Deep Learning-Assisted Thermography-Derived Crop Water Stress Index for Drought Stress Detection in Potato Plants

This study investigated the feasibility of using thermography-derived crop water stress index (CWSI) images combined with deep learning for irrigation-level identification in potato plants as a proof of concept. Potato plants (Solanum tuberosum L. cv. Sumi) were cultivated under greenhouse conditions and subjected to six irrigation regimes (i20, i33, i46, i60, i80, and i100) based on soil field capacity. Thermal images were acquired over four imaging weeks from late April to early May 2024. CWSI images were generated from the microclimate conditions and subsequently used for image-based classification. A lightweight convolutional neural network (CNN) consisting of four convolutional layers was developed, and a convolutional block attention module (CBAM) was further incorporated as an attention-based comparison model. The classification performance varied across imaging weeks and model architectures. Across repeated runs, the baseline CNN achieved better overall performance than CNN-CBAM, with ACC, PRE, REC, and F1S values ranging from 0.569 to 0.781, 0.534 to 0.828, 0.570 to 0.781, and 0.523 to 0.770, respectively. These results indicate that the incorporation of CBAM did not provide a consistent performance advantage under the current dataset and training conditions. Feature visualisation using t-distributed stochastic neighbour embedding (t-SNE) further showed the distribution of CNN- and CNN-CBAM-extracted features across irrigation levels, while local interpretable model-agnostic explanations (LIME) indicated that model-relevant regions for lower irrigation treatments were frequently located around the canopy margins. Overall, this study provides a proof-of-concept demonstration that thermography-derived CWSI images combined with a lightweight CNN can support rapid, non-contact irrigation-level identification in potato plants, while also highlighting current limitations and future directions for more robust field application. Future studies should validate the method using larger datasets collected across different cultivars, seasons, greenhouse and field environments, and independent plant- or plot-level biological replicates.