Crop mapping using high resolution optical and radar satellite data: case study of maize in Trans Nzoia County, Kenya

Abstract

Agriculture and food security are some of the main drivers of strong societies, essentially propelling countries’ economies. However, the reality is that buffeted by climate change, urbanization and accompanying unsustainable human activities, crop yields have been on a monotonic decline. With the advent of freely available high resolution optical and radar images, a paradigm change in agriculture mapping and crop monitoring has been witnessed. The objective of this study was to map the maize fields in Trans Nzoia county and monitor the growth conditions from optical images (Landsat-8– L-8, RapidEye- RE, and Sentinel-2– S-2), and radar images (TerraSAR-X- TSX, and Sentinel-1- S-1) during the 2015 cropping season, and explore the integration of the results in policy advise. 18 large scale maize fields with existent field management data were selected, and their field leaf area index (LAI) modeled. The field areas were extracted from the satellite images and compared to the field areas reported by the farmers. Classification of various combined sets of the optical and radar images was carried out as well, for crop type mapping. Monitoring of the maize using the optical images involved relating the vegetation indices (VIs) (EVI2, SAVI, NDVI, NDVIre and gNDVI) from L-8 and RE images, whereas the 2016 S-2 LAI values validated the field modelled LAI. On the other hand, monitoring of the maize with radar images involved analysing the image backscatter (BS) values for the cropping season. This was threefold: the analysis of the TSX and S-1 BS values separately; comparison of TSX and S-1 BS values; comparison of the 2015 and 2016 S-1 BS values. The maize phenological growth stages were described by the Biologische Bundesanstalt, Bundessortenamt und CHemische Industrie (BBCH) scale. From the results, 38% of the maize fields reported similar areas between the satellites measured areas, and the reported areas recorded by the farmers, 33% of the fields’ areas were overestimated while the remaining 29% were underestimated. Combining the TSX, RE, and RE-NDVIre images achieved the best classification results (user accuracy -72%, producer accuracy -89%). The line of best fit for modeled LAI against the VIs plots was linear with R2 values of 0.88 (0.07 RMSE), 0.81 (0.09 RMSE), 0.8 (0.07 RMSE), and 0.82 (0.04 RMSE) for EVI2, SAVI, NDVI and gNDVI respectively. Due to unavailability of images during the cloudy early season, the conventional exponential LAI vs VIs model fits was not realized. Validation of the 2015 modeled LAI with 2016 S-2 LAI achieved an R2 of 0.54 (RMSE 0.31). Fewer validation points were available during the early crop season due to cloud influence on the S-2 images. The fields S-1 BS values were grouped according to the cultivated maize varieties, the fields’ orientation, and the planting dates. Of the three factors, the fields’ orientation had the greatest influence on the BS curves. The backscatter values increased during the rapid growth stages, before saturating at BBCH 7, and later decreased towards senescence. The phenological stages were identifiable from the image backscatter values, with either a sudden increase or decrease in the backscatter values at the main growth stages. Comparing S-1 to TSX, similar BS curves were observed for the various phenological stages, though TSX BS values were higher than the S-1 values. For the 2015 and 2016 S-1 BS values comparison, the VV ascending IW1 mode produced the best results in comparing the maize phenological characteristics from one year to the next. Estimation of the field sizes from satellite images provided a fast, accurate, and cost effective method of acreage estimation which would improve yields estimations. With the increase in the number of freely available high-resolution optical and radar satellite images, and with a high repeat cycle, crops can be monitored for the entire cropping season. The phenological analysis of the backscatter values provides a tool in the monitoring and evaluation of the maize fields for the entire season, independent of weather conditions. A reference baseline from S-1 backscatter values can be formulated for maize monitoring during subsequent years. The study provides evidence and results in support of the implementation of satellite data in the formulation of policy towards crop monitoring and food security. Policy areas include: elimination of fraud in the provision of government subsidies occasioned by the overestimation of field sizes; making timely decisions on the need for maize importation in the event of a poor crop season; adoption of irrigation policies to complement the rains in cases of poor rain seasons