Original papersValue of dimensionality reduction for crop differentiation with multi-temporal imagery and machine learning

- A diverse set of features dramatically increases crop classification accuracy.
- SVM is not greatly affected by high-feature dimensionality for crop differentiation.
- The SVM algorithm can be applied to the full set of generated features.
- Principal Components Analysis holds much potential for crop-type classification.

This study evaluates the use of automated and manual feature selection - prior to machine learning - for the differentiation of crops in a Mediterranean climate (Western Cape, South Africa). Five Landsat-8 images covering the different crop class phenological stages were acquired and used to generate a range of spectral and textural features within an object-based image analysis (OBIA) paradigm. The features were used as input to decision trees (DTs), k-nearest neighbour (k-NN), support vector machine (SVM), and random forest (RF) supervised classifiers. Testing was done by performing classifications (using all spatial variables) and then incrementally reducing the feature counts (based on importance allocated to features by filters), feature extraction, and manual (semantic) feature selection. Classification and regression trees (CART) and RF were used as methods to filter feature selection. Feature-extraction methods employed include principal components analysis (PCA) and Tasselled cap transformation (TCT). The classification results were analysed by comparing the overall accuracies and kappa coefficients of each scenario, while McNemar's test was used to assess the statistical significance of differences in accuracies among classifiers. Feature selection was found to improve the overall accuracies of the DT, k-NN, and RF classifications, but reduced the accuracy of SVM. The results showed that SVM with feature extraction (PCA) on individual image dates produced the most accurate classification (96.2%). Semantic groupings of features for classification also revealed that using the image bands and indices is not sufficient for crop classification, and that additional features are needed. The accuracy differences of the classifiers were, however, not statistically significant, which suggests that, although dimensionality reduction can improve crop differentiation when multi-temporal Landsat-8 imagery is used, it had a marginal effect on the results. For operational crop-type classification in the study area (and similar regions), we conclude that the SVM algorithm can be applied to the full set of features generated.