Multiclass Crop Interpretation via a Lightweight Attentive Feature Fusion Network Using Vehicle-View Images

Automatic crop interpretation can provide important reference information for national agricultural decision-making. However, due to the diverse characteristics and complex spatial relationship of crops, remote sensing images taken from a bird's eye view are insufficient in vertical features of crops, making it difficult to interpret crop types and locations accurately. The similar features and blurred edges between different crops make it difficult to extract crop boundaries accurately. Due to the high memory and computational costs, most of the deep learning-based models face efficiency limitations in real-scenario crop interpretation. To address the abovementioned issues, this article proposes a novel lightweight neural network, namely the CropNet, for crop interpretation. Aiming at the problem of feature similarity among different categories of crops, this article designs a global-local path aggregation (GLPA) mechanism, which uses shallow and deep neural networks to extract global detail information and local high-level information to enhance feature discrimination. An edge context feature enhancement module (ECFEM) is proposed to enhance edge and context feature extraction to address the problem of ambiguous crop edges. Finally, a feature fusion module based on an attention mechanism is used to automatically weigh different feature channels to enhance the crop semantics. To demonstrate the effectiveness of the CropNet proposed in this article, we constructed several sets of comparison experiments comparing it with state-of-the-art deep learning models on a manually labeled vehicle-view crop image dataset. The experimental results show that CropNet has better semantic segmentation results with fewer model parameters and lower computational costs.