Thermal imaging provides key physiological metrics such as canopy temperature, stomatal conductance, and crop water stress index. However, high-resolution radiometric thermal cameras remain cost-prohibitive for large-scale agricultural monitoring. Consumer-grade thermal sensors present an affordable alternative, but suffer from low spatial resolution and thermal measurement noise.
We present VisTA-SR, a multimodal deep learning framework designed to enhance both the spatial resolution and temperature measurement accuracy of low-cost thermal cameras. By leveraging high-resolution RGB imagery as a structural guide, VisTA-SR aligns multimodal image pairs and super-resolves low-resolution thermal features. The framework integrates a radiometric calibration pipeline to correct sensor temperature drift. Evaluated on field data collected from garbanzo bean plots, VisTA-SR enables organ-level canopy temperature extraction at a fraction of the cost of industrial thermal imaging systems.
