Hyperspectral sensors record surface reflectance in hundreds of narrow, contiguous spectral bands and therefore allow crops that appear almost identical in broadband imagery to be discriminated. This spectral richness, however, introduces the Hughes phenomenon: as the number of bands grows relative to the number of labelled training samples, the accuracy of conventional classifiers first improves and then deteriorates, while the strong correlation between neighbouring bands adds redundancy and computational cost. Dimensionality reduction, either by feature extraction such as principal component analysis (PCA) or by band selection, is therefore an essential pre-processing step. Classical learners such as artificial neural networks (ANN) and support vector machines (SVM) remain standard baselines and recent deep architectures exceed 99% accuracy on benchmark scenes, although they require large training sets, heavy computation and complex tuning and offer limited interpretability. Two issues have received comparatively little attention in the benchmark literature: the strong sensitivity of shallow-network accuracy to training hyperparameters in high-dimensional spaces, and the use of unsupervised clustering to reveal the intrinsic spectral-affinity structure among crops without any labelled data.


The standard Salinas Valley hyperspectral image acquired by the AVIRIS sensor was used. The scene covers 512 by 217 pixels at a spatial resolution of 3.7 m and, after the removal of the water-absorption bands 108-112, 154-167 and 224, retains 204 effective spectral bands spanning 400-2500 nm; it contains 54,129 labelled pixels distributed very unevenly among sixteen crop classes. Band collinearity was quantified before any reduction: the mean absolute correlation over all band pairs was 0.65 and that between adjacent bands 0.99, with 33% of all pairs exceeding 0.90, and the participation ratio of the covariance eigenvalues was 1.6, so the effective dimensionality is close to two. Three input scenarios were then defined: all 204 bands, all principal components, and the first two components, which are the only ones above the 2% variance threshold and together account for 98.3% of the total variance. The labelled pixels were split 70/30 by stratified random sampling, and a spatially disjoint split along a column cut served as a complementary test of spatial information leakage. An ANN and an RBF support vector machine were trained on identical training pixels and scored on identical test pixels in every scenario, using overall accuracy, the kappa coefficient, producer's and user's accuracy and the F1 score. The ANN sensitivity analysis covered the number of hidden nodes (2 to 50), the number of layers (one to three), the early-stopping patience (1 to 12) and five training algorithms, searched one factor at a time with ten repetitions per configuration. In the unsupervised stage, K-means and self-organizing maps were applied for cluster numbers from 2 to 16 and assessed with the silhouette, Davies-Bouldin and Calinski-Harabasz indices, while the spectral angle between class centroids provided an independent measure of spectral affinity.


Training hyperparameters proved to be as decisive as the choice of classifier. Increasing the early-stopping patience raised the ANN test accuracy on PCA features from about 66% to 90.3%, a change exceeding 24 percentage points; across the three scenarios the range of accuracy over the patience sweep was 20.1, 24.4 and 23.0 percentage points, with Spearman correlations of 0.62 (p = 0.031) for all bands and 0.64 (p = 0.026) for two components, so the effect is large but not monotonic. In the high-dimensional space the scaled conjugate gradient and one-step secant algorithms performed best, whereas the Levenberg-Marquardt algorithm failed completely, which is attributable to its heavy memory demand when the number of parameters is large. In the like-for-like comparison the support vector machine ran without difficulty on the full 204-band space, contrary to the initial analysis, and reached 93.9% overall accuracy with a kappa of 0.932, against 91.1% and 0.901 for the ANN; on all principal components the two reached 92.7% and 90.0%, and on the first two components only 83.4% and 81.8%. The earlier finding that the support vector machine was ineffective therefore reflected the restriction of its input to two components and untuned kernel parameters rather than any property of the method. Dimensionality reduction brought no gain in accuracy on this scene: retaining all components, an invertible rotation of the original space, left accuracy almost unchanged, while discarding components on a variance criterion cost about ten percentage points. Under the spatially disjoint split the mean F1 fell from 97.2% to 83.4%, showing that spatial autocorrelation does inflate the conventional random-split estimate. The clustering analysis showed that the intrinsic spectral structure is coarser than the nominal sixteen classes: the silhouette peaked and the Davies-Bouldin index reached its minimum at the same seven clusters. The stubble class formed the purest cluster with more than 99% of its pixels in a single cluster, while the spectral angle confirmed the overlaps independently, the closest pairs being fallow and developing vineyard soil at 1.1 degrees and untrained grapes and untrained vineyard at 1.4 degrees.


In the supervised recognition of crop patterns from hyperspectral imagery with classical methods, the tuning of the training parameters is at least as important as the choice of the learning method: the early-stopping rule alone moved accuracy by more than 24 percentage points, whereas the gap between the two classifiers never exceeded three. Dimensionality reduction on a variance criterion did not improve accuracy on this scene; the value of principal component analysis here is computational, and the number of retained components should be tied to classification performance rather than to a fixed variance threshold. The intrinsic spectral structure of the scene is described by seven clusters rather than sixteen nominal classes, and the classes prone to confusion can be identified before any classification, without labelled data, from the spectral angle between class centroids. These results give three practical recommendations for crop-recognition systems: budget and document the parameter tuning before selecting a classifier,, choose the number of components by classification performance, and concentrate ground sampling on the class pairs that the spectral angle identifies as close.