Drone-based imaging is a non-invasive technique that plays a critical role in monitoring plant health, stress levels, and soil moisture. Accurate transformation of RGB images to thermal maps is a challenging task due to various factors, including environmental noise, crop density, and variable lighting. Existing RGB-to-thermal translation approaches typically depend on implicit convolutional representations that inadequately extract directional crop structures, soil background variations, and multi-scale spatial variability, leading to reduce generalization across agricultural environments. To address these limitations, we advance a novel RGB to thermal map transformation approach (DTBQW-2BCLIFT) that integrates Dual-Tree Biquaternion Wavelet Transform (DTBQWT) with a newly designed two-branch convolutional attention layers framework based on Continual Learning with Informed Feature Transfer (2-BCLIFT). RGB drone images are decomposed into 18 feature maps to capture multi-resolution spatial and directional frequency information. Then, DTBQWT coefficients and raw RGB image are fused to 2BCLIFT. Unlike traditional deep learning approaches, the proposed 2-BCLIFT model is specifically optimized for field-level variation and is evaluated on drone imagery collected from four different agricultural fields in Prince Edward Island, Canada, under potato crop cultivation and various environmental conditions. This enables the proposed model to generalize effectively across diverse fields. Several metrics, including RMSE, PSNR, SSIM, and per-pixel accuracy, are employed to evaluate the proposed model against U-Net, Pix2Pix-GAN, Sparse GAN, Thermal-GAN, Modified Pix2Pix, and Cycle GAN. The results demonstrated that the proposed DTBQW-2BCLIFT model provided performance that is well suited for real world agricultural applications and compares favourably with existing state-of-the-art approaches. It can be combined with drones to detect early signs of water stress and nutrient deficiencies through variations in canopy temperature
Diykh et al. (Sun,) studied this question.