Semantic segmentation of very-high spatial resolution satellite images: A comparative analysis of 3D-CNN and traditional machine learning algorithms for automatic vineyard detection

Abstract

Erzincan (Cimin) grape, which is an endemic product, makes a significant contribution to the economy of both the region where it is grown and the country's economy. For this reason, the follow-up and dissemination of this product have become increasingly important. The study aims to determine the spatial distribution of vineyards by classifying very-high spatial resolution satellite images using machine learning and deep learning algorithms. A 3D Convolutional Neural Network (CNN) based deep learning model was created for classification vineyards. Then, this model has compared machine learning algorithms, namely support vector machine (SVM), Random Forest (RF), and Rotation Forest (RTF). Post-classification accuracy analyses were performed using error matrices, kappa analysis and McNemar tests, and the best overall classification accuracies and kappa values were obtained for 3D CNN and RF as 86.47% (0.8308) and 70.53% (0.6279), respectively. It was seen that the accuracy of RF increased to 75.94% (0.6364) when Gabor texture features were used, yet the highest classification accuracy was obtained with the 3D CNN classifier with an 11% higher score (86.47%). The χ^2 values calculated with the McNemar's statistical hypothesis test for all classification results are higher than 3.84 at the 95% confidence interval, indicating that the 3D CNN classifier has improved the classification accuracy, significantly.