The increasing market interest in coffee beverage, lead coffee growers around the world to adopt more efficient methods to select the best-quality coffee beans. Currently, coffee beans selection is carried out either manually, which is a costly and unreliable process, or using electronic sorting machines, which are often inefficient because some coffee beans defects, such as sour and immature beans, have similar spectral response patterns. In this sense, the present work aimed to assess coffee beans quality using both computer vision and machine learning techniques, such as Support Vector Machine (SVM), Deep Neural Network (DNN) and Random Forest (RF). For this purpose, an algorithm written in Python language was developed to extract shape and color features from coffee beans images. The obtained dataset was then used as input to the machine learning algorithms. The data reported in this study pointed to the importance of color descriptors for classifying coffee beans defects. Among the variables used, the components from RGB (Red, Green and Blue) and HSV (Hue, Saturation and Value) color spaces presented the most relevant contribution for the classification models. Also, the results reported in this study provides evidence that computer vision along with machine learning algorithms can be used to identify and classify coffee beans with a very high accuracy (> 90%). Key words: Deep neural network; classification; artificial intelligence; image processing; granulometry.

The characterization of soil attribute variability often requires dense sampling grids, which can be economically unfeasible. A possible solution is to perform targeted sampling based on previously collected data. The objective of this research was to develop a method for mapping soil attributes based on Management Zones (MZs) delineated from Sentinel-1 radar data. Sentinel-1 images were used to create time profiles of six indices based on VV (vertical–vertical) and VH (vertical–horizontal) backscatter in two agricultural fields. MZs were delineated by analyzing indices and VV/VH backscatter bands individually through two approaches: (1) fuzzy k-means clustering directly applied to the indices’ time series and (2) dimensionality reduction using deep-learning autoencoders followed by fuzzy k-means clustering. The best combination of index and MZ delineation approaches was compared with four soil attribute mapping methods: conventional (single composite sample), high-density uniform grid (one sample per hectare), rectangular cells (one composite sample per cell of 5 to 10 hectares), and random cells (one composite sample per cell of varying sizes). Leave-one-out cross-validation evaluated the performance of each sampling method. Results showed that combining the VV/VH index and autoencoders for MZ delineation provided more accurate soil attribute estimates, outperforming the conventional, random cells, and often the rectangular cell method. In conclusion, the proposed methodology presents scalability potential, as it does not require prior calibration and was validated on soil types commonly found across Brazil’s agricultural regions, making it suitable for integration into digital platforms for broader application in precision agriculture.

Characterization of soil attribute variability often requires dense sampling grids, which can be economically unfeasible. A possible solution is to perform targeted sampling based on previously collected data. The objective of this research was to develop a method for mapping soil attributes based on Management Zones (MZs) delineated from Sentinel-1 radar data. Sentinel-1 images were used to create time profiles of six indices based on VV (vertical-vertical) and VH (vertical-horizontal) backscatter in two agricultural fields. MZs were delineated by analyzing indices and VV/VH backscatter bands individually through two approaches: (1) fuzzy k-means clustering directly applied to the indices' time series, and (2) dimensionality reduction using deep-learning autoencoders followed by fuzzy k-means clustering. The best combination of index and MZs delineation approach was compared with four soil attribute mapping methods: conventional (single composite sample), high-density uniform grid (one sample per hectare), rectangular cells (one composite sample per cell of 5 to 10 hectares), and random cells (one composite sample per cell of varying sizes). Leave-one-out cross-validation evaluated the performance of each sampling method. Results showed that combining VV/VH index and autoencoders for MZs delineation provided more accurate soil attribute estimates, outperforming the conventional, random cells, and often the rectangular cell method.