Multi-frequency SAR time series for the detection of sowing events and early-season crop classification

In the context of the sustainable development goals driving 21st-century debates, agriculture faces the challenge of reducing environmental impact while meeting global consumption needs. Tillage activities influence various terrestrial dynamics, including soil erosion, runoff, carbon sequestration and water and carbon dioxide exchange. Monitoring tillage events on a regional scale is therefore essential to facilitate the transition to more sustainable agricultural systems. Synthetic aperture radars (SARs), such as Sentinel-1 and TerraSAR-X, can retrieve information on soil roughness in agricultural fields despite cloud cover, and are particularly sensitive to soil roughness in agricultural fields. The objective of this study is to assess the individual and combined contributions of X-band and C-band SAR data for sowing events detection and early-season crop classification for the most common spring crops grown in Wallonia (Belgium), i.e., beet, potato, and maize. Along with SAR acquisitions, in situ data were collected during a field campaign carried out at the beginning of the 2021 growing season. Between March 26th and May 31st, 166 parcels were visited 7 times, with an average of 11 days between each visit to match the revisit time of TerraSAR-X. At each visit and for each plot, the soil cover, i.e. bare, covered, presence of crop residues, and its roughness, i.e. shallow or deep furrows, crop mounds, or seedbed, were recorded. A sowing period detection method based on VV coherence drop showed promising results in highlighting homo- and heterogeneity in spatial (regional mapping) and temporal distribution of sowing events for studied cultures. Although the available validation data for sowing detection has been limited, expanding the study to a regional and national scale within the context of the BELCAM (BELgian Collaborative Agriculture Monitoring) project will enable greater validation opportunities. Despite these limitations, the results provided valuable insights into crop classification, revealing that the sowing period for each crop type significantly influences classification performance when using C-band or X-band data. More globally, Random Forest classifications have primarily highlighted three key parameters that contribute to better classifications: the necessity of a roughly balanced dataset, the significance of the sowing period, and the critical role of high temporal resolution. While Sentinel-1 outperformed TerraSAR-X for beet and potato fields classification, the best overall results were achieved using a combination of both sensors acquisitions, which leveraged the small-scale roughness sensitivity of X-band data and the sensitivity to sub-surface soil layers of C-band data. In this context, beet fields were accurately classified in early April, potato fields in late April, and maize fields in the final weeks of May. This underscores the importance of utilizing multi-frequency information to accurately classify selected crops as early as possible in the growing season. The sowing events detection and early-season crop classification act as valuable tools for assisting farm advisory services, agricultural insurance companies, improving or build new agricultural policies, running yield or nitrogen uptake models, etc. Results also offer promising perspective for further improvement and analysis in the monitoring of tillage practices. Further efforts targeting the mapping of tillage or no-tillage practices could improve reporting of greenhouse gas emissions at a regional level, taking a better account of the diversity of farming practices, as well as the annual assessment of farmers' efforts to move towards conservation agriculture. These results could also be used by public administration as part of the Common Agricultural Policy land monitoring system.