Coffee mapping by remote sensing and machine learning

Abstract

Coffee cultivation is an important element in the Brazilian economy, producing export revenues, jobs and driving the local economy. Brazil is the largest producer and exporter of coffee in the world and its cultivation is an important element in our cultural identity. Knowing precise locations of coffee growing allows a better assessment of the balance between supply and demand for the product, makes it easier to monitor problems associated with cultivation, helping to take preventive measures or determine public policies to guide this activity. Mapping, done as automatically as possible, would help keep up-to-date data on growing regions across large tracts of land. On the other hand, automatic mapping of this particular crop faces several challenges related to the variety of coffee growth systems in different locations. Variations range from the use of different species and varieties of plants, such as visual differences related to the age of the plants, intercropping with other crops, and cultivation and management techniques. This work investigates the mapping of coffee crops, in the form of sun exposed monoculture, in the municipality of Lavras, MG, using images from the Sentinel-2 MSI satellite and the Random Forest classification algorithm. Random Forest is a machine learning algorithm and therefore “learns” to classify through examples, which requires a little manual classification in order to generate examples that intend to cover the various possible cases of classification. Producing adequate sampling to create a classification creates several practical problems, whose impact on classification still needs to be better studied. In this work, we observed that some practical aspects have much more significant effects than others. Classification tests were carried out, showing that the choice of classification examples ends up producing more significant effects than the choice of electromagnetic bands sampled by the satellite. The inclusion of noise (shadows, planting failures, road tracks) in the crop samples did not lead to a bad classification. A technique was also developed to eliminate common noise in pixel-based classifications, producing more continuous areas of classification, more suitable for geometric demarcation. The accuracy analysis focused on the classification on an area distinct from the training region, an uncommon feature in previous works, but which is important for the practical feasibility of the classification, since it is not feasible to produce a manual classification in a large region to be used. in order to train the classifier. Classification results were obtained with accuracy of up to 94.4%, with Kappa of 0.761, for classification in a region other than the training one. The classification system was all implemented with free software, using satellite data that are publicly available, using the R language and its libraries, including a Random Forest implementation of the ranger library.