Enhancing agricultural semantic segmentation: architectural innovations and strategies for reducing model complexity

Abstract

Agricultural image semantic segmentation plays a vital role in precision agriculture, enabling accurate analysis of visual data to enhance crop management and optimize resource use. However, achieving high segmentation accuracy while maintaining computational efficiency remains challenging, particularly for real-time systems and edge devices. This thesis presents a two-phase research effort toward an efficient and scalable segmentation framework for high-resolution agricultural imagery. In the first phase, an effective model was developed using a novel Dual Atrous Separable Convolution (DAS-Conv) module integrated into a DeepLabV3 backbone. The DAS-Conv module optimizes the balance between dilation rates and padding size to enhance contextual representation without extra computational cost, while a skip connection between encoder and decoder stages improves fine-grained feature recovery. Despite its lightweight design, the model achieved strong results on the Agriculture-Vision benchmark, demonstrating over 66% higher efficiency compared to transformer-based state-of-the-art models. In the second phase, the framework was extended to DAS-SK, which integrates Selective Kernel (SK) attention into the DAS-Conv module to strengthen multi-scale feature learning and adaptability. The enhanced Atrous Spatial Pyramid Pooling (ASPP) module captures both fine local structures and global context, while a dual-backbone design (MobileNetV3-Large and EfficientNet-B3) further improves representation and scalability. Across three benchmark datasets—LandCover.ai, VDD, and PhenoBench—DAS-SK consistently demonstrates superior efficiency–accuracy trade-offs and notable improvements over its predecessor, DAS. On LandCover.ai, DAS-SK achieves 86.25% mIoU, surpassing DAS by +3.17%, using 10.68M parameters and 11.25 GFLOPs. Although it employs slightly more parameters than DAS, the model remains far lighter than hybrid systems such as Ensemble UNet and transformer models like SegFormer MiT-B2. DAS-SK also achieves higher overall efficiency compared with DAS, demonstrating that the added SK attention and dual-backbone design translate directly into improved segmentation quality. A similar trend is observed on the VDD dataset. DAS-SK attains 79.45% mIoU, improving on DAS by +2.25%, while operating with 10.68M parameters and 43.52 GFLOPs. Although the smaller DAS backbone enables slightly higher FPS, DAS-SK delivers the best accuracy–efficiency balance overall, achieving the highest efficiency score of 9.12%, outperforming transformer models whose parameter counts range from 27M to 234M. On the PhenoBench dataset, DAS-SK again provides the highest performance, reaching 85.55% mIoU compared to DAS at 82.53% (+3.02% improvement). The computational cost remains moderate at 45.00 GFLOPs, versus 25.23 GFLOPs for DAS, but the efficiency gain is substantial—10.09% for DAS-SK versus 7.43% for DAS—highlighting the benefits of improved feature selection and multi-scale fusion. Despite introducing a modest computational increase (typically 1.8× more GFLOPs), DAS-SK consistently delivers 2–3% higher mIoU and markedly stronger multi-scale feature discrimination than its ancestor DAS. Combined with parameter counts that remain significantly below those of modern transformer and hybrid models, DAS-SK offers a practical, lightweight, and high-performing solution for real-time agricultural monitoring and remote sensing in resource-limited environments, where accuracy, efficiency, and scalability are equally critical.