Fusion architectures for soft rot detection in melon plants using hyperspectral and multicolor fluorescence imaging

Abstract

The wide range of non-destructive sensors currently available in agriculture encourages the development of fusion models for better monitoring tools than individual devices. In this work, two deep-learning fusion architectures that combine hyperspectral and multicolor fluorescence imaging were designed and tested for the early detection of soft rot caused by Botrytis cinerea in melon (Cucumis melo) leaves, and compared them with four single-sensor architectures. Melon leaf sampling yielded 739 samples (431 infected, 308 healthy). Each sample combined an hyperspectral imaging spectrum (400–1000 nm, 300 bands) with four fluorescence images taken at 440, 520, 680, and 740 nm. Sampling was performed 1 and 2 days post-inoculation (dpi), when no visual symptoms were present, to train deep-learning models discriminating between infected and healthy plants. The six architectures were evaluated by training a large number of models, resulting from a combination of 48 hyperparameter fits and repeated 5-fold cross-validation, for a total of 4320 models. Statistical analyses showed that one fusion architecture significantly outperformed all others, achieving the highest mean cross-validated accuracy (0.8068), with individual runs reaching a maximum accuracy of 0.8919 for the significantly better-performing fusion architecture. These results confirm that the combination of reflectance spectra with four channels of fluorescence improves early soft rot detection in melon compared with single-sensor solutions.