A Novel Hybrid Approach for Foreground Extraction in Precision Agriculture

Modernizations in precision agriculture have transformed data analyses and crop management strategies. It redefines the farming practices, enhancing productivity and sustainable resource management. Foreground segmentation is crucial to empower the exact identification for the accurate crop analysis and other significant elements in agricultural practices. This work involves the unsupervised approach combined with the pretrained network UNet, VGG16, ResNet50 followed by the clustering method K Means and DBSCAN to extract the foreground regions from plant images. Furthermore Class Activation Mapping were employed to generate heatmaps to highlight the area of high importance assisting in accurate segmentation. The result revealed that this collaborated work brings the output with the accuracy and F1 score above 70%for the complex maize weed dataset, and above 85%for crop weed detection dataset and crop weed field images dataset.