Non-invasive monitoring of lettuce growth in IoT aeroponics through thermal-sensing-driven transformer network

Transformer-guided thermal imagery assists in effective observation of aeroponic plant growth. These networks enable fast and accurate analysis of sensor data, yet the precision relies heavily on architectural selection, hyper-parameter tuning and the quality of the data input. This research tackles these constraints by formulating a transformer-based approach that leverages tri-band thermal imaging for precise lettuce growth assessment in IoT-based aeroponic systems. The thermal images are adjusted to three distinct sets of properties which are compatible with transformer networks. The initial features encompass raw temperature (RT), normalized temperature (NT), and Sobel edge gradient (SEG); the second set embraces raw and normalized temperature, and CLAHE (contrast limited adaptive histogram equalization); and the third set assembles raw temperature, temperature difference (TD), and local average filtering (LAF). Multiple transformer-based approaches, such as the mobile vision transformer (MobileViT), the vision transformer (ViT), the efficient attention network (EANet), and the shifted window transformer (Swin), were applied to analyze the proposed hybrid thermal spectra. The patch-based grid masking technique was deployed to augment the dataset and optimize training. The findings indicated that the RT-TD-LAF-derived features, when fused as thermal signals in ViT-driven networks, significantly boosted the precision of plant growth status evaluation. The developed network attained 99.8% for classification precision (Pr), validation accuracy (Ac), F-measure (Fm), and recall (Re), and 99.5% in intersection over union (IoU). Moreover, it surpassed a traditional RGB-based ViT model (Ac, Pr, Fm = 0.895, Re = 0.898, and IoU = 0.810). Lastly, this work could support the farming community by facilitating accurate observing of plant growth and ensuing proactive decisions for sustainable aeroponic farming.