A point-supervised algorithm with multiscale semantic enhancement for counting multiple crop plants from aerial imagery

Abstract

Counting crop plants is important for agricultural activities such as crop breeding and yield prediction. Numerous studies have developed methods for counting individual crop plants or those with similar morphological characteristics. However, these methods often face challenges of low accuracy and poor generalization when counting multiple crop plants with significant scale variations in complex backgrounds. Hence, we proposed MCPCNet, a point-supervised algorithm that enhances multiscale semantics for counting multiple crop plants from aerial imagery. We also constructed the first dataset of multicategory crop plant counting (MCPC-Dataset). We developed a concurrent spatial group enhancement module, a residual dynamic dilated convolution module, and introduced the contextual transformer module with self-attention mechanism. These modules can reduce the impact of background, adapt to scale variations of multiple crops, and enhance the robustness of our algorithm, respectively. The experiment results on the MCPC-Dataset indicate that MCPCNet achieves excellent performance, with a mean absolute error (MAE) of 2.577, a mean square error (MSE) of 14.289, and a coefficient of determination ($R^2$) of 0.991. MCPCNet also has a clear advantage over the state-of-the-art (SOTA) point-supervised counting algorithm. In conclusion, MCPCNet provides a robust solution for high-precision counting of multiple crop plants and is a vital reference for future related research.