Deep Learning-Based Semantic Segmentation of Turmeric Crop Images Using the DeepLabV3+ Architecture

In this research, we focus on semantic segmentation in the domain of agricultural image analysis, employing an advanced deep learning architecture known as DeepLabV3+. The objective is to accurately classify pixels in high-resolution images into distinct classes, specifically targeting the segmentation of Space, Disease, and Crops within agricultural portraits. The Atrous Spatial Pyramid Pooling (ASPP) module handles multi-scale context aggregation, the Squeeze-and-Excite block handles channel-wise attention, and the encoder, which is based on ResNet50, extracts features. Additionally, skip connections from lower-level encoder blocks enhance the spatial resolution of the final segmentation map. The input images and corresponding mask images constitute the data set employed for training the model. The model which is trained, is then applied to unseen images for inference, and the results are analyzed, visualized, and discussed. Our research provides insights into the capabilities of state-of-the-art semantic segmentation for precision agriculture applications, showcasing the potential for automating crop monitoring and disease detection.