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

The Erzincan (Cimin) grape, which is an endemic product, plays a significant role in the economy of both the region it is cultivated in and the overall country. Therefore, it is crucial to closely monitor and promote this product. The objective of this study was to analyze the spatial distribution of vineyards by utilizing advanced machine learning and deep learning algorithms to classify high-resolution satellite images. A deep learning model based on a 3D Convolutional Neural Network (CNN) was developed for vineyard classification. The proposed model was compared with traditional machine learning algorithms, specifically Support Vector Machine (SVM), Random Forest (RF), and Rotation Forest (ROTF). The accuracy of the classifications was assessed through error matrices, kappa analysis, and McNemar tests. The best overall classification accuracies and kappa values were achieved by the 3D CNN and RF methods, with scores of 86.47% (0.8308) and 70.53% (0.6279) respectively. Notably, when Gabor texture features were incorporated, the accuracy of the RF method increased to 75.94% (0.6364). Nevertheless, the 3D CNN classifier outperformed all others, yielding the highest classification accuracy with an 11% advantage (86.47%). The statistical analysis using McNemar's test confirmed that the χ2 values for all classification outcomes exceeded 3.84 at the 95% confidence interval, indicating a significant enhancement in classification accuracy provided by the 3D CNN classifier. Additionally, the 3D CNN method demonstrated successful classification performance, as evidenced by the minimum-maximum F1-score (0.79-0.97), specificity (0.95-0.99), and accuracy (0.91-0.99) values.