High-resolution mapping of paddy rice fields from unmanned airborne vehicle images using enhanced-TransUnet

Abstract

Through remote sensing to obtain accurate information on the area of rice fields is of great significance for precision agriculture. Currently, rice extraction is primarily based on multi-temporal but low spatial resolution remote sensing images, which are unsuitable for a wide range of applications in efficient agricultural management and production. Exploring new methods for acquiring very-high processing resolution (VHR) images from Unmanned Aerial Vehicles (UAV) is a viable research avenue. Given that emerging deep learning networks have shown potential in image processing and object detection, this research proposed a deep learning network named Enhanced-TransUnet (ETUnet) for identifying paddy rice fields from VHR images. The developed network utilizes a dilated convolution approach and introduces the Convolutional Block Attention Module (CBAM) to the feature extraction layer in the convolutional neural networks to reduce unnecessary feature extractions by combining the self-attention mechanism in the Transformer. We applied the developed deep-learning network to extract rice fields from UAV images at three different growth stages, including transplanting, tilling, and maturing in Guangzhou city in China. The results demonstrate that ETUnet can accurately extract paddy fields during the phases of transplanting, tillering, and maturing, where the attained F1 scores are 94.87 %, 95.05 %, and 92.95 %, respectively. The attained IoUs are 90.24 %, 90.55 %, and 87.84 %, respectively, and the Kappa coefficients obtained are 93.13 %, 93.07 %, and 90.15 %, respectively. We identified that training samples had a substantial impact on the performance of the deep neural networks. The study revealed that both the timing of image acquisition and the model architecture affected paddy rice mapping using deep learning networks based on UAV data. It provides reference and help for studying the changes of crop phenology.