Evaluation of a newly developed ploughshare: An ensemble deep learning approach for soil surface roughness classification with explainable artificial intelligence

The roughness of the soil surface directly affects seedbed preparation, sowing operations and crop yield. The roughness of the soil surface depends on the equipment used in the tillage process. Evaluating the performance of this equipment is important for increasing agricultural productivity. Recent advances in artificial intelligence (AI) and image processing technologies use convolutional neural network (CNN) models that can automatically learn meaningful patterns from images, enabling the objective analysis of soil surface roughness. Rather than analyzing a single CNN model, ensemble deep learning (EDL) approaches combining predictions from multiple CNN models improve classification accuracy. The aim of this study is to classify soil surface images obtained from primary tillage operations performed with a newly developed mouldboard ploughshare (NDP) and a conventional mouldboard ploughshare (CP) using the EDL method. Images were taken from the field to evaluate the soil surface structure using NDP and CP after primary tillage. These images were then enhanced using the gradient-weighted class activation mapping (Grad-CAM) method to generate heat maps, and the generated images were classified using known CNN and EDL models. The hard voting technique was used for the EDL approach. The Voting 6 model achieved 98.3 % accuracy. In addition, according to the results of the profilometer measurement, it was determined that the soil cutting and fragmentation performance of NDP was higher than that of CP. As a result, this image-based method can significantly contribute to the testing process of primary tillage tools.