Spectral–Spatial Feature Learning in Multispectral Plant Detection: A Comparison of 2D+ and 3D CNN Architectures

Abstract

Plant detection in multispectral images is an important process requiring effective analysis of spectral signatures and spatial patterns. In this study, the detection of horsetail (Equisetum telmateia), a plant species valuable for both medicinal and agricultural purposes, is addressed using multispectral aerial imagery. A detection system focused on achieving a high recall rate is proposed, and the images are subjected to binary classification through local patches. For performance analysis, a 3D convolutional neural network (3D CNN) that processes spectral and spatial dimensions simultaneously is compared with a 2D+ hybrid CNN architecture that handles these dimensions sequentially. Experiments conducted on approximately 300 multispectral images obtained from the Black Sea region show that the 3D model achieves higher recall in plant detection, but its precision is lower than that of the 2D+ model. In addition, it was observed that the 2D+ model has a shorter inference time.