Cross-batch calibration of sugarcane disease classification models based on visible and near-infrared spectroscopy using deep learning-based domain adaptation.

Visible-near infrared (Vis-NIR) spectroscopy provides rapid crop disease assessment; however, poor model generalizability remains a major limitation when models developed for a specific period are applied to batches collected at different times, primarily due to variations in physicochemical properties such as chlorophyll content, moisture level, surface texture, and tissue structure, which induce shifts in spectral distributions across batches. This study investigates deep domain adaptation to enhance cross-batch transferability for sugarcane disease classification. Two batches of healthy and symptomatic leaves were collected at different times using the same spectrometer. A customized one-dimensional convolutional neural network (1D-CNN) was trained on Batch 1 and adapted to Batch 2 using labelled samples through two strategies: retraining only the fully connected layers or fine-tuning all network parameters. Both strategies achieved 94% accuracy, with precision 0.92, sensitivity 0.97 and specificity 0.98, outperforming the non-adapted model and standard-free calibration transfer methods, namely Correlation Alignment and Transfer Component Analysis. These findings demonstrate that deep domain adaptation substantially improves the robustness and transferability of Vis-NIR classification models across heterogeneous sampling batches.