Advancing agricultural remote sensing: A comprehensive review of deep supervised and Self-Supervised Learning for crop monitoring

Abstract

Deep Learning based on Remote Sensing has become a powerful tool to increase agricultural productivity, mitigate the effects of climate change, and monitor deforestation. However, there is a lack of standardization and appropriate taxonomic classification of the literature available in the context of informatics. Taking advantage of the categories already available in the literature, this paper provides an overview of the relevant literature categorized into five main applications: Parcel Segmentation, Crop Mapping, Crop Yielding, Land Use and Land Cover, and Change Detection. We review notable trends, including the transition from traditional to deep learning, convolutional models, recurrent and attention-based models, and generative strategies. We also map the use of Self-Supervised Learning through contrastive, non-contrastive, data masking and hybrid semi-supervised pretraining for the aforementioned applications with an experimental benchmark for Post-Harvest Crop Mapping models, and present our solution, SITS-Siam, which achieves top performance on two of the three datasets tested. In addition, we provide a comprehensive overview of publicly available datasets for these applications and also unlabeled datasets for Remote Sensing in general. We hope that our work can be useful as a guide for future work in this context. The benchmark code and the pre-trained weights are available in https://github.com/mateuspinto/rs-agriculture-survey-extended.