Assessment of water stress in rainfed maize using RGB and thermal imagery

Non-invasive techniques such as RGB, thermal imaging, and optical sensing can provide information on plant water stress before it occurs. In the present study, the maize crop (Var. DMRH 1301) was cultivated with three different sowing dates (1, 15, and 30 June) under rainfed condition. The RGB and thermal images along with crop growth parameters were recorded periodically along with soil moisture in maize. A deep learning model (DarkNet53) was applied on RGB as well as thermal images for the identification of water stress. The occurrence of rainfall and thereby soil moisture had a significant effect on crop canopy temperature under different dates of sowing. The canopy temperature histogram under water stress showed bimodal distribution whereas unimodal distribution was observed under non-stress condition. The maximum accuracy of the DarkNet53 model at 20 epochs and 8 batch size for the identification of water stress in RGB and thermal imagery was 94.1 and 99.6%, respectively. The testing accuracy of identification of water stress for RGB imagery were 94.1, 90.5, and 92.3% and for thermal imagery 95.7, 91.5, and 99.6% in early, timely and late sown maize, respectively. Overall, thermal imagery input had higher accuracy for water stress identification. The present study showed that non-destructive, rapid vision-based techniques using digital imagery can helpto characterize plant growth, water stress, and thereby, yield.