Abstract. Thermal imaging plays a critical role in agriculture, particularly in calculating crop evapotranspiration (ET). However, many drone and satellite systems lack thermal imaging sensors and are limited to capturing RGB (Red, Green, and Blue) color images and near-infrared images or red edge images. Recent advancements in deep learning have demonstrated the potential of RGB-to-thermal image translation, but research specifically targeting crop images remains scarce. This paper introduces a multimodal color-to-thermal translation model based on a Generative Adversarial Network (GAN) architecture. The proposed model generates thermal images from multispectral color inputs, including RGB, RGB+NIR (Near InfraRed), and RG+NIR+Red Edge images through encoding-decoding process (with more details on the innovation part). To evaluate the model‘s performance, quantitative metrics such as per-pixel root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) were employed to compare the synthesized thermal images against actual thermal imagery. These generated thermal images were subsequently used to estimate crop ET and water use efficiency in intercropping systems during the 2023–2024 growing seasons. Additionally, the statistical significance of the results was analyzed across different cropping systems and irrigation schemes. The results shows that the proposed GAN architecture and multimodal image translation algorithm can achieve RMSE of 0.0076, PSNR of 48.7051, and SSIM of 0.4677 from RGB to thermal imagery, RMSE of 0.0067, PSNR of 50.0192, and SSIM of 0.4699 from RGB-NIR to thermal imagery, RMSE of 0.0075, PSNR of 48.8856, and SSIM of 0.4668 from RGB-IR-RedEdge to thermal imagery. The findings demonstrate that the developed multispectral-to-thermal computing framework is a promising tool for estimating ET and optimizing water and nutrient management in agricultural fields.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2025) studied this question.