Towards identification and explainable localization of slopes in autonomous excavation: a feature fused CAM approach

Autonomous earthmoving requires excavators to identify and localize slopes within complex environments while operating with limited computational resources. To address this challenge, we propose an explainable localization method that leverages the explainability of machine learning (ML) models for slope identification and localization, which also guides the excavator in optimal digging point determination. Our approach integrates a modified residual neural network with joint features derived from Class Activation Mapping (CAM), enhanced through transfer learning to fine-tune a pre-trained model for the target task. Evaluations on public SODA dataset demonstrate significant improvements in localization performance, with a 45.6% increase in the Intersection over Union (IoU) metric compared to the original CAM. Further performance gains are observed when preprocessing based on identification precedes localization, with IoU improving by over 70%. Furthermore, we constructed a few-shot slope dataset to validate the method’s efficacy under low-cost and resource-constrained conditions. The results indicate that our approach enables continuous explainable localization, effectively guiding an unmanned excavator in autonomous earthmoving. The proposed approach proves highly practical for engineering applications, addressing the challenges of large-scale datasets and high computational resource demands, thereby providing an effective technical pathway for applying ML methods to the automation of construction machinery.