Digital mapping of coffee ripeness using UAV-based multispectral imagery

Timely and accurate monitoring of coffee ripeness is essential for harvest planning, especially in mountainous areas where the harvest is performed manually due to the limited use of agricultural mechanization. The increasing temporal and spatial resolutions of remote sensing based on low-altitude unmanned aerial vehicles (UAV) provides a feasible way to monitor the fruit ripeness variability. Due to these facts, this study was aimed to: (1) predict the fruit ripeness using spectral and textural variables; and (2) to determine the best variables for developing spatio-temporal variability maps of the fruit ripeness. To do so, an experiment with six arabica coffee fields was set up. During the coffee ripeness stage in the 2018-2019 and 2020-2021 seasons, seven flights were carried out using a quadcopter equipped with a five-band multispectral camera. After that, 12 spectral and 64 textural variables composed of bands and vegetation indices were obtained. For the same period, the percentage of unripe fruits (fruit ripeness) was determined using an irregular grid on each field. Then, the fruit ripeness was predicted with six machine learning (ML) algorithms using as input (1) the spectral variables and (2) the combination of spectral and textural variables. Among the evaluated ML algorithms, the random forest presented the higher accuracy, in which the model using the spectral and textural variables (r2 = 0.71 and RMSE = 11.47%) presented superior performance than the model based solely on spectral variables (r2 = 0.67 and RMSE = 12.09%). Finally, this study demonstrated the feasibility of using spectral and textural variables derived from UAV imagery for mapping and monitoring the spatiotemporal changes in the fruit ripeness at a fine scale.

Automated summary

This study aimed to predict coffee ripeness using spectral and textural variables and determine the best variables for developing spatio-temporal variability maps. They used UAV imagery to obtain the variables and applied six machine learning algorithms to predict fruit ripeness. The results showed that the model using spectral and textural variables had better performance.