Segmentation of Agricultural Fields in Aerial Imagery Using Enhanced Deep Learning Models

Precision agriculture utilizes advanced technologies to enhance crop management and optimize agricultural yield. A crucial component involves segmenting agriculture fields in aerial imagery, which is critical for accurate identification of crops, weeds, and other agricultural traits. However, aerial image segmentation is challenging due to diverse agricultural patterns and visually similar elements, which results in pixel misclassification. This study focuses on segmenting agriculture fields from aerial imagery using UNet with ResNet block, UNet with DenseNet-121 block, and DeepLabV3-Plus models. Despite their robust frameworks, existing models struggle to distinguish between comparable classes due to differences in plant appearance, lighting conditions, and soil backgrounds. The results show that the UNet-ResNet and UNet-DenseNet models provide more accurate and visually consistent segmentations than DeepLabV3-Plus. UNet-ResNet outperformed the other models in terms of accuracy, with 88.16 percent and the lowest loss of 0.48. This study presents insights and recommendations to enhance model performance for accurate agriculture patterns segmentation. Future research should focus on creating models capable of handling nuanced variances in agricultural datasets to promote efficient and sustainable farming techniques.