Research on the Remote Sensing Change Detection of Land Cover Types in Agricultural-Forestry Transition Zone Based on CNN-Transformer Hybrid Model

In the task of land cover change detection in complex agricultural, forestry, and mountainous areas (such as karst landform areas in mountainous agricultural and forestry area), Convolutional Neural Networks (CNNs) are limited by their local receptive fields, making it difficult to effectively integrate global contextual information. Conversely, Transformer models have deficiencies in expressing fine-grained local features. These limitations make it challenging for existing methods to simultaneously address issues such as highly fragmented land features, spectral-spatial heterogeneity, and complex topographic effects. This paper proposes a hybrid model, CT-CDNet, which deeply integrates the advantages of CNNs in fine-grained local feature extraction and Transformers in global contextual dependency modeling. To systematically evaluate the model's performance, this study established a sample subset containing the aforementioned typical complex cover characteristics of the mountainous agricultural and forestry area regions based on the CLCD/px-CLCD dataset (and specifically introduced DEM data as an auxiliary input) for core experimental verification. Simultaneously, its generalization ability was tested on the LEVIR-CD and WHU-CD public benchmark datasets. The experimental results show that CT-CDNet performs outstandingly on all key evaluation indicators. For example, on the core pxCLCD mountainous agricultural and forestry area characteristic sample subset, its F1 score and Intersection over Union (IoU) reached 86.2% and 75.7%, respectively, significantly outperforming various comparative benchmark models. It also demonstrated strong competitiveness on general datasets such as LEVIR-CD (F1 of 91.9%) and WHU-CD (F1 of 89.6%). These data fully demonstrate the advancement of CT-CDNet in improving the accuracy and robustness of land cover change detection in complex agricultural, forestry, and mountainous areas, especially for challenging spectral and topographical land cover.