Abstract:

Objective Farmland resources are an important foundation for social and economic development. In order to fully utilize and protect farmland and lay a solid foundation for the sustainable use of land, it is particularly important to obtain real-time and accurate information on farmland area, distribution, and other factors. The use of remote sensing technology to obtain farmland data can meet the requirements of large-scale coverage and timeliness, and it has been widely applied in the field of farmland information surveys. However, the current research and application of deep learning methods in remote sensing for cultivated land identification still have room for improvement in terms of depth. The accuracy of identification needs to be enhanced to ensure reliable results. The objective of this study is to explore the potential application of deep learning methods in the field of remote sensing for the hilly areas of Southwest China identification. Methods This study focused on utilizing high-resolution Gaofen-6 (GF-6) remote sensing images to identify cultivated land in Santai county, Mianyang city, Sichuan province, China. The study employed several state-of-the-art deep learning models, namely UNet++, DeeplabV3+, UNet, and PSPNet. These models were widely recognized for their effectiveness in image segmentation tasks. By utilizing these models, the researchers aimed to accurately identify and delineate cultivated land areas within the study area. The use of high-resolution GF-6 remote sensing images allowed for detailed and precise analysis of the land features. Deep learning models were well-suited for this task as they could learn complex patterns and relationships within the images, enabling accurate identification of cultivated land. By applying these deep learning models to the remote sensing images, the study aimed to provide insights into the potential of using such methods for cultivated land identification. Results and Discussions (1) The performance of the newly developed deep learning models showcased remarkable advancements in accuracy evaluation metrics, including the F₁ score, overall accuracy (OA), Kappa coefficient, and more, surpassing the capabilities of traditional machine learning approaches such as Radom Forest (RF) and the latest land cover products such as SinoLC-1 Landcover. With an astounding accuracy improvement of nearly 20% for machine learning approach and 50% for land cover products, these models proved to be a game-changer. (2) Among the various models evaluated, the UNet++ model stood out with its exceptional identification capabilities and achieved unparalleled results. The mean intersection over union (MIoU) metric, a measure of segmentation accuracy, reaches an impressive 81.93%. Furthermore, the OA metric demonstrates a remarkable 90.6% accuracy, while the Kappa coefficient, which assessed the agreement between predicted and actual labels, attains a high value of 0.800 6. These outstanding performance indicators highlighted the superiority of the UNet++ model in accurately identifying and segmenting cultivated land. (3) Representative local regions were selected to visually compare the results of cultivated land identification. It could be observed that the deep learning models generally exhibit consistent distribution patterns with the satellite imageries, accurately delineating the boundaries of cultivated land and demonstrating overall satisfactory performance. However, due to the complexity of identification features in remote sensing imageries, the deep learning models still encountered certain issues of omission and misclassification in cultivated land extraction. Among them, the UNet++ model showed the closest overall extraction results to ground truth, and it also exhibited advantages in terms of completeness of cultivated land extraction, discrimination between cultivated land and other land classes, and boundary extraction compared to other models. In contrast, the RF method performed poorly in identifying fragmented cultivated land areas, often misclassifying them as non-cultivated land. (4) UNet++ was an improved network architecture based on UNet model. UNet++ significantly enhanced the performance of the network by introducing more upsampling nodes and skip connections, as well as effectively integrating different depths of UNet. This enabled the network to better extract semantic information and fuse features from different levels. The UNet++ model has demonstrated outstanding performance in cultivated land extraction, producing more comprehensive results with clearer edges. Compared to the traditional UNet model, UNet++ successfully addressed the issue of information bottleneck, effectively enhancing the accuracy of segmentation. The accuracy of cultivated land extraction results was evaluated using metrics such as OA and Kappa coefficient etc. It was found that the four deep learning models outperformed the machine learning Random Forest (RF) algorithm. The improvement in OA and Kappa coefficient reached a maximum of 15.8% and 0.3486, respectively. Among the deep learning models, the UNet++ model performed the best, and DeeplabV3+, UNet, and PSPNet methods followed suit. Conclusions This study focuses on the practicality and reliability of automatic cultivated land extraction using four different deep learning models, based on high-resolution satellite imagery from the GF-6 in Santai county in China. This indicates that deep learning methods have certain advantages and practical value in rapidly and accurately obtaining land information from high-resolution remote sensing imagery. Based on the cultivated land extraction results in Santai county and the differences in network structures among the four deep learning models, it was found that the UNet++ model, based on UNet, can effectively improve the accuracy of cultivated land extraction by introducing the mechanism of skip connections.. Overall, this study demonstrates the effectiveness and practical value of deep learning methods in obtaining accurate farmland information from high-resolution remote sensing imagery, with the UNet++ model showing the highest performance among the models evaluated.

Key words: deep learning, remote sensing images, cultivated land identification, accuracy evaluation, hilly region of Sichuan Basin