Assessment of CNN-based methods for discrimination of olive planting systems with Sentinel-2 images

• Deep Learning algorithms were trained to discriminate olive planting systems with satellite images. • The use of Sentinel-2 imagery posed a challenge due to their spatial resolution. • Three convolutional neural networks (CNN) approaches were proposed to address the input size variability. • The best-fit approach yielded an accuracy of over 80%. • This methodology makes it possible to optimize the frequency of management of olive plantation systems. Traditional low-density old olive groves are being replaced by high-density or super-high-density young orchards, entailing important changes in crop management to increase productivity. However, this transformation has significant agronomic, economic, environmental, and sociocultural consequences. This study aims to evaluate deep learning techniques based on convolutional neural networks (CNN) to automatically identify different olive grove intensification patterns using medium-spatial resolution Sentinel-2 imagery. The use of these images has posed a challenge due to their spatial resolution (10 × 10 m). Three distinct CNN approaches were proposed to address the varying sizes and morphologies of the input: (1) Resizing, (2) Cropping, and (3) Fully Convolutional Networks (approaches A, B, and C, respectively). These three methods discriminated the crop patterns with accuracies higher than 60%. Although it was expected that approach C, being more complex, would yield the best results, the higher results (80.6%) were ultimately achieved with the second approach B (Cropping). The most frequent misclassifications were observed in intensive densities (approaches A and C) or superintensive patterns (approach A) predicted as traditional olive groves, followed by traditional cropping systems classified as superintensive (approach C). Approach B, based on Sentinel-2 image partition into sub-images, improved the temporal resolution of the analysis (5-day satellite imagery compared to 3-year orthophotos), reaching accuracies of 80%, similar to the averages obtained in previous studies with freely available very high-resolution images (0.5 × 0.5 m). This methodology provides an advanced approach for the discrimination of planting systems to allow optimization of frequencies for the management of olive groves management.