A Quantitative Interpretability-Guided Network for Enhanced Wheat Seedling Detection

Abstract

Seedling count is a key indicator of wheat population during the seedling stage. Accurate seedling detection is therefore vital. Recently, deep learning techniques have been widely used to detect wheat seedlings and estimate seedling numbers from UAV images. However, existing models lack interpretability, resulting in unclear internal processes and making reliable optimization difficult. Consequently, directly applying deep learning for wheat seedling detection often does not yield optimal results. To facilitate this study, we constructed an RGB image dataset captured by the DJI ZENMUSE X4S during the wheat emergence stage. We introduced a model interpretation method that employs network dissection and Gradient-weighted Class Activation Mapping (Grad-CAM) to quantitatively assess the contribution of each network layer to wheat seedling detection. The results show that grayscale inputs elicited the highest responses in the early layers. Based on this, we propose a dual-input wheat seedling detection model that uses both the original image and its grayscale version, fusing their features in the early network layers. The combination of grayscale images improves wheat seedling detection performance. Experimental results demonstrate that our method outperformed the benchmark and traditional object detection techniques, achieving an AP50 of 90.2%, with a recall of 0.88 and a precision of 0.90. The proposed model not only improved AP50 but also reduced the missed detection rate, making wheat detection results more reliable. This approach enhances the robustness, generalization, and interpretability of deep learning models, offering an innovative and highly effective solution for increasing the accuracy and reliability of wheat seedling detection.

Abstract from the paper’s OpenAlex record; open copy licensed cc-by. Read at the source ↗

The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.