Deep Learning-based Detection of Feature Changes in Arable Land

Abstract

In economically developed regions, the occupation of arable land has long been a significant concern, and identifying cultivated land accurately is crucial for mitigating this issue. To address the problem of "pseudo-change" caused by complex backgrounds, uneven brightness, and natural growth, we propose a change detection algorithm based on twin neural networks. Our algorithm comprises two sub-networks and a decision layer that use a multi-scale feature model and an improved spatio-temporal attention module to extract and process features from two temporal images. We validate our algorithm using high-resolution UAV aerial photography data with a resolution of 0.3m and 0.05m. Our proposed method achieves an F1 coefficient of over 87% in the cultivated land feature change detection dataset, which is more than 45% higher than the traditional STANet. The contribution of our research lies in developing a more effective method for identifying cultivated land in economically developed regions, which can help mitigate the serious "man-made disaster" of arable land occupation. Our proposed method has the potential to contribute significantly to the field of precision agriculture and land resource management.