Boosting multiclass crop interpretation from vehicle-view images by fusing visual, geometry, and texture information using a multimodal progressive fusion network

Accurate and refined crop mapping is important for precision agriculture and global food security monitoring. Proximal sensing data, such as vehicle-view images, provide important support for this task. However, crop interpretation from such images remains difficult due to complex backgrounds, variable viewing angles, and high visual similarity among crop types. Most existing methods rely solely on RGB (Red, Green, and Blue) visual information and therefore have limited accuracy under these challenging conditions. To address these issues, this paper proposes a multimodal progressive fusion network, termed TriFusionNet. We integrate three complementary sources of information: visual information from RGB images, geometric information from AI-generated depth images to handle complex structures and varying viewing angles, and textural information from wavelet-extracted features to distinguish visually similar crops. The proposed model adopts an HRNet (High-Resolution Network) backbone to preserve spatial details, progressively fuses multimodal features at multiple scales, and introduces a Transformer module to capture global context. Experiments on a self-collected dataset show that the proposed method significantly outperforms unimodal baselines and several mainstream segmentation methods, achieving a mIoU (mean Intersection over Union) of 66.62%. The results provide valuable ground-level crop recognition support for remote sensing validation.