Hyperspectral imaging and one–dimensional convolutional neural network (1D-CNN) based detection of strawberry powdery mildew under polytunnel conditions

Powdery mildew ( Podosphaera aphanis ) threatens strawberry yield and quality, demanding early, nondestructive detection in field-like conditions. We integrated hyperspectral imaging (350-1000 nm) on a rail-mounted system under a polytunnel with spectral preprocessing (Savitzky-Golay smoothing, SNV/MSC) and Isolation Forest outlier removal, and ROI-based spectrum extraction. A One-Dimensional Convolutional Neural Network (1D CNN) was trained on balanced spectra (n = 1,940; 70/15/15 split) and evaluated against RF, DT, PLS‑DA, and SVM. On the held-out test set, the full-spectrum 1D CNN achieved 84.2% accuracy (Healthy: precision/recall/ F 1 -score = 0.93/0.74/0.82; Mildew: 0.78/0.94/0.86) with AUC = 0.933, favoring high mildew sensitivity for screening. Results demonstrate the feasibility of HSI coupled with deep learning for reliable PM detection under semi‑controlled, near‑field conditions and highlight a practical path toward compact multispectral implementations and threshold‑calibrated operation in precision agriculture.