Crop Area Mapping in Southern and Central Malawi With Google Earth Engine

Abstract

Agriculture in sub-Saharan Africa consists primarily of smallholder farms of rainfed crops. Historically, satellite data were too coarse to account for the heterogeneity in these landscapes. Sentinel-2 data have improved spectral resolution and much higher spatial resolution (10 m) than previously available satellites with global coverage, such as Landsat or MODIS, making mapping smallholder farms possible. Spectral mixture analysis was used to convert the Sentinel-2 signal into fractions of green vegetation, non-photosynthetic vegetation, soil, and shade endmembers. Very high spatial resolution imagery in Google Earth Pro was used to identify locations of crop and natural vegetation classes, with over 20,000 reference points interpreted. The high temporal resolution of Sentinel-2 (5 days repeat) allows for classification of landcover based on the phenological signal, with natural areas having smoothly varying amounts of photosynthetic vegetation annually, while cropped areas show more abrupt changes, and also the presence of bare soil due to agricultural activity at some point during the year. We summarized the endmember values using monthly medians, extracted values for the reference data points, randomly split them into training and test data sets, and input the training data into the random forests algorithm in Google Earth Engine to map crop area. We divided southern and central Malawi into tiles, and found crop/no crop classification accuracies on the test data for each tile to be between 87 and 93%. The 10 m map of crop area was aggregated to the district level and showed an R2 of 0.74 with ground-based statistics from the Malawi government and 0.79 with a remotely sensed product developed by the USGS.

Crop area change in the context of civil war in Tigray, Ethiopia

“…The initial effort to estimate cropped area in Tigray follows the technique successfully used in Malawi (Peterson and Husak 2021). Reference data points for crops, natural vegetation (sparse, shrub, tree/forest), and water were identified by hand using all available VHR true color imagery in GEPro, progressing from West/Northwest, Central, East, to Southeast/South Tigray (Peterson 2024).…”

“…Assessing the extent of agricultural land from satellite information requires data and techniques that are tuned to this smallholder, subsistence agriculture environment. The introduction of 10 m Sentinel-2 data made it much more feasible to map crop area in these smallholder landscapes compared to 30 m Landsat data (Delrue et al 2013, Jolivot et al 2021, and much work has been done in this area (Defourny et al 2019, Masiza et al 2020, Olsen et al 2021, Peterson and Husak 2021. While some of these studies are solely performed remotely, cropped area assessments can be informed with the aid of field campaigns, surveys, and work with local extension officers to provide on-the-ground feedback.…”

“…Machine learning-based studies of crop area will often include vegetation indices, such as NDVI, as well as raw bands as predictors (Jin et al 2019, Karlson et al 2020, Kerner et al 2020. An important part of machine learning in landcover classification using remote sensing is human intervention to ensure that the results are logical/interpretable (Peterson and Husak 2021). For instance, studying greenup (increase in GV fraction), crop residues (NPV fraction) or plowing (soil, shade) is much more readily undertaken with EM fractions than raw bands.…”

“…Sentinel-2 data have a high temporal (5 d repeat), spatial (10-20 m), and spectral (10 bands) resolution to generate high quality timeseries of GV, NPV, soil, and shade which allow for good understanding of agricultural activity remotely. VHR data (∼1 m, RGB), such as that available in Google Earth Pro (GEpro), allow for the identification of agricultural fields via observing the presence of plowed rows, harvest residue, regular field boundaries and the like, thus a reference dataset can be generated to train and test the model (Witmer 2008, Peterson andHusak 2021).…”

“…A methodology for crop mapping was developed for Malawi that combined training data from the inspection of VHR imagery in GEpro, and timeseries of GV, NPV, soil, and shade EM fractions from Sentinel-2 as predictor variables in the random forests (RFs) algorithm in Google Earth Engine (GEE) (Peterson and Husak 2021). The work presented here investigates the applicability of this methodology to the conflict-stricken region of Tigray, Ethiopia, with the goal of quantifying the impact that conflict had on change in cropped area.…”