Scale-space-based patch selection for crop classification in satellite images

Crop classification is a critical task in agriculture, enabling the efficient management of soil and resources. Despite significant advancements using machine learning (ML) and deep learning (DL), existing approaches often fail to capture the intrinsic properties of crops necessary for achieving high classification accuracy. To address this limitation, this study introduces a scale analysis technique to determine the optimal patch size for image processing, enhancing the extraction of relevant features. By segmenting images into patches of appropriate size, a convolutional neural network (CNN) is employed to perform binary classification for each crop class. This tailored approach ensures that each patch focuses on distinguishing the target crop from others, improving classification performance and decreasing computational cost. The methodology is tested on the Campo Verde dataset, where it achieves superior results in terms of accuracy and sensitivity. The proposed model demonstrates an overall accuracy (A) of 96.46%, sensitivity (S) of 95.88%, and F1-score (F1) of 67.53%, establishing its effectiveness for crop classification tasks.