Application of a convolutional neural network to land use classification based on GF-2 remote sensing imagery

Traditional remote sensing-based land use classification methods mainly focus on pixel-based unsupervised and supervised approaches and object-based image analysis (OBIA). The primary objective of this study is to improve the land use classification accuracy by introducing a convolutional neural network (CNN). A 1-m resolution fused Chinese GaoFen-2 (GF-2) remote sensing image was used as the validation data, and the classification scheme included six land use types were identified, namely wetland, cultivated land, forestland, artificial surface, fallow land, and others. To augment the training samples, the rotation operation of 90° and 180° was carried out. A ratio of 3:1 of total samples was randomly assigned as the training set and validation set. Considering the spectral bands and spatial resolution of GF-2 imagery as well as classification scheme, a seven-layer CNN model was constructed including two convolutional layers with the kernel size of 3 × 3, two pooling layers with the kernel size of 2 × 2, and a fully connected layer. To avoid the vanishing gradient problem and accelerate model training, the ReLU activation function, adaptive moment estimation (Adam) algorithm, and cross-entropy objective function were introduced. The parameter optimization and ablation study were performed to obtain the optimal values including learning rate of 0.0001, weight decay of 0.4, batch size of 16, and iteration of 100. The novelty of our method is that four sizes of divided image block (n = 5, 7, 9, 11) were used as the input to optimally select the best classification effect. Support vector machine (SVM) classifier was comparatively used to validate the proposed method. The results showed that the best overall accuracy (OA) and Kappa coefficient (k) were respectively 94.68% and 0.9351 when n was set to 9. All the user’s and producer’s (UA and PA) accuracies were greater than 90% for each type. Conversely, for the SVM-based classification method, the OA and k were just 77.92% and 0.7328, and most of the UA and PA were less than 85%. It is obvious that our proposed CNN-based method has greatly improved the land use identification accuracy.