Extracting cropland accurately and efficiently from high-resolution remote sensing images is of great significance to agricultural production and agricultural resource investigation. Cropland with different areas, ground covers and cultivation types in remote sensing images have large differences in features, whereas the insufficient generalization ability of traditional supervised learning models also leads to poor extraction of heterogeneous cropland with the above features. To solve this problem, the semi-supervised semantic segmentation with mutual knowledge distillation (SSS-MKD) model as the base model and incorporated an online hard example mining strategy based on a weighted loss function. The proposed model was evaluated on the Vaihingen dataset and achieved the highest overall accuracy of 87.1% and an average F1 score of 85.0%, The model had the best extraction accuracy compared with other semi-supervised models. In addition, for the task of large-area cropland extraction, the feature information of the unannotated images that were heterogeneous and homogeneous with the annotated images were added to the training of the semi-supervised learning model by designing two sets of experiments using Jilin-1 cropland image dataset, respectively, in order to improve the cropland extraction accuracy in the proposed extraction area. The experimental results showed that the proposed model could achieve the highest overall accuracy of 84.0% by using the cropland images to be extracted for assisted training. In addition, by using unlabeled images with strong similarity to the cropland in the target region, the overall accuracy could be further improved by 2.1~6.1 percentage points. The maximum overall accuracy achieved using unlabeled images with strong similarity to the cropland features in the training set was 81.6%, which was 6.6~8.5 percentage points higher than the accuracy achieved by using conventional supervised learning. Taking Xian County area in Hebei Province as an example, the model used the Jilin-1 cropland image dataset (part 1) as the labeled training set, and achieved the highest accuracy of 88.7% overall accuracy after training with both the Jilin-1 cropland image dataset (part 2) and the unlabeled images from the Xian County areas Gaofen-2 image dataset, compared with the extraction accuracy by using these two datasets alone and without the unlabeled images, respectively 1.0 percentage points, 4.9 percentage points, and 8.1 percentage points. Under the realistic background of abundant cropland remote sensing data and insufficient corresponding labeled information. The semi-supervised model learned the feature information in the images of cultivated land in the proposed extraction area and the training set area, and improved the extraction accuracy of heterogeneous cultivated land. A method that maximized the learning of heterogeneous cropland unlabeled images and a minimum amount of labeled data using a semi-supervised learning model for the task of wide-area cropland extraction was proposed, and the extraction results were effective.