Enhancing Crop Discrimination And Croplands Mapping In Africa Using Satellite Data For Food Security

Food security is important in Africa given the rapid annual population growth rate of 2.7%,and low agricultural productivity.Decision‐making regarding food security issues requires spatial information on crop type distribution and cropland extent.Zimbabwe and Africa,in general,have an ostensible lack of spatially explicit,reliable,precise,and up‐to‐date information pertaining to crop type distribution and cropland extent.The mapping of subtropical,humid,and tropical regions is a daunting task due to the inescapable cloud cover problem.Radar imagery has few polarization bands that can limit the ability to do traditional digital classification.While the importance of cropland extent maps is widely recognized,the production of these maps remains a challenge due to a number of technical and non‐technical reasons.The overarching objective of this study is to develop scalable,robust and repeatable crop classification approaches exploiting the spectral,textural,and temporal features of remote sensing datasets to enhance crop discrimination and cropland mapping in Zimbabwe and the rest of Africa for food security.To achieve this general objective,(i)cropping patterns were determined using Sentinel‐1 SAR data.(ii)A decision‐level data fusion approach was implemented to ensemble classify crops using multi‐source optical imagery(Landsat 8,Sentinel‐2 and Landsat 7).(iii)A decision‐level integration of Sentinel‐1 SAR and Landsat 8 OLI texture features approach was developed for crop discrimination and classification.(iv)The cropland extent and change detection in Zimbabwe from 2000 to 2018 was executed.(v)The cropland extent and change analysis in Africa from 2000 to 2018 was also executed.For the mapping of cropping patterns on smallholder scale croplands,two study sites were considered namely: International Maize and Wheat Improvement Center(CIMMYT)research station and a few neighboring fields,and the Middle Sabi Estate.Fourier time series analysis was employed to smooth and identify the various trends transpiring on both sites under investigation.Stacked time‐series images were classified using K‐means and random forest algorithms.Results exhibited that Sentinel‐1 SAR time‐series data permit mapping the crop distribution,cropping patterns and determination of subtle changes on the crops and field parcels irrespective of parcel size.Cropping pattern maps were synthesized from the classification of stacked images.Random forest classification of the multi‐temporal image stacks attained high overall accuracies of 99% and 95% on the respective study sites.For the integration of multi‐source optical sensors,the heterogeneous ensemble classifier employs a parallel and concatenation approach.Pixel‐ and decision‐level fusions approaches were designed and their performances on discriminating and mapping of crop types were compared.The base classifiers in the multi‐classifier system were Support Vector Machines,Spectral Information Divergence,and Maximum Likelihood.The plurality voting method was implemented to achieve the decision‐level fusion.Landsat 7 or Sentinel‐2 cloud‐free pixels were appended to Landsat 8 gaps by mosaicking to achieve the pixel‐level fusion.Classification accuracies of both fused classified images were compared and evaluated.Decision‐level fusion achieved an overall classification accuracy of 85.4% and kappa coefficient of 0.84,whereas pixel‐level fusion classification attained 82.5%,and kappa coefficients of 0.80,however Z‐test at α = 0.05 showed that the results from the two approaches are not significantly different.F1‐test was computed on all individual classes from both approaches,and decision‐level outperformed pixel‐level on most of the individual classes.A regression coefficient of 0.99 was obtained between the planted areas statistics extracted from the two approaches.Nevertheless,Support Vector Machines outperformed the other base classifiers.For the integration of SAR and optical data,texture features were derived from Landsat 8 OLI and dual‐polarized Sentinel‐1 SAR speckle filtered and unfiltered backscatter.Their respective classification results were aggregated using the decision‐level fusion.The accuracy of the decision‐level fused maps was evaluated.Gray level co‐occurrence matrix(GLCM)was employed to derive sets of seven texture features for Landsat 8 bands and VV+VH backscatter using 5 × 5,7 × 7,9 × 9 and 11 × 11 window sizes.To avoid overfitting,each feature is stacked with respective source images and classified independently using support vector machines algorithm.Classified maps from the best three performing textures from both speckle filtered and unfiltered are aggregated with classified maps from Landsat 8’s best three performing textures using plurality voting algorithm and the two seamless classified maps are compared using Z‐test.Results attained indicate an overall classification accuracy of 96.02% from decision‐level fused images of Landsat 8 + Sentinel‐1’s non‐speckle filtered maps,whereas Landsat + speckle filtered achieved 94.69%.Best texture information from Landsat 8 was derived from the blue band followed by a red band,whereas Sentinel‐1 SAR’s speckle unfiltered textures performed better than speckle filtered textures.We conclude that integration of Landsat 8 and Sentinel‐1,either speckle filtered or unfiltered,improves crop overall classification.However,speckles do not have statistically significant effects(p=0.1208)on texture feature derivation.For the mapping of cropland extent in Zimbabwe,three approaches(i)automatic classification(ii)multi‐classifier system(MCS)(iii)NDVI‐BSI threshold were employed.The spatio‐temporal cropland changes were determined using the best performing approach.The change detection was achieved by performing a post‐classification statistical method.The quality of classified observations was assessed by comparing them with google earth imagery,GFSAD30 AFCE cropland layer,cropland class from ESA and SADC land cover products.Observations executed revealed that performances of MCS and NDVI‐BSI were comparable and better than automated classification.They attained overall accuracies of 80.54% and 79.32% for 2013 and the accuracies achieved for 2018 were 87.90% and 88.56% respectively.Respective mean cropland areas extracted for automated classification,MCS,and NDVI‐BSI threshold were 3416396 Ha,10346778 Ha,and 9788833 Ha.Visual evaluation of the observations revealed that NDVI‐BSI threshold outperformed the other two approaches.A further comparison was performed using total cropland areas extracted from MCS and NDVI‐BSI threshold techniques for Zimbabwe’s ten provinces for the years 2013 and 2018.The coefficients of determination of 0.8404 and 0.9619 were observed respectively.Change detection revealed a general expansion of the cropland area resulting from human activities regardless of the prolonged drought.A stratum‐specific cropland mapping approach using a logical combination of spectral indices(LCoSI)that delineates both rainfed and irrigated cropland extent for Africa was proposed.The stratification is based on the agro‐ecological zones(AEZ)of Africa.Optimal threshold values for BSI,DBI,GNDVI,GVI,MNDWI,NBLI,NBRI,NDMI,and NDVI for the AEZ based on a thousand samples were computed and implemented using conditional equations.The cropland extents for the different AEZ were integrated into a seamless cropland layer.The 2018 cropland layer was synthesized from Landsat 8 data,whereas the 2000 cropland layer was synthesized from Landsat 7 data.Accuracy assessment of the 2018 cropland layer was achieved by using overall accuracy,kappa coefficient,user’s accuracy,producer’s accuracy,quantity disagreement,and allocation disagreement.Accuracies of 97.5%,93%,88.94%,100%,2.5%,and 0.0% were achieved when referenced to the samples from the ESRI world map were employed.It was also compared with the GFSAD30 AFCE cropland layer,IWMI cropland product,and ESA CCI LC prototype cropland layer and produced higher accuracy.Change analysis was performed between the 2000 and 2018 cropland extents and a general expansion of cropland was observed.The expansion is due to the shifting cultivation(by slash‐and‐burn)cropping pattern practiced,population growth and government polices promoting crop cultivation.For Zimbabwe,cropland extent increase of 7.63% was observed from 2000 to 2018 using areas extracted from the LCoSI approach.A cropland expansion of 7.36% was computed from 2001 to 2018 using areas extracted from the NDVI‐BSI approach.The LCoSI approach has great potential to map cropland at large scales irrespective of landscape or AEZ or field size.Further research should focus on automating the approach.Overall,the findings of this study showed that the recently launched Sentinel‐1provides enough spatial resolution that allows the mapping of cropping patterns irrespective of farm size.The spatial coverage due to inescapable cloud problems can be overcome by the integration of multi‐sensor data and decision‐level fusion since it does not require prior knowledge pertaining to the spectral configuration of the various sensors involved.Ensemble classification enhances crop classification.Speckle filtering is not necessary when extracting SAR‐based texture features prior to crop classification.The stratum‐specific mapping approach using logical combination of spectral indices(LCoSI)offers greater potential to the mapping of croplands.However,the expansion in cropland extent is influenced by shifting cultivation cropping patterns,policies regarding crop production and human activities regardless of the prolonged drought.Limitations of the study included inadequate in‐situ reference observations.The major drawback of google earth engine in the classification of large areas such as Africa is due to the limited size of training samples acceptable by the platform.Recommendations for future studies include the discrimination of fallow land from abandoned land.