Cropland Mapping over Sahelian and Sudanian Agrosystems: A Knowledge-Based Approach Using PROBA-V Time Series at 100-m

Large-area mapping of active cropland and short-term fallows in smallholder landscapes using PlanetScope data

Cropland mapping in smallholder landscapes is challenged by complex and fragmented landscapes, labor-intensive and unmechanized land management causing high within-field variability, rapid dynamics in shifting cultivation systems, and substantial proportions of short-term fallows. To overcome these challenges, we here present a large-area mapping framework to identify active cropland and short-term fallows in smallholder landscapes for the 2020/2021 growing season at 4.77 m spatial resolution. Our study focuses on Northern Mozambique, an area comprising 381,698 km2. The approach is based on Google Earth Engine and time series of PlanetScope mosaics made openly available through Norwaýs International Climate and Forest Initiative (NICFI) data program. We conducted multi-temporal coregistration of the PlanetScope data using seasonal Sentinel-2 base images and derived consistent and gap-free seasonal time series metrics to classify active cropland and short-term fallows. An iterative active learning framework based on Random Forest class probabilities was used for training rare classes and uncertain regions. The map was accurate (area-adjusted overall accuracy 88.6% ± 1.5%), with the main error type being the commission of active cropland. Error-adjusted area estimates of active cropland extent (61,799.5 km2 ± 4,252.5 km2) revealed that existing global and regional land cover products tend to under-, or over-estimate active cropland extent, respectively. Short-term fallows occupied 28.9% of the cropland in our reference sample (13% of the mapped cropland), with consolidated agricultural regions showing the highest shares of short-term fallows. Our approach relies on openly available PlanetScope data and cloud-based processing in Google Earth Engine, which minimizes financial constraints and maximizes replicability of the methods. All code and maps were made available for further use.

Thirty Years of Land Cover and Fraction Cover Changes over the Sudano-Sahel Using Landsat Time Series

Historical land cover maps are of high importance for scientists and policy makers studying the dynamic character of land cover change in the Sudano-Sahel, including anthropogenic and climatological drivers. Despite its relevance, an accurate high resolution record of historical land cover maps is currently lacking over the Sudano-Sahel. In this study, 30 m resolution historically consistent land cover and cover fraction maps are provided over the Sudano-Sahel for the period 1986–2015. These land cover/cover fraction maps are achieved based on the Landsat archive preprocessed on Google Earth Engine and a random forest classification/regression model, while historical consistency is achieved using the hidden Markov model. Using these historical maps, a multitude of variability in the dynamic Sudano-Sahel region over the past 30 years is revealed. On the one hand, Sahel-wide cropland expansion and the re-greening of the Sahel is observed in the discrete land cover classification. On the other hand, subtle changes such as forest degradation are detected based on the cover fraction maps. Additionally, exploiting the 30 m spatial resolution, fine-scale changes, such as smallholder or subsistence farming, can be detected. The historical land cover/cover fraction maps presented in this study are made available via an open-access platform.

Comparison of Different Cropland Classification Methods under Diversified Agroecological Conditions in the Zambezi River Basin

Having updated knowledge of cropland extent is essential for crop monitoring and food security early warning. Previous research has proposed different methods and adopted various datasets for mapping cropland areas at regional to global scales. However, most approaches did not consider the characteristics of farming systems and apply the same classification method in different agroecological zones (AEZs). Furthermore, the acquisition of in situ samples for classification training remains challenging. To address these knowledge gaps and challenges, this study applied a zone-specific classification by comparing four classifiers (random forest, the support vector machine (SVM), the classification and regression tree (CART) and minimum distance) for cropland mapping over four different AEZs in the Zambezi River basin (ZRB). Landsat-8 and Sentinel-2 data and derived indices were used and synthesized to generate thirty-five layers for classification on the Google Earth Engine platform. Training samples were derived from three existing landcover datasets to minimize the cost of sample acquisitions over the large area. The final cropland map was generated at a 10 m resolution. The performance of the four classifiers and the viability of training samples were analysed. All classifiers presented higher accuracy in cool AEZs than in warm AEZs, which may be attributed to field size and lower confusion between cropland and grassland classes. This indicates that agricultural landscape may impact classification results regardless of the classifiers. Random forest was found to be the most stable and accurate classifier across different agricultural systems, with an overall accuracy of 84% and a kappa coefficient of 0.67. Samples extracted over the full agreement areas among existing datasets reduced uncertainty and provided reliable calibration sets as a replacement of costly in situ measurements. The methodology proposed by this study can be used to generate periodical high-resolution cropland maps in ZRB, which is helpful for the analysis of cropland extension and abandonment as well as intensity changes in response to the escalating population and food insecurity.

A High-Resolution Cropland Map for the West African Sahel Based on High-Density Training Data, Google Earth Engine, and Locally Optimized Machine Learning

The West African Sahel Cropland map (WASC30) is a new 30-m cropland extent product for the nominal year of 2015. We used the computing resources provided by Google Earth Engine (GEE) to fit and apply Random Forest models for cropland detection in each of 189 grid cells (composed of 100 km2, hence a total of ~1.9 × 106 km2) across five countries of the West African Sahel (Burkina Faso, Mauritania, Mali, Niger, and Senegal). Landsat-8 surface reflectance (Bands 2–7) and vegetation indices (NDVI, EVI, SAVI, and MSAVI), organized to include dry-season and growing-season band reflectances and vegetation indices for the years 2013–2015, were used as predictors. Training data were derived from an independent, high-resolution, visually interpreted sample dataset that classifies sample points across West Africa using a 2-km grid (~380,000 points were used in this study, with 50% used for model training and 50% used for model validation). Analysis of the new cropland dataset indicates a summed cropland area of ~316 × 103 km2 across the 5 countries, primarily in rainfed cropland (309 × 103 km2), with irrigated cropland area (7 × 103 km2) representing 2% of the total cropland area. At regional scale, the cropland dataset has an overall accuracy of 90.1% and a cropland class (rainfed and irrigated) user’s accuracy of 79%. At bioclimatic zones scale, results show that land proportion occupied by rainfed agriculture increases with annual precipitation up to 1000 mm. The Sudanian zone (600–1200 mm) has the highest proportion of land in agriculture (24%), followed by the Sahelian (200–600 mm) and the Guinean (1200 +) zones for 15% and 4%, respectively. The new West African Sahel dataset is made freely available for applications requiring improved cropland area information for agricultural monitoring and food security applications.

Vegetation Fraction Images Derived from PROBA-V Data for Rapid Assessment of Annual Croplands in Brazil

This paper presents a new method for rapid assessment of the extent of annual croplands in Brazil. The proposed method applies a linear spectral mixing model (LSMM) to PROBA-V time series images to derive vegetation, soil, and shade fraction images for regional analysis. We used S10-TOC (10 days synthesis, 1 km spatial resolution, and top-of-canopy) products for Brazil and S5-TOC (five days synthesis, 100 m spatial resolution, and top-of-canopy) products for Mato Grosso State (Brazilian Legal Amazon). Using the time series of the vegetation fraction images of the whole year (2015 in this case), only one mosaic composed with maximum values of vegetation fraction was generated, allowing detecting and mapping semi-automatically the areas occupied by annual crops during the year. The results (100 m spatial resolution map) for the Mato Grosso State were compared with existing global datasets (Finer Resolution Observation and Monitoring—Global Land Cover (FROM-GLC) and Global Food Security—Support Analyses Data (GFSAD30)). Visually those maps present a good agreement, but the area estimated are not comparable since the agricultural class definition are different for those maps. In addition, we found 11.8 million ha of agricultural areas in the entire Brazilian territory. The area estimation for the Mato Grosso State was 3.4 million ha for 1 km dataset and 5.3 million ha for 100 m dataset. This difference is due to the spatial resolution of the PROBA-V datasets used. A coefficient of determination of 0.82 was found between PROBA-V 100 m and Landsat-8 OLI area estimations for the Mato Grosso State. Therefore, the proposed method is suitable for detecting and mapping annual croplands distribution operationally using PROBA-V datasets for regional analysis.

Automatic Proba-V Processor: TREX—Tool for Raster Data Exploration

The processing tool TREX, standing for ‘Tool for Raster data EXploration’ is presented and evaluated in the Biebrza wetlands in northeastern Poland. TREX was designed for the automatization of processing satellite data from the Proba-V satellite into maps of NDVI or LAI in any defined by the user projection, spatial resolution, or extent. The open source and access concept of TREX encourages the potential community of users to collaborate, develop, and integrate the tool with other satellite imagery and models. TREX reprojects, shifts, and resamples original data obtained from the Proba-V satellite to deliver reliable maps of NDVI and LAI. Validation of TREX in Biebrza wetlands resulted in correlations between 0.79 and 0.92 for NDVI data (measured with ASD Field Spec 4) and 0.92 for LAI data (measured with LiCOR—LAI-2000 Plant Canopy Analyzer).

Sub-Pixel Crop Type Classification Using PROBA-V 100 m NDVI Time Series and Reference Data from Sentinel-2 Classifications

This paper presents the results of a sub-pixel classification of crop types in Bulgaria from PROBA-V 100 m normalized difference vegetation index (NDVI) time series. Two sub-pixel classification methods, artificial neural network (ANN) and support vector regression (SVR) were used where the output was a set of area fraction images (AFIs) at 100 m resolution with pixels containing estimated area fractions of each class. High-resolution maps of two test sites derived from Sentinel-2 classifications were used to obtain training data for the sub-pixel classifications. The estimated area fractions have a good correspondence with the true area fractions when aggregated to regions of 10 × 10 km2, especially when the SVR method was used. For the five dominant classes in the test sites the R2 obtained after the aggregation was 86% (winter cereals), 81% (sunflower), 92% (broad-leaved forest), 89% (maize), and 67% (grasslands) when the SVR method was used.

Evaluation of PROBA-V Collection 1: Refined Radiometry, Geometry, and Cloud Screening

PROBA-V (PRoject for On-Board Autonomy–Vegetation) was launched in May-2013 as an operational continuation to the vegetation (VGT) instruments on-board the Système Pour l’Observation de la Terre (SPOT)-4 and -5 satellites. The first reprocessing campaign of the PROBA-V archive from Collection 0 (C0) to Collection 1 (C1) aims at harmonizing the time series, thanks to improved radiometric and geometric calibration and cloud detection. The evaluation of PROBA-V C1 focuses on (i) qualitative and quantitative assessment of the new cloud detection scheme; (ii) quantification of the effect of the reprocessing by comparing C1 to C0; and (iii) evaluation of the spatio-temporal stability of the combined SPOT/VGT and PROBA-V archive through comparison to METOP/advanced very high resolution radiometer (AVHRR). The PROBA-V C1 cloud detection algorithm yields an overall accuracy of 89.0%. Clouds are detected with very few omission errors, but there is an overdetection of clouds over bright surfaces. Stepwise updates to the visible and near infrared (VNIR) absolute calibration in C0 and the application of degradation models to the SWIR calibration in C1 result in sudden changes between C0 and C1 Blue, Red, and NIR TOC reflectance in the first year, and more gradual differences for short-wave infrared (SWIR). Other changes result in some bias between C0 and C1, although the root mean squared difference (RMSD) remains well below 1% for top-of-canopy (TOC) reflectance and below 0.02 for the normalized difference vegetation index (NDVI). Comparison to METOP/AVHRR shows that the recent reprocessing campaigns on SPOT/VGT and PROBA-V have resulted in a more stable combined time series.

A Phenology-Based Method to Map Cropping Patterns under a Wheat-Maize Rotation Using Remotely Sensed Time-Series Data

Agricultural land use and cropping patterns are closely related to food production, soil degradation, water resource management, greenhouse gas emission, and regional climate alterations. Methods for reliable and cost-efficient mapping of cropping pattern, as well as their changes over space and time, are therefore urgently needed. To cope with this need, we developed a phenology-based method to map cropping patterns based on time-series of vegetation index data. The proposed method builds on the well-known ‘threshold model’ to retrieve phenological metrics. Values of four phenological parameters are used to identify crop seasons. Using a set of rules, the crop season information is translated into cropping pattern. To illustrate the method, cropping patterns were determined for three consecutive years (2008–2010) in the Henan province of China, where reliable validation data was available. Cropping patterns were derived using eight-day composite MODIS Enhanced Vegetation Index (EVI) data. Results show that the proposed method can achieve a satisfactory overall accuracy (~84%) in extracting cropping patterns. Interestingly, the accuracy obtained with our method based on MODIS EVI data was comparable with that from Landsat-5 TM image classification. We conclude that the proposed method for cropland and cropping pattern identification based on MODIS data offers a simple, yet reliable way to derive important land use information over large areas.

"Reduction of tree cover in West African woodlands and promotion in semi-arid farmlands"

Woody vegetation in farmland acts as a carbon sink and provides ecosystem services for local people, but no macro-scale assessments of the impact of management and climate on woody cover exists for drylands. Here we make use of very high spatial resolution satellite imagery to derive wall-to-wall woody cover patterns in tropical West African drylands. Our study reveals a consistently high woody cover in farmlands along all semi-arid and sub-humid rainfall zones (16%), on average only 6% lower than in savannas. In semi-arid Sahel, farmland management increases woody cover to a greater level (12%) than found in neighbouring savannas (6%), whereas farmlands in sub-humid zones have a reduced woody cover (20%) as compared to savannas (30%). In the region as a whole, rainfall, terrain and soil are the most important (80%) determinants of woody cover, while management factors play a smaller (20%) role. We conclude that agricultural expansion cannot generally be claimed to cause woody cover losses, and that observations in Sahel contradict simplistic ideas of a high negative correlation between population density and woody cover.

National-scale cropland mapping based on spectral-temporal features and outdated land cover information

The lack of sufficient ground truth data has always constrained supervised learning, thereby hindering the generation of up-to-date satellite-derived thematic maps. This is all the more true for those applications requiring frequent updates over large areas such as cropland mapping. Therefore, we present a method enabling the automated production of spatially consistent cropland maps at the national scale, based on spectral-temporal features and outdated land cover information. Following an unsupervised approach, this method extracts reliable calibration pixels based on their labels in the outdated map and their spectral signatures. To ensure spatial consistency and coherence in the map, we first propose to generate seamless input images by normalizing the time series and deriving spectral-temporal features that target salient cropland characteristics. Second, we reduce the spatial variability of the class signatures by stratifying the country and by classifying each stratum independently. Finally, we remove speckle with a weighted majority filter accounting for per-pixel classification confidence. Capitalizing on a wall-to-wall validation data set, the method was tested in South Africa using a 16-year old land cover map and multi-sensor Landsat time series. The overall accuracy of the resulting cropland map reached 92%. A spatially explicit validation revealed large variations across the country and suggests that intensive grain-growing areas were better characterized than smallholder farming systems. Informative features in the classification process vary from one stratum to another but features targeting the minimum of vegetation as well as short-wave infrared features were consistently important throughout the country. Overall, the approach showed potential for routinely delivering consistent cropland maps over large areas as required for operational crop monitoring.

Combining global land cover datasets to quantify agricultural expansion into forests in Latin America: Limitations and challenges

While we know that deforestation in the tropics is increasingly driven by commercial agriculture, most tropical countries still lack recent and spatially-explicit assessments of the relative importance of pasture and cropland expansion in causing forest loss. Here we present a spatially explicit quantification of the extent to which cultivated land and grassland expanded at the expense of forests across Latin America in 2001-2011, by combining two "state-of-the-art" global datasets (Global Forest Change forest loss and GlobeLand30-2010 land cover). We further evaluate some of the limitations and challenges in doing this. We find that this approach does capture some of the major patterns of land cover following deforestation, with GlobeLand30-2010's Grassland class (which we interpret as pasture) being the most common land cover replacing forests across Latin America. However, our analysis also reveals some major limitations to combining these land cover datasets for quantifying pasture and cropland expansion into forest. First, a simple one-to-one translation between GlobeLand30-2010's Cultivated land and Grassland classes into cropland and pasture respectively, should not be made without caution, as GlobeLand30-2010 defines its Cultivated land to include some pastures. Comparisons with the TerraClass dataset over the Brazilian Amazon and with previous literature indicates that Cultivated land in GlobeLand30-2010 includes notable amounts of pasture and other vegetation (e.g. in Paraguay and the Brazilian Amazon). This further suggests that the approach taken here generally leads to an underestimation (of up to ~60%) of the role of pasture in replacing forest. Second, a large share (~33%) of the Global Forest Change forest loss is found to still be forest according to GlobeLand30-2010 and our analysis suggests that the accuracy of the combined datasets, especially for areas with heterogeneous land cover and/or small-scale forest loss, is still too poor for deriving accurate quantifications of land cover following forest loss.

Crop Phenology Detection Using High Spatio-Temporal Resolution Data Fused from SPOT5 and MODIS Products

Timely and efficient monitoring of crop phenology at a high spatial resolution are crucial for the precise and effective management of agriculture. Recently, satellite-derived vegetation indices (VIs), such as the Normalized Difference Vegetation Index (NDVI), have been widely used for the phenology detection of terrestrial ecosystems. In this paper, a framework is proposed to detect crop phenology using high spatio-temporal resolution data fused from Systeme Probatoire d'Observation de la Tarre5 (SPOT5) and Moderate Resolution Imaging Spectroradiometer (MODIS) images. The framework consists of a data fusion method to produce a synthetic NDVI dataset at SPOT5’s spatial resolution and at MODIS’s temporal resolution and a phenology extraction algorithm based on NDVI time-series analysis. The feasibility of our phenology detection approach was evaluated at the county scale in Shandong Province, China. The results show that (1) the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) algorithm can accurately blend SPOT5 and MODIS NDVI, with an R² of greater than 0.69 and an root mean square error (RMSE) of less than 0.11 between the predicted and referenced data; and that (2) the estimated phenology parameters, such as the start and end of season (SOS and EOS), were closely correlated with the field-observed data with an R² of the SOS ranging from 0.68 to 0.86 and with an R² of the EOS ranging from 0.72 to 0.79. Our research provides a reliable approach for crop phenology mapping in areas with high fragmented farmland, which is meaningful for the implementation of precision agriculture.