Disruptive events such as conflicts or the COVID-19 pandemic can have adverse effects on regional agricultural production and food security, while at the same time rendering those regions inaccessible to organizations that aim to measure the impact of the events on food security to inform humanitarian support. Satellite Earth observation (EO) data can provide a remote look into crop planting and conditions in regions not accessible by traditional means of assessment on the ground (e.g., windshield or household surveys), while machine learning (ML) enables rapid analysis of large EO datasets. Here, we present our approach for estimating annual planted area and inter-annual planted area change using Earth observations and machine learning, which we implemented in support of the USAID Famine Early Warning Systems Network (FEWS NET).
The first step in our approach is to generate a cropland probability map where each pixel contains the probability of a pixel containing active cropland as predicted by our ML classifier (Kerner & Tseng et al., 2020). Our classification can be applied to past seasons or within an ongoing growing season. The ML classifier is trained using labeled samples located in the study region as well as globally and may come from a variety of data sources. To ensure the resulting map will have sufficient quality for downstream analysis, we iteratively improve the model classification performance using a held-out validation set of 525 samples with locations chosen using a random uniform sampling strategy in the study region. These validation samples are labeled as either planted crop or not-planted/non-crop by manually inspecting monthly high-resolution PlanetScope basemaps from the mapping year; multiple individuals label every point so that we can assess label confidence and review non-consensus labels. When estimating inter-annual change, we then difference the maps from each year to obtain a change map with four classes: stable planted, stable not-planted, planted gain, and planted loss.
Reliable area estimates cannot be obtained by pixel counting (i.e., summing the pixels in each class and multiplying by the area covered by each pixel) from predicted maps. We use the best practices outlined in Olofsson et al., 2015 to adjust our map-based areas using a reference sample drawn from the predicted maps using a random sample stratified by the map classes. We label these reference samples using the same procedure described for the validation samples, taking care to increase label confidence through consensus and thorough review. Finally, we adjust the map-based areas for each class using the confusion matrix between the predicted map classes and reference sample labels. This results in class-wise annual or inter-annual change estimates of planted area with associated standard error for a 95% confidence interval.
We present the results from applying this approach to estimate planted area change in multiple regions, including the Tigray region of Ethiopia where the ongoing civil war has been reported to have disrupted crop production and put residents at risk of famine. We estimated the change in planted area between 2020 and 2021 and found that there was some planted area loss in 2021, particularly in localized regions that may have been impacted by conflict events, but there was also significant planted area gain, likely due to lower crop production in 2020 due to drought conditions that improved in 2021.
References
- Kerner, H. R., Tseng, G., Becker-Reshef, I., Barker, B., Munshell, B., Paliyam, M., Hosseini, M. (2020). Rapid Response Crop Maps in Data Sparse Regions. Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining Workshops.
- Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E., & Wulder, M. A. (2014). Good practices for estimating area and assessing accuracy of land change. Remote Sensing of Environment, 148, 42-57.
# Abstract
Thanks to its primary ability of restoring soil fertility, fallow practices are an integral part in cropping systems worldwide. More recently, their importance has been further pointed out, considering the related implications on climate change mitigation through carbon sequestration and biodiversity conservation (Dayamba et al., 2016; Ringius, 2002). Mapping fallow land has hence become a key challenge to assess the impact of such practices on the sustainability of agricultural systems.
Sentinel-2 missions have enhanced the ability for mapping agricultural land due to its high revisit time (5-day) and spatial resolution (10-meter pixels), easing the task of monitoring cropping practices and allowing land mapping in areas with relatively small plots. This has an even greater importance in West African countries where political and socioeconomic instability renders traditional methods for land mapping/inventory expensive and dangerous (Nakalembe, 2021; Sahajpal et al., 2020). Nonetheless, fallow land mapping has been, for the most part, overlooked in mainstream land cover products with little to no discrimination between active cropped land and fallowed land. Still, these global/regional land cover products are widely used as a basic information for many crop monitoring tasks, including yield estimation and forecasting in food security early warning systems (Nakalembe et al., 2021).
Few studies have proposed a land cover mapping methodology specifically for fallow fields and, to the best of our knowledge, only one has provided tests of such a method across the Sahel, reporting that more than 50% of cropland class in most of the global land cover products are fallow fields (Tong et al., 2020). However, the unsupervised methodology they used relies on a strong hypothesis on seasonal NDVI profiles (i.e. cropped fields have in general a lower NDVI compare to fallow fields across the cropping season) whose pertinence at both local scales and outside the Sahelian area may be questionable.
In this study, we present the outcomes of a first exploratory analysis on fine scale, remote sensing based characterization of fallow practices, carried out over a study case located in the Sudanian region of Burkina Faso. Leveraged data consist of a Sentinel-2 multi-year image time series, appropriately pre-processed and coupled with detailed, small scale in-situ data derived from a recently published agricultural land cover database for the 2015-2021 period (Jolivot et al., 2021), built up during the JECAM experiment of the GEOGLAM network (http://jecam.org/). In order to test the suitability of the Tong et al. (2020) underlying NDVI hypothesis for our study area, we first replicated the aforementioned reference methodology, but did not reach satisfying accuracies, with both producer and user accuracy below 50% and highly overestimating the proportion of fallow land when validating with JECAM database.
We then provide an expertise-based exploratory analysis of fallow-field NDVI profiles in order to come up with a more suitable set of hypotheses which could be used in the definition of a novel methodology for remote sensing based fallow mapping. Our preliminary results highlight that seasonal NDVI-based fallow discrimination approaches are not sufficient for discriminating fallow fields from other cropped areas. Conversely, we come out with several evidences that multi-year NDVI fallow characterization might a more suitable approach, for example by showing that transitions of fields from cropped to fallow and vice versa may have a measurable impact on vegetation index dynamics over multiple years (see attach figures).
In parallel, we also performed a data-driven analysis in which we use common machine learning techniques to provide automatic fallow mapping through supervised image classification. The rationale of this part of the study is two-fold : (1) assess the potential of supervised classification and build a “baseline” set of fallow maps for the period covered by the reference database, and (2) exploit a larger sets of variables, radiometric (i.e. derived from Sentinel-2 time series - vegetation, water and soil indices) as well as other types of geo-spatial data (such as soil type or rainfall data), and explore their correlation with the dynamics of agricultural practices and crop vegetation development. Although no clear hypothesis can be made yet for the design of a novel methodology suitable for upscaling, we come up with further clues that integrating multi-year strategies into state-of-the-art land cover mapping techniques may be a promising approach in tackling the complex, yet key task of fallow monitoring in West African agrosystems.
# References
Dayamba, S. D., Djoudi, H., Zida, M., Sawadogo, L., & Verchot, L. (2016). Biodiversity and carbon stocks in different land use types in the Sudanian Zone of Burkina Faso, West Africa. Agriculture, Ecosystems & Environment, 216, 61–72. https://doi.org/10.1016/j.agee.2015.09.023
Jolivot, A., […], Gaetano, R., […] Leroux, L., […] Bégué, A. (2021). Harmonized in situ datasets for agricultural land use mapping and monitoring in tropical countries. Earth System Science Data [Preprint]. https://doi.org/10.5194/essd-2021-125
Nakalembe C et al. (2021). A review of satellite-based global agricultural monitoring systems available for Africa, Global Food Security, 29 : 100543. https://doi.org/10.1016/j.gfs.2021.100543
Ringius, L. (2002). Soil carbon sequestration and the CDM: Opportunities and challenges for Africa. Climatic Change, 54(4), 471–495.
Sahajpal, R., Fontana, L., Lafluf, P., Leale, G., Puricelli, E., O’Neill, D., Hosseini, M., Varela, M., & Reshef, I. (2020). Using machine-learning models for field-scale crop yield and condition modeling in Argentina. 49º Jornadas Argentinas InformáTica, Congr. Argentino AgroinformáTica, 1–6.
Tong, X., Brandt, M., Hiernaux, P., Herrmann, S., Rasmussen, L. V., Rasmussen, K., Tian, F., Tagesson, T., Zhang, W., & Fensholt, R. (2020). The forgotten land use class: Mapping of fallow fields across the Sahel using Sentinel-2. Remote Sensing of Environment, 239, 111598. https://doi.org/10.1016/j.rse.2019.111598
The United States (US) Department of Agriculture has been producing 30-meter annual crop cover classifications over the conterminous US dating back to 2008. The product, coined the Cropland Data Layer (CDL), has been derived from a variety of satellite imagery sources but primarily includes Landsat and, since 2016, Sentinel-2. The CDL has provided the US agricultural community a rich dataset for exploring topics such as crop extent monitoring, rotation pattern modeling, and small area estimation. The now 13-year CDL archive has also provided a robust “ground truth” dataset to help extrapolate classifiers forward in time for use in real-time US crop type identification or, conversely, applied retrospectively to generate CDL-like maps prior to 2008.
What has not been explored is the potential use of spectral-temporal training signatures derived from the CDL to be applied to crop areas internationally. The idea being the development of generic training datasets, or “spectral libraries”, in the US from the CDL that can be applied, or “transfer learned”, to similar crop regimes globally. These classifications are highly desirable because there are no systematic crop type maps are being generated annually outside of North America.
To test this concept, archived Landsat 7/8 and Sentinel-2a/b imagery was used as a basis in parallel with the CDL. In short, cloud-free imagery composites were built consistently throughout each of the past five growing seasons, stacked by year, and then intersected with the corresponding CDL. Next, training signatures were randomly drawn from each of the five years, using the CDL as the label information. Then all years’ samples were merged as one to better captures across year cropping and imagery quality variability. Next, the samples were sent to decision tree style classifiers. Tested were the common machine learning tools of classification and regression tree (CART) and random forests (RF). Finally, the derived rulesets outputs from those classifiers were applied across space to the similarly composited predictor imagery for the years 2018, 2019, and 2020 and maps generated.
Crop intensive areas of the US were used to generate the training data with the focus on the ubiquitous commodities of maize, soybeans, wheat, and cotton. Example CART and RF based classifications were generated in a sampling of industrialized cropping areas around the world, including regions from both the northern and southern hemispheres. Quantitative map accuracies were unable to be produced since there is no validation data in the foreign test cases. However, results were subjectively inspected and often look quite reasonable, particularly for the RF cases. Perceived robustness and utility of the methodology will be shown as well.
Detailed parcel-level crop type mapping for the whole European Union (EU) is necessary for the evaluation of agricultural policies. The Copernicus program, and Sentinel-1 (S1) in particular, offers the opportunity to monitor agricultural land at a continental scale and in a timely manner.
However, so far, the potential of S1 has not been explored at such a scale. Capitalizing on the unique LUCAS 2018 Copernicus in-situ survey, we present the first continental crop type map at 10-m spatial resolution for the EU based on S1A and S1B Synthetic Aperture Radar observations for the year 2018. Random Forest (RF) classification algorithms are tuned to detect 3 land cover types and 19 different crop types. We assess the accuracy of this EU crop map with three approaches. First, the accuracy is assessed with independent LUCAS core in-situ observations over the continent. Second, an accuracy assessment is done specifically for main crop types from farmers declarations from 6 EU member countries or regions totaling more than 3 million parcels and 8.21 Mha. Finally, the crop areas derived by classification are compared to the subnational (NUTS 2) area statistics reported by Eurostat.
The overall accuracy for the crop type map is reported as 80.3% when grouping main crop classes and 76% when considering all 19 crop type classes separately. Highest accuracies are obtained for rape and turnip rape with user and produced accuracies higher than 96%. The correlation between the remotely sensed estimated and Eurostat reported crop area ranges from 0.93 (potatoes) to 0.99 (rape and turnip rape). We discuss how the framework presented here can underpin the operational delivery of in season high-resolution crop mapping.
The perspectives of the EU crop map are multifold, both using the final product to extract further information on the agricultural landscape and investigating future developments. We are currently evaluating the fit for purpose of the EU crop map to derive indices of pesticide risk exposure and crop diversity in space. Based on the final map, we are also investigating how to use crop area statistics at the subnational level and RF class-specific probabilities to improve the accuracy of the EU crop map. In terms of future development, the main efforts are on developing multi-year mapping approaches on specific areas.
Satellite imagery has shown large capabilities in agriculture for monitoring crop conditions. Various vegetation indices are in use for the purpose of better decision taking in agricultural practices regarding irrigation, agricultural inputs, seeding and harvesting. Most common is the use of NDVI. Though being a proven index for monitoring photosynthetic activity in the field, the possibility of consistent and continuous monitoring of crops is greatly hampered by the presence of clouds. This is why others have engaged in deriving vegetation indices based on Synthetic Aperture Radar (SAR) observations.
By being able to penetrate clouds, SAR has the potential to be a temporally consistent source of information for crop health monitoring. However, SAR vegetation monitoring comes with its own challenges, mostly being connected to high noise-to-signal ratios causing accuracy of spatial patterns to be low when compared to optical remote sensing.
The Biomass Proxy algorithm is made to integrate the temporal consistency of SAR with the spatial accuracy of optical remote sensing. A novel fusion methodology has been developed into an operational near real time product that now serves more than 20 countries spread over the globe. This methodology uses a dynamic fusion of both temporal and spatial signals of NDVI and SAR derived vegetation index at field level and produces daily biomass indications at a 10m x 10m resolution. It has been evaluated and tested in Germany, Ukraine, United Kingdom, Canada and Brazil where there are various conditions in terms of cloud cover, agricultural practices as well as satellite coverage.
As expected, performance is directly linked to the number of satellite coverages available. However, the algorithm demonstrated to be able to mitigate biomass indications with potentially lower quality, both spatially and temporally, in periods in which no optical information was available by making effective use of SAR. It also showed that noise in SAR observations did hardly propagate into temporal and spatial output using the built-in dynamic fusion weights mechanism of the algorithm. The biomass proxy has shown to be robust under various climatic conditions such as heavy rainfall, snow and drought, and consistent values have been observed for various crops such as cereals, sugarcane, potatoes, sunflowers and alfalfa. Furthermore, the methodology has been designed to be independent of the input signal source and is able to incorporate future sources and therefore different overpass frequencies such as the inclusion of Sentinel-1C and -D and Sentinel-2C and -D or other optical sources such as Planet scope.
The growing world population and increasing weather extremes pose a serious threat to food security. To meet the requirements set by SDG2 Zero Hunger, we need more than ever a timely and global view of crop conditions. To do so, we need a good understanding of where these crops are planted. Satellite remote sensing has nowadays become the most important data source to automate the mapping of agricultural areas and identification of crop types. Existing initiatives are however often characterized by one or more of the following limitations: (i) they provide only a one-time product; (ii) they do not generalize well across space and time; (iii) they do not explicitly account for local growing seasons; (iv) they do not scale well in cloud environments.
Within the ESA WorldCereal project, all these limitations are addressed in order to build the first global classification system that can produce seasonal crop mapping products at field scale within one month after the end of the growing season. The focus products of WorldCereal are: (i) 10m annual cropland extent maps, generated at the end of the major agricultural season; (ii) 10m seasonal wheat and maize maps; (iii) 10m seasonal active cropland maps; (iv) 10m seasonal active irrigation maps.
Making use of the WorldCereal community-based open reference data repository, newly created global crop calendars for summer and winter cereal crops and derived agro-ecological zones (AEZ), a total of more than 170,000 globally distributed reference samples were used to benchmark a variety of classification algorithms with a focus on spatiotemporal robustness and transferability. Inputs consist of Sentinel-1, Sentinel-2, Landsat 8, Copernicus 30m DEM and AgERA5 meteorological variables, subsequently processed into robust and normalized temporal features for further classification. The winning algorithm, a CatBoost classifier, was deployed in a hierarchical setup with a parent model trained on the entire global reference database and subsequent continued training towards local models per AEZ.
While extensive validation is still ongoing, preliminary test results over the priority areas show robust cropland (F1 0.88), wheat (F1 0.80) and maize (F1 0.85) detectors. The next step is to run the system globally for a full year, taking into account the different growing seasons and generating all above mentioned products. Results of this demonstration will be presented during LPS22. Finally, the modular nature of the WorldCereal classification module – which will become available as open-source software – allows to plugin a wide range of classification algorithms, ranging from pixel-based decision trees to convolutional deep neural networks. This ensures the future-proof nature of the system in many diverse applications.