Interseasonal transfer learning for crop mapping using Sentinel-1 data

Abstract

Crop maps are highly desired information in modern agriculture as they enable possessors to manage their business in the most optimal way. Usually in remote sensing, crop mapping is performed using satellite images and within-season ground truth samples that are collected in extensive survey campaigns every year, neglecting information and knowledge that past seasons’ classification models provided. This paper assessed different temporal transferring approaches, including transfer learning, together with traditional crop mapping approach to provide an exhaustive comparison. Transferring approaches differed in portion of knowledge utilized from a historical model and that coming from a target season dataset. Three distinct algorithms, Random Forest, Convolutional Neural Network and Transformer, were employed and evaluated using highly dense time series of Sentinel-1 data. Source and target domain were respectively represented by two sets, 2017–2020 and 2021 season data, and 9 different crop types were classified. Results showcased that transferring a model has a great potential in crop mapping when little to no ground truth data is available for the target season. However, traditional approach catches up rather quickly and even surpasses transfer learning approach in terms of performance after a certain portion of target domain data is collected. Without target season ground truth data, model transferring can yield modest crop mapping performance of 78% for F1 score, between 84% and 86% F1 score with transfer learning employed in conjunction with limited target season ground truth (i.e. between 120 and 720 parcels), and 88% F1 score at best with traditional approach (ca. 720 parcels). Even though a good discriminatory is found between different crop types, there is still a room for improvement regarding the least represented classes in the dataset. The study significantly contributes to the area of agricultural monitoring and management by demonstrating the effectiveness of transfer learning while lessening the necessity for extensive and labor-intensive data collection, thereby fostering cost and time efficiency. Utilizing Sentinel-1 data, it provides a practical and efficient solution for agricultural analysis worldwide regardless of cloudiness.