Aprendizado profundo (Deep learning) na identificação de estresse hídrico em meloeiro

Abstract

Remote monitoring through imaging is a powerful tool for evaluating the performance of melon crops, which exhibit high water demand and are primarily grown under irrigation. This study aimed to analyze the physiological and yield behavior of melon plants under different irrigation depths, as well as to build a database of spectrum visible and thermal images for water status prediction via deep learning. The experiment followed a randomized complete block design with four irrigation depths (50%, 75%, 100%, and 125% of crop evapotranspiration - ETc) and five replications. Drip irrigation was employed, and the Class A pan method was used to determine evapotranspiration. The length of the main branch showed a linear relationship with the increase in the applied irrigation depth. Physiological results at 65 days after transplanting (DAT) showed that transpiration and stomatal conductance exhibited quadratic behavior, with maximum values near 100% of ETc, while the photosynthetic rate and fruit mass increased linearly with the irrigation depth. Conversely, total soluble solids and pH decreased with increasing water depth. The thermal index showed a negative linear relationship with the applied depth. Regarding the computational aspect, InceptionV3, MobileNetV3, and ResNet50V2 networks were evaluated. It was found that visble spectrum images outperformed the thermal index by 9% in accuracy, and that segmentation reduced network performance by 6%. Although global metrics were similar across architectures, InceptionV3 was the most efficient, achieving the best results in only 19 minutes, compared to 104 minutes for ResNet50V2, as well, the worst coefficients of determination for the applied water depth in relation to the thermal index were observed in the morning period.