HPAC: a forest tree species recognition network based on multi- scale spatial enhancement in remote sensing images

Abstract

ABSTRACT Forest tree species recognition is a pivotal subject in the field of remote sensing. To address this, deep learning has been extensively applied. Thus far, most classification methods have generally relied on learning certain global features, yet often overlook the characteristics of specific regions, thereby struggling to adequately handle the similarity between classes. Furthermore, due to the singular nature of features, effectively representing the attributes of tree species images becomes challenging, consequently impacting classification performance. To tackle these issues, a novel approach for forest tree species classification in remote sensing images is proposed, based on the Hollow Pyramid Attention Combination (HPAC) network. Initially, a Shallow Multi-scale Hollow Fusion (SMHF) module is introduced before the 7 × 7 convolution in the ResNet-50 network and the first residual block’s first layer. This module employs dilated convolutions to achieve varying receptive fields. Moreover, it incorporates positional feature information, significantly enhancing the shallow-level feature extraction capabilities, resulting in a richer feature representation. Subsequently, to minimize network parameters and computational workload while bolstering the capacity to recognize deep-level features, the last residual block of the ResNet-50 differentiation is substituted with a Maxpool Avgpool Fusion (MAF) module. This replacement serves to enhance classification accuracy. The classification process is ultimately concluded with a Softmax classifier. Experimental results underscore the effectiveness of the proposed method, achieving a classification accuracy of 95.89% on the PCANDVI dataset of forest tree species data (FTSD). In summary, the introduced HPAC network proves to be both feasible and effective.