Assessment of remotely sensed inventories for land cover classification of public grasslands in Manitoba, Canada

Land cover classification is one of the most common applications of remote sensing and is used for developing and modifying land management policies on agricultural landscapes to achieve conservation and economic goals, such as reducing grassland degradation and improving livestock and crop production. In this study, the grassland classification of the crown lands (public grasslands in Canada) from a newly developed remotely sensed dataset in the Prairie Province of Manitoba (i.e., the Manitoba Grassland Inventory, MGI) was assessed in terms of accuracy by comparison to non-spatial government records. The analysis consisted of (i) converting non-spatial records from the provincial crown land database to spatially-defined parcels by performing parcel delineations using geographic information system (GIS) and R programming tools, (ii) summarising the MGI classification at the same spatial scale, and (iii) comparing the agreement between MGI and the crown land database. The most common land cover types identified were: forest (30%) and shrubland (25%), followed by native (10%) and tame (9%) grasslands. However, the class agreements between woody (i.e., forests and shrublands) and grassy (i.e., native and tame grasslands) vegetation classes were low between these datasets because of their spectral similarities. Based on these results, we suggest additional refinements on both sensor and ground data to improve the classification agreement between these datasets. This study is one of the first attempts to compare ground-collected government records against a remotely sensed product in Manitoba.

Automated summary

Land cover classification is a common application of remote sensing and can help achieve conservation and economic goals. This study assessed the accuracy of grassland classification in Canadian crown lands using a remotely sensed dataset and compared it with government records. The results showed low agreement between woody and grassy vegetation classes, indicating the need for improvements in sensor and ground data for better classification accuracy.