Arable Land Change Detection Using Landsat Data and Deep Learning

Mei Huang, Wenzhong Yang

Abstract: Arable land is closely related to people’s livelihood. Protecting arable land is very urgent. Thus, rapid and accurate detection of arable land changes is especially important for arable land protection. However, most existing deep learning-based methods can easily lead to the accumulation of errors, low accuracy, and have poor anti-noise ability. In this study, we proposed an improved U-Net model for arable land change detection. This is an end-to-end network that is briefer and more intuitive. The model was trained and tested on three arable land areas in Xinjiang. We trained Landsat 8 images of exuberant arable land areas with RGB and 15 m spatial resolution. The improved U-Net model has some advantages compared to other methods: the deeper U-Net has a larger field of perception, with greater noise immunity, and deep convolution can capture more complex spectral features, thus improving feature differentiation. Considering that the deeper the network, the easier the gradient disappears, we use residual units to prevent gradients from disappearing. Moreover, the model parameters were adjusted to reduce the complexity of the model. The experimental results show superior performance on change detection tasks compared to other traditional models with 96.00% accuracy, precision, recall, and FI score, 93.54%, 85.07%, 88.29%. Through experiments, we found that the network can detect the change of cultivated land well. Thus, the proposed model can effectively implement arable land change detection.