Multilabel land cover aerial image classification using convolutional neural networks

Classifying the remote sensing images requires a deeper understanding of remote sensing imagery, machine learning classification algorithms, and a profound insight into satellite images’ know-how properties. In this paper, a convolutional neural network (CNN) is designed to classify the multispectral SAT-4 images into four classes: trees, grassland, barren land, and others. SAT-4 is an airborne dataset that captures the images in 4 bands (R, G, B, infrared). The proposed CNN classifier learns the image’s spectral and spatial properties from the ground truth samples provided. The contribution of this paper is three-fold. (1) A classification framework for feature extraction and normalization is built. (2) Nine different architectures of CNN models are built, and multiple experiments are conducted to classify the images. (3) A deeper understanding of the image structure and resolution is captured by varying different optimizers in CNN. The correlation between images of varying classes is identified. The experimental study shows that vegetation health is predicted most accurately by the proposed CNN models. It significantly differentiates the grassland vegetation from tree vegetation, which is better than other classical methods. The tabulated results show that a state-of-the-art analysis is done to learn varying land cover classification models.