Characterizing agri-forestry systems in Portugal through high-resolution orthophotos and convolutional neural networks

The Portuguese agri-forestry system Montado occupies about 730,000 hectares, which is about 8% of total area of Portugal. The maintenance of this biodiverse and complex land cover system is threatened, among other causes, due to frequent tillage to manage shrubs encroachment. In order to characterize Montado areas, we develop a neural network algorithm for identifying regions with trees, shrubs, covered and/or bare soil in grasslands. For this purpose, we used high-resolution RGB orthophotos (spatial resolution of 25 cm) that cover mainland Portugal. They were collected during the summer and autumn of 2018. The labelling of the used images was performed through an unsupervised method (Gaussian mixtures), which was validated through visual interpretation. The deep convolutional neural networks architecture used was U-net, which has been used in the literature to segment remote sensing images with a high performance. To train models, 800 orthophotos with 10,000 m2 each were used. They were divided between training and test set. A hyperparameter tuning was performed, namely the number of filters, dropout rate, batch size and the training/test partition percentage. In the best model, the overall classification performance (measured on the test set) was 89%, the recall 90% and the mean intersection of the union of 79%. Nevertheless, identification of shrubs had the lowest performance (accuracy of 85%), which are mainly confused with trees that have similar spectral signature. This model enables the identification of the status of Montado ecosystem regarding shrub encroachment for better future management.