Machine learning integration of UAS-based thermal and multispectral imaging for precision irrigation in corn

Abstract

Water scarcity and climate variability are intensifying the challenge of sustaining corn (Zea mays L.) yield while minimizing irrigation inputs. The study employed Unmanned Aerial System (UAS)-based thermal and multispectral remote sensing with advanced machine and deep learning methods to create a comprehensive, field-scale framework for precision irrigation and crop productivity forecasting. UAS-based thermal imaging was deployed to quantify the spatial variability of crop water stress under three irrigation treatments—33% (P33), 67% (P67), and 100% (P100) of full crop water requirement across six reproductive stages (R1—R6). High resolution canopy temperature maps were converted to the Crop Water Stress Index (CWSI), a standardized indicator of water stress revealing severe stress in the P33 treatment (CWSI > 0.80, canopy temperature > 45°C), moderate stress in the P67 treatment (CWSI 0.40-0.75), and optimal conditions in the P100 treatment (CWSI < 0.40). Temporal analysis confirmed peak stress during early reproductive stages, and non-parametric statistics (Friedman χ² = 6.33, p = 0.042) verified significant treatment effects, indicating the value of UAS-derived thermal data for site-specific irrigation scheduling. Temporal-Spatial Fusion Network (TSFN) was developed to predict corn yield by integrating thermal and multispectral imagery collected by unmanned aerial systems (UAS) across six maize reproductive stages (R1 to R6) with deep learning modules: convolutional neural networks (CNNs) to learn intra-plot spatial features, Graph Neural Networks (GNNs) to model inter-plot spatial dependencies, and gated Recurrent Units (GRUs) with attention mechanisms to capture sequential temporal dynamics. High resolution imagery was acquired weekly and used to derive vegetation indices (NDVI, NDRE, GCI, MTCI, EVI) and crop water stress Index (CWSI), enabling spatially explicit tracking of crop health and water stress responses. TSFN combines the model significantly outperformed traditional machine learning methods and standalone deep learning models. It achieved an R² of 0.79 and RMSE of 13.10 bu/acre, surpassing CNN-GNN and ensemble baselines. Later reproductive stages (R4-R6) contributed most to predictive accuracy, while early-stage data (R1-R3) supported timely in-season diagnostics. The model performed best under moderate and deficit irrigation (P67 and P33) due to increased spectral variability caused by crop stress. Unmanned aerial system (UAS) multispectral imagery was combined with machine-learning and deep learning models to generate high-resolution evapotranspiration (ET) estimates. NDVI-derived crop coefficient (Kc) and on-site ET-gage reference evapotranspiration (ETo) informed model training. Random Forest delivered near perfect accuracy across irrigation levels (R² = 1.00 at R1 to R3, RMSE = 0.00 mm day -1) and only small reduction at R4 (R² = 0.97). ResNet18 consistently outperformed the CNN and achieved strong agreement with field measurements (R² ≈ 0.95–1.00, RMSE ≈ 0.00-0.17 mm day -1). Stage-wise evaluation revealed ResNet18 remained highly accurate during most reproductive phase, but showed moderate dip at R4 (R² = 0.95). Ablation analysis confirmed that late season imagery was essential for model performance. Removing R5, which contains the highest ETc values (≈ 6-8 mm day-1), produced largest decline in predictive accuracy, reducing R² to 0.93 and raising RMSE to 0.28 mm day-1), whereas omitting R4 had minimal impact and in some cases slightly improved accuracy. These studies effectively delivered information that integrating UAS thermal and multispectral sensing with advanced analytics enables zone-level mapping of crop water stress, accurate yield prediction, and operational ET estimation. The resulting framework provides actionable, data-driven irrigation strategies that reduce water use while sustaining corn productivity under variable irrigation regimes.