A multi-angle comprehensive solution based on deep learning to extract cultivated land information from high-resolution remote sensing images

Abstract

Accurate acquisition of cultivated land area and location information is of great significance to agricultural management, agro-ecological environment monitoring, and national food security. The rapid development of deep learning technology provides a new way to extract cultivated land information. However, there are many parameters involved in deep learning, so it is time-consuming to find the optimal parameters. In order to simplify the complex parameter tuning process and explore the main parameters that affect the classification accuracy of deep learning, this study uses Gaofen-2 images as data support and based on U-Net semantic segmentation model to carry out an experiment of cultivated land extraction in Henan Province, China. This study designs model training in different scenarios based on four bands(near-infrared (Nir), red, green, and blue), three backbone models, and six patch sizes (pixel resolution), and analyzes the training model and classification results from multiple perspectives such as model training time, sample size, epoch, and classification categories and quantity. After analysis and evaluation, we conducted cultivated land extraction experiments in three years and different phenological periods, and finally tested the performance of U-Net, PSPNet, Deeplab, and Random Forest in cultivated land extraction. The results show that: (1) the patch size in the model has the greatest influence on the classification accuracy, which is affected by image resolution, geographic objects and batch size. When the patch size was smaller than 112 × 112 pixels, the model recall and F1 decreased substantially, with a decrease of 23.91% and 22.64%, respectively. (2) For U-Net model, after setting reasonable band combination, backbone model, patch size, epoch and sample category, the segmentation and classification accuracy of predicted results increased by 3.27% (Intersection over Union, IoU), 2.62%(overall accuracy, OA), 3.89%(F1) and 7.89%(Kappa) on average. (3) We also determined the optimal band combination (Nir, Red, and Green), patch size (224 × 224 pixels), and backbone (Resnet34) for cultivated land extraction from Gaofen-2 images. The appropriate patch size of cultivated land extraction in the study is between 224 × 224 and 256 × 256 pixels. (4) For cultivated land extraction, relying only on binary classification is not as effective as multi-classification.