A novel deep learning model for extracting arable land from high-resolution remote sensing images in hilly areas: a case study in the Sichuan Basin of Southwest China

Arable land is the fundamental guarantee of agricultural production, and accessing accurate arable land information is particularly crucial. A novel deep learning model named CNX-eMLP with ConvNeXt as the backbone and an enhanced Multilayer Perceptron (eMLP) as the decoder was proposed for arable land extraction. The model was employed to extract arable land using high-resolution satellite imagery in a case study at Pengxi County of Southwest China and compared its performance with six deep learning models, a machine learning-based algorithm, and SinoLC-1. The study results show the CNX-eMLP significantly achieved the highest accuracy, with MIoU and OA of 75.21 and 87.9, highlighting a trade-off between computational complexity and accuracy. The CNX-eMLP model reveals arable land is predominantly found in low-altitude areas (below 400 m), with most plots being 0-5 hectares. The study presents an efficient and feasible method for accurate high-resolution remote sensing monitoring of arable land parcels in hilly regions. The eMLP decoder effectively boosts semantic segmentation performance via its innovative bidirectional fusion architecture. This design incorporates a bottom-up path augmentation within the Feature Pyramid Network, enhancing the propagation of low-level details. Original MLP layers are replaced by high-performance eMLP-Blocks that introduce batch normalization and GELU activation, enabling smoother nonlinearity. CARAFE is used for content-aware feature reassembly instead of conventional bilinear interpolation, thereby markedly enhancing both the efficiency and accuracy of semantic segmentation tasks.The CNX-eMLP model significantly outperforms other models and SinoLC-1 in terms of precision and efficiency. The use of deep learning techniques, particularly the innovative combination of ConvNeXt as a backbone and eMLP as a decoder, leads to a noteworthy improvement in accuracy metrics for arable land extraction.By integrating spectrum-texture features, the CNX-eMLP model enhances the accuracy of arable land extraction, demonstrating the substantial potential of combining multiple features to enhance the results of semantic segmentation tasks in complex environments.