Farmland non-agriculturalization is a serious threat to global food security and ecological stability. Remote sensing change detection technology has become a core tool for identifying the process of farmland non-agriculturalization by virtue of its advantage of large-scale dynamic monitoring. However, existing methods face challenges in balancing the extraction of fine edge details in fragmented farmland with the maintenance of global semantic consistency in large-scale fields, often resulting in blurred edges and lost local features. To address these issues, a farmland change detection algorithm based on improved BIT, named Far-CDNet, is proposed. Firstly, a detail enhancement convolution module that connects ordinary convolution and multiple differential convolutions in parallel is introduced, and the edge detail representation capability of the feature extraction network is enhanced through dynamic weighting and residual connection. Secondly, the ordinary convolution of the semantic tokenizer in the BIT module is replaced by a deep separable convolution to enhance the local feature capture ability and generate output features with higher-level semantics, so as to improve the overall feature expression ability of the model. Finally, a residual branch is added to further integrate the local and global information before and after the Transformer. The experimental results show that the improved model F1 score is 79.18%, and IoU is 69.32%. Compared with the BIT model, the F1 score is increased by 4.17%, and IoU is increased by 4.24%.

In order to effectively reduce and prevent the damage to the natural environment caused by illegal land reclamation and mineral excavation, an ETS-YOLO small target monitoring and recognition algorithm for the detection of various types of engineering vehicles in complex environments is proposed using cameras deployed to high towers. Firstly, the EfficientViT network is used to replace the backbone feature extraction network of YOLOv5s in order to improve the attention diversity and significantly reduce the number of model parameters. Secondly, a small target detection layer is added to enhance the network’s extraction of shallow semantic information to improve the performance of small target detection. Finally, the original NMS function is replaced with the soft non-maximal suppression algorithm (soft-NMS) to effectively recognize occluded and overlapped targets. The experimental results show that the improved model has a mean average precision (mAP) of 93.3%, a parameter count of 5.90 M, and a detection speed of 52 f/s. Compared with the YOLOv5s model, the mAP is improved by 2.6% and the parameter count is decreased by 16.1%.