Detecting and correcting logically inconsistent crop rotations and other land-use sequences

Authors: Mueller Warrant, George; SULLIVAN, CLARE - Oregon State University; ANDERSON, NICOLE - Oregon State University; Whittaker, Gerald

Submitted to: International Journal of Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 3/31/2016
Publication Date: 5/31/2016
Citation: Mueller Warrant, G.W., Sullivan, C., Anderson, N., Whittaker, G.W. 2016. Detecting and correcting logically inconsistent crop rotations and other land-use sequences. International Journal of Remote Sensing. 37:29-59.

Interpretive Summary: Knowledge of crop rotation patterns and perennial crop stand durations has great value for a wide range of agriculture research. Such information can provide richer landuse data for inclusion in Soil-Water-Assessment Tool (SWAT) analyses of the impact of agriculture on water quality. It can also help identify particular fields departing from industry norms in terms of duration of economically viable perennial crop production and ease or difficulty in establishing new stands, the most expensive phase in the life cycle of perennial crops commonly grown in the Pacific Northwest. Such information can be extracted from multi-year landuse data, but is only reliable if certain standards for accuracy and consistency are met. We tested 8 consecutive years of previously published landuse data describing 57 categories ranging from annual crops such as cereals and Italian ryegrass grown for seed to forests and urban development for consistency of apparent year-to-year transitions, finding inconsistencies in nearly 27% of agriculture crop classifications. We used these year-to-year inconsistencies to target replacement of the original classifications with alternate category values, primarily the second place classes of objects (known fields) that had received majority-rule classification in the original analysis. Substantial progress was made in reducing year-to-year inconsistency in landuse classification through an iterative process of identifying inconsistencies, changing classification values for one or both years where inconsistencies existed, and then retesting the entire 8-year long dataset for new year-to-year inconsistencies. A total of 130 rounds of this process were required to reach equilibrium where no further improvement in year-to-year consistency could be achieved. The corrected landuse rasters provided better quality data when converted to crop stand durations than the original landuse rasters did, and should facilitate the next phase of our research as we attempt to identify reasons for variability in useful stand life of crops such as perennial ryegrass. Our approach should be useful for a wide group of researchers collecting and processing multi-year landuse data in settings were some crop rotation patterns are highly likely to occur and others are rare or entirely forbidden by logic.

Technical Abstract: Multi-year landuse data of adequate duration and quality has the potential to identify crop rotation history on individual fields. In the diverse landscape of western Oregon where many crops are established perennials whose stands can remain in production for multiple years, our interests included measuring ages of perennial crops in given years along with overall durations of crop stands in cases where the period covered by data preceded planting of a specific crop and extended past its eventual destruction. Such knowledge would have considerable utility in monitoring an entire industry for signs of variability among fields and farmers in managing pests, establishing new stands, and maintaining economically productive crops. Conversion of multi-year landuse data into field history is relatively simple in the absence of classification errors, but severely compromised in their presence. Many classification errors can be detected by applying a matrix of forbidden year-to-year landuse transitions (e.g., most established perennial crops cannot be changed without one or more intervening years for growing annual [rotational] crops or planting of new stands of grass). Similar restrictions exist regarding transitions among forest and urban development classes. We categorized 730 of 3249 potential year-to-year transitions among 57 landuse classes as being logically permissible, with the remaining 77.5% forbidden. Applying these restrictions to 8 years of landuse raster data in western Oregon revealed that an average of 26.7% of the apparent year-to-year transitions among agricultural classes were illogical, in contrast to only 2.5% of the urban development and 0.6% of the forest transitions. Selecting results from two separate year-to-year transitions with a common middle year identified that middle year as the one most likely in error on an average of 12.4% of agricultural landuse. Corrections applied to the data included replacement of original majority-rule values (generated during prior pixel- to object-based conversion of rasters intersecting 91,442 known fields) with second place classification categories for fields with inconsistent landuses at the beginning, middle, or ending years of sequences. This approach reduced year-to-year landuse inconsistency to 20.1% of the agricultural area, while improving overall classification accuracy by an average of 0.2%. All subsequent corrections involved alternating substitutions of either majority-rule values or second place classification categories within fields when multiple selection criteria were met: year-to-year transitions were inconsistent even after previous rounds of correction, the current classification categories were overrepresented on a county-wide basis relative to independent crop production data, and the second place or majority-rule substitution categories were crops underrepresented on the same county-wide basis. Loosening this third requirement after corrections had stabilized in 13 rounds was necessary in order to make further progress in reducing year-to-year inconsistency. A lengthy series of iterations involving substitution of either majority-rule or second place classification categories (randomized within years and counties) in all cases of year-to-year inconsistency from the previous iteration stabilized at 17.4% inconsistency by iteration number 128. The total reduction in year-to-year inconsistency from 26.7 to 17.2% achieved over 130 iterations came with little loss in overall classification accuracy. Our imposition of multi-year (crop rotation) rules improved our ability to measure duration of perennial crop stands in the complex landscape of the Pacific Northwest. Similar approaches should work to identify and partially correct classification errors in other settings wherever specific year-to-year transitions are either entirely illogical or at least highly