Dynamic monitoring of cropland using high spatial resolution remote sensing images is a powerful means to protect cropland resources. However, when a change detection method based on a convolutional neural network employs a large number of convolution and pooling operations to mine the deep features of cropland, the accumulation of irrelevant features and the loss of key features will lead to poor detection results. To effectively solve this problem, a novel cropland change detection network (CroplandCDNet) is proposed in this paper; this network combines an adaptive receptive field and multiscale feature transmission fusion to achieve accurate detection of cropland change information. CroplandCDNet first effectively extracts the multiscale features of cropland from bitemporal remote sensing images through the feature extraction module and subsequently embeds the receptive field adaptive SK attention (SKA) module to emphasize cropland change. Moreover, the SKA module effectively uses spatial context information for the dynamic adjustment of the convolution kernel size of cropland features at different scales. Finally, multiscale features and difference features are transmitted and fused layer by layer to obtain the content of cropland change. In the experiments, the proposed method is compared with six advanced change detection methods using the cropland change detection dataset (CLCD). The experimental results show that CroplandCDNet achieves the best F1 and OA at 76.04% and 94.47%, respectively. Its precision and recall are second best of all models at 76.46% and 75.63%, respectively. Moreover, a generalization experiment was carried out using the Jilin-1 dataset, which effectively verified the reliability of CroplandCDNet in cropland change detection.

To address the challenge of low automation accuracy in cropland extraction caused by complex mountainous terrain, severe cropland fragmentation, and ambiguous boundaries, this study proposes a novel semantic segmentation model for cropland in high-resolution remote sensing imagery, termed DEDANet, which integrates detail enhancement and distance attenuation mechanisms. Into the encoding stage, introduced is a Detail Enhancement Convolution (DEConv) module that amplifies high-frequency edge information; through a multi-branch feature extraction pathway fusing five distinct convolutional types, the model significantly enhances its sensitivity and representational capacity for irregular cropland boundaries. Particularly in the decoder stage, embedded is a Distance-Attenuated Transformer module, wherein an attenuation matrix assigns differentiated attention weights to pixels across spatial locations-thereby suppressing irrelevant background interference and reinforcing the contextual coherence of adjacent regions. By jointly optimizing and integrating global dependencies with local fine-grained details, not only are long-range contextual relationships captured, but also the precision of localized boundaries is substantially refined, leading to notable improvements in cropland boundary delineation and internal connectivity under complex terrains. To evaluate the model's performance, a mountainous cropland dataset encompassing diverse cropland types was constructed. Experimental results demonstrate that DEDANet surpasses mainstream models across key metrics, achieving an overall accuracy of 95.76%, an F1-score of 95.15%, and a mean Intersection over Union (mIoU) of 90.79%. Furthermore, ablation studies and cross-regional validations on the Hi-CNA and iFLYTEK datasets verify the model's robust generalization ability, thereby offering an efficient and scalable technical solution for refined cropland management in mountainous regions.

Deep learning techniques have become the mainstream approach for fine-grained crop classification in unmanned aerial vehicle (UAV) remote sensing imagery. However, a significant challenge lies in the long-tailed distribution of crop samples. This imbalance causes neural networks to focus disproportionately on majority class features during training, leading to biased decision boundaries and weakening model performance. We designed a minority sample enhanced sampling (MES) method with the goal of addressing the performance limitations that are caused by class imbalance in many crop classification models. The main principle of MES is to relate the re-sampling probability of each class to the sample pixel frequency, thereby achieving intensive re-sampling of minority classes and balancing the training sample distribution. Meanwhile, during re-sampling, data augmentation is performed on the sampled images to improve the generalization. MES is simple to implement, is highly adaptable, and can serve as a general-purpose sampler for semantic segmentation tasks, functioning as a plug-and-play component within network models. To validate the applicability of MES, experiments were conducted on four classic semantic segmentation networks. The results showed that MES achieved mIoU improvements of +1.54%, +4.14%, +2.44%, and +7.08% on the Dali dataset and +2.36%, +0.86%, +4.26%, and +2.75% on the Barley Remote Sensing Dataset compared with the respective benchmark models. Additionally, our hyperparameter sensitivity analysis confirmed the stability and reliability of the method. MES mitigates the impact of class imbalance on network performance, which facilitates the practical application of deep learning in fine-grained crop classification.