Search

Sentinel-2 has opened a new era for the remote sensing community where 10-m imagery is freely available with a 5-day revisit frequency and a systematic global coverage. Having both frequent and detailed observations across large geographic areas are ideal characteristics that can potentially revolutionize applications such as crop mapping and monitoring. However, such large volumes of high-resolution data pose challenges to users in terms of problem complexity, computational resources and processing time, beckoning the increasingly relevant question: at which resolution should this imagery be processed? Here, we develop a methodology to characterize resolution-dependent errors in cropland mapping and explore their behavior when we move across spatial scales and landscapes, taking special care to include the effects of the instrument's Point Spread Function (PSF). Results show how local upscaling of 10-m imagery, e.g., from Sentinel-2, to 30 m mitigates most the adverse effects generated by the PSF when comparing it to native 30-m imagery, e.g., from Landsat-8. Extending this logic, we demonstrate for two nationwide cases how maps can be calculated showing the optimal spatial resolution that keeps resolution-dependent errors below a user-defined threshold. Based on these maps, we estimate that 31% of Belgium and 59% of South Africa could be processed at 20 m instead of 10 m, while keeping the increase of resolution-dependent errors below 3%. These local resolution adjustments lead to a reduction in data volume and processing time by 23% and 44%, respectively.

Pixel counting is probably the most popular way to estimate class areas from satellite-derived maps. It involves determining the number of pixels allocated to a specific thematic class and multiplying it by the pixel area. In the presence of asymmetric classification errors, the pixel counting estimator is biased. The overarching objective of this article is to define the applicability conditions of pixel counting so that the estimates are below a user-defined accuracy target. By reasoning in terms of landscape fragmentation and spatial resolution, the proposed framework decouples the resolution bias and the classifier bias from the overall classification bias. The consequence is that prior to any classification, part of the tolerated bias is already committed due to the choice of the spatial resolution of the imagery. How much classification bias is affordable depends on the joint interaction of spatial resolution and fragmentation. The method was implemented over South Africa for cropland mapping, demonstrating its operational applicability. Particular attention was paid to modeling a realistic sensor's spatial response by explicitly accounting for the effect of its point spread function. The diagnostic capabilities offered by this framework have multiple potential domains of application such as guiding users in their choice of imagery and providing guidelines for space agencies to elaborate the design specifications of future instruments.

A precise knowledge of the crop distribution in the landscape is crucial for the agricultural sector to inform better management and logistics. Crop-type maps are often derived by the supervised classification of satellite imagery using machine learning models. The choice of data sampled during the data collection phase of building a classification model has a tremendous impact on a model's performance, and is usually collected via roadside surveys throughout the area of interest. However, the large spatial extent, and the varying accessibility to fields, often makes the acquisition of appropriate training data sets difficult. As such, in situ data are often collected on a best-effort basis, leading to inefficiencies, sub-optimal accuracies, and unnecessarily large sample sizes. This highlights the need for new more efficient tools to guide data collection. Here, we address three tasks that one commonly faces when planning to collect in situ data: which survey route to select among a set logistically feasible routes; which fields are the most relevant to collect along the chosen survey route; and how to best augment existing in situ data sets with additional observations. Our findings show that the normalised Moran's I index is a useful indicator for choosing the survey route, and that sequential exploration methods can identify the most important fields to survey on that route. The provided recommendations are flexible, overcome the main logistical constraints associated with in situ data collection, yield accurate results, and could be incorporated in a mobile application to assist data collection in real-time.