Deep 2D Convolutional Neural Network Architecture for Hyperspectral Land Cover Classification: A Comparative Study with KNN

Abstract

In recent years, deep learning techniques have received a great deal of attention in the context of hyperspectral
image (HSI) classification, particularly with regard to land cover mapping. Although 2D convolutional
neural networks (CNNs) are now widely used in this field, this study presents a refined, deeply structured
2D-CNN architecture that is specifically designed for spatial–spectral integration. Rather than introducing
a novel concept, the contribution lies in the balanced design of the architecture, which integrates dropout
and batch normalisation to enhance accuracy and generalisability on benchmark datasets. The proposed
network includes 10 convolutional layers organized into three blocks, each followed by max-pooling, batch
normalization, and dropout layers to reduce overfitting and improve model robustness. A fully connected
classifier with Softmax activation performs the final prediction. We trained the architecture using the Salinas
Valley dataset, which contains 54,129 labeled pixels across 16 land cover classes. The data were
meticulously segmented into two distinct components: the initial segment encompassed the primary data
set, while the subsequent segment comprised the ensuing data. It is noteworthy that 70% of the data was
allocated for training purposes. The remaining 30% of the budget was allocated for testing purposes. The
training was executed for 100 epochs by employing the Adam optimizer and categorical cross-entropy loss
function. The 2D-CNN model demonstrated superior performance in terms of classification accuracy when
compared with the KNN approach. The 2D-CNN model attained a classification accuracy of 94%, while
the KNN method achieved 88%. The findings indicate the efficacy of deep 2D-CNNs (Convolutional Neural
Networks) in the classification of hyperspectral land cover. The results also demonstrate the networks’
suitability for implementation in large-scale remote sensing projects.