Efficient Super-Resolution for Resource-Constrained Precision Agriculture: A Loss Function Optimization Approach

In smart agriculture, vision-based crop monitoring systems require high-quality imagery along with efficient computation and low power consumption to support accurate analysis and real-time decision-making in resource-constrained edge environments. However, in digital image processing, image quality and computational performance present a trade-off, where increasing reconstruction quality typically increases model complexity and resource requirements. This study addresses this challenge by proposing a lightweight super-resolution (SR) approach optimized for real-world edge applications in agriculture. Unlike existing SR methods that rely on complex architectures or synthetic datasets, this work focuses on loss function-level optimization to improve perceptual image quality without increasing computational cost and power consumption, making it applicable to support resource-constrained precision smart agriculture. ESPCN was chosen as the baseline due to its efficiency, and was further optimized by replacing the conventional Mean Squared Error (MSE) loss with L1 loss to improve robustness to noise and lighting variations. Experimental results on a real-world lettuce dataset captured using ESP32-CAM show that the proposed method produces better texture quality in both day and night conditions. Structural evaluation using Laplacian-based edge density and sharpness, supported by statistical analysis, confirmed improved preservation of high-frequency details without additional computational or energy costs. These findings demonstrate that lightweight loss function optimization provides a practical and energy-efficient solution for improving image quality in edge-based agricultural monitoring systems.