Machine Learning-Based Measurement Forecasting Approach for Smart Agriculture
TRANSACTIONS on ENVIRONMENT and DEVELOPMENT 2025
Artūrs Ķempelis, Andrejs Romānovs, Antons Patļins, Rasa Bruzgiene

This work explores whether a low-resolution thermal camera can estimate three discrete sensor measurements on a resource-constrained IoT node. Correlation analysis showed that individual thermal pixels correlate strongly with air temperature, negatively with relative humidity and positively with light intensity. Three lightweight regressors VGG CNN, ViT-Tiny and CvT-Tiny were trained from 1 053 single channel 120 x 160-pixel thermal frames to estimate sensor measurements. Experimental tests confirmed the CNN superiority as it achieved RMSE of 2.29 °C and R² of 0.978 (estimating air temperature), RMSE of 0.075 %RH and R² 0.897 (estimating relative air humidity) and RMSE of 0.059 lux and R² of 0.924 (estimating light intensity), outperforming ViT-Tiny and CvT-Tiny on humidity and light intensity estimation. The findings demonstrate that convolutional models remain critical for lightweight and accurate environmental measurement estimation in edge deployments.


Atslēgas vārdi
Convolutional Network Regression, Vision Transformers, Plant Sensor Data Estimation, Thermal imagery, Deep learning, Measurement Forecasting, Precision Agriculture
DOI
10.37394/232015.2025.21.78
Hipersaite
https://wseas.com/journals/ead/2025/b585115-034(2025).pdf

Ķempelis, A., Romānovs, A., Patļins, A., Bruzgiene, R. Machine Learning-Based Measurement Forecasting Approach for Smart Agriculture. TRANSACTIONS on ENVIRONMENT and DEVELOPMENT, 2025, Vol. 21, No. 78, 937.-949.lpp. e-ISSN 2224-3496. Pieejams: doi:10.37394/232015.2025.21.78

Publikācijas valoda
English (en)
RTU Zinātniskā bibliotēka.
E-pasts: uzzinas@rtu.lv; Tālr: +371 28399196