Farmland Segmentation in Landsat 8 Satellite Images Using Deep Learning and Conditional Generative Adversarial Networks

Leveraging mid-resolution satellite images such as Landsat 8 for accurate farmland segmentation and land change monitoring is crucial for agricultural management, yet is hindered by the scarcity of labelled data for the training of supervised deep learning pipelines. The particular focus of this study is on addressing the scarcity of labelled images. This paper introduces several contributions, including a systematic satellite image data augmentation approach that aims to maintain data population consistency during model training, thus mitigating performance degradation. To alleviate the labour-intensive task of pixel-wise image labelling, we present a novel application of a modified conditional generative adversarial network (CGAN) to generate artificial satellite images and corresponding farm labels. Additionally, we scrutinize the role of spectral bands in satellite image segmentation and compare two prominent semantic segmentation models, U-Net and DeepLabV3+, with diverse backbone structures. Our empirical findings demonstrate that augmenting the dataset with up to 22.85% artificial samples significantly enhances the model performance. Notably, the U-Net model, employing standard convolution, outperforms the DeepLabV3+ models with atrous convolution, achieving a segmentation accuracy of 86.92% on the test data.