A Low-Cost IoT Imaging and Machine-Learning Framework for Morphology-Based Estimation of Crop Coefficient and Crop Evapotranspiration
IEEE Internet of Things Journal, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/jiot.2026.3735753
- Dergi Adı: IEEE Internet of Things Journal
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, ABI/INFORM, Aerospace Database, Compendex, INSPEC, Technology Collection (ProQuest)
- Anahtar Kelimeler: Crop coefficient, crop evapotranspiration, image processing, Internet of Things (IoT), plant morphology, precision agriculture, water-stress detection
- Atatürk Üniversitesi Adresli: Evet
Özet
Determining how much water a crop actually requires remains a practical bottleneck in irrigation management, since prevailing methods depend either upon dense meteorological and soil instrumentation that is costly to install and maintain, or upon aerial and satellite imagery whose resolution cannot resolve the individual plant. This study addresses that gap by estimating the crop coefficient (Kc) and crop evapotranspiration (ETc) directly from the morphology of individual plants observed by low-cost Internet of Things (IoT) cameras, without any meteorological or soil sensor. Distributed ESP32 and ESP32-CAM nodes imaged common bean grown under four controlled deficit-irrigation levels throughout one season and transmitted the images to a cloud store. A You Only Look Once version 12 (YOLOv12) pipeline combining instance segmentation with pose estimation extracted leaf, bean, and whole-plant areas together with the leaf inclination angle; this joint use of segmentation and pose estimation for the extraction of plant traits has not previously been reported. Nine regression models were trained upon these features under a day-based grouped cross-validation protocol in which all images acquired on a given day were confined to a single fold. Tree-based ensembles proved the most accurate, with extremely randomised trees attaining R2 = 0.984 for Kc and R2 = 0.935 for ETc and demonstrating a statistically significant advantage over every competing model, whereas models restricted to meteorological variables reached only 0.482 and 0.400, respectively, under the same protocol. A leaf inclination threshold of 170°, justified here by a sensitivity analysis rather than merely asserted, separated stressed from unstressed plants without error. Since the reference values are derived from a Penman–Monteith-based decision-support system rather than from lysimetry, and the evidence covers a single crop, site, and season, the results establish sensorless visual monitoring as a practical and scalable basis for estimating crop water requirements.