Accurate automated segmentation of remote sensing data could benefit applications from land cover mapping and agricultural monitoring to urban development surveyal and disaster damage assessment. While convolutional neural networks (CNNs) achieve state-of-the-art accuracy when segmenting natural images with huge labeled datasets, their successful translation to remote sensing tasks has been limited by low quantities of ground truth labels, especially fully segmented ones, in the remote sensing domain. In this work, we perform cropland segmentation using two types of labels commonly found in remote sensing datasets that can be considered sources of “weak supervision”: (1) labels comprised of single geotagged points and (2) image-level labels. We demonstrate that (1) a U-Net trained on a single labeled pixel per image and (2) a U-Net image classifier transferred to segmentation can outperform pixel-level algorithms such as logistic regression, support vector machine, and random forest. While the high performance of neural networks is well-established for large datasets, our experiments indicate that U-Nets trained on weak labels outperform baseline methods with as few as 100 labels. Neural networks, therefore, can combine superior classification performance with efficient label usage, and allow pixel-level labels to be obtained from image labels.

Feeding a growing population while reducing negative environmental impacts is one of the greatest challenges of the coming decades. We show that microsatellite data can be used to detect the impact of sustainable intensification interventions at large scales and to target the fields that would benefit the most, thereby doubling yield gains. Our work reveals that satellite data provide a scalable approach to sustainably increase food production.

Wheat is the most important staple crop grown in Australia, and Australia is one of the top wheat exporting countries globally. Timely and reliable wheat yield prediction in Australia is important for regional and global food security. Prior studies use either climate data, or satellite data, or a combination of these two to build empirical models to predict crop yield. However, though the performance of yield prediction using empirical methods is improved by combining the use of climate and satellite data, the contributions from different data sources are still not clear. In addition, how the regression-based methods compare with various machine-learning based methods in their performance in yield prediction is also not well understood and needs in-depth investigation. This work integrated various sources of data to predict wheat yield across Australia from 2000 to 2014 at the statistical division (SD) level. We adopted a well-known regression method (LASSO, as a benchmark) and three mainstream machine learning methods (support vector machine, random forest, and neural network) to build various empirical models for yield prediction. For satellite data, we used the enhanced vegetation index (EVI) from MODIS and solar-induced chlorophyll fluorescence (SIF) from GOME-2 and SCIAMACHY as metrics to approximate crop productivity. The machine-learning based methods outperform the regression method in modeling crop yield. Our results confirm that combining climate and satellite data can achieve high performance of yield prediction at the SD level (R2 ˜ 0.75). The satellite data track crop growth condition and gradually capture the variability of yield evolving with the growing season, and their contributions to yield prediction usually saturate at the peak of the growing season. Climate data provide extra and unique information beyond what the satellite data have offered for yield prediction, and our empirical modeling work shows the added values of climate variables exist across the whole season, not only at some certain stages. We also find that using EVI as an input can achieve better performance in yield prediction than SIF, primarily due to the large noise in the satellite-based SIF data (i.e. coarse resolution in both space and time). In addition, we also explored the potential for timely wheat yield prediction in Australia, and we can achieve the optimal prediction performance with approximately two-month lead time before wheat maturity. The proposed methodology in this paper can be extended to different crops and different regions for crop yield prediction.

Accurate measurements of maize yields at field or subfield scales are useful for guiding agronomic practices and investments and policies for improving food security. Data on smallholder maize systems are currently sparse, but satellite remote sensing offers promise for accelerating learning about these systems. Here we document the use of Google Earth Engine (GEE) to build “wall-to-wall” 10 m resolution maps of (i) cropland presence, (ii) maize presence, and (iii) maize yields for the main 2017 maize season in Kenya and Tanzania. Mapping these outcomes at this scale is extremely challenging because of very heterogeneous landscapes, lack of cloud-free satellite imagery, and the low quantity of quality ground-based data in these regions.

First, we computed seasonal median composites of Sentinel-1 radar backscatter and Sentinel-2 optical reflectance measures for each pixel in the region, and used them to build both crop/non-crop and maize/non-maize Random Forest (RF) classifiers. Several thousand crop/non-crop labels were collected through an in-house GEE labeler, and thousands of crop type labels from the 2015–2017 growing seasons were obtained from various sources. Results show that the crop/non-crop classifier successfully identified cropland with over 85% out-of-sample accuracy in both countries, with Sentinel-1 being particularly useful for prediction. Among the cropped pixels, the maize/non-maize classier had an accuracy of 79% in Tanzania and 63% in Kenya.

To map maize yields, we build on past work using a scalable crop yield mapper (SCYM) that utilizes simulations from a crop model to train a regression that predicts yields from observations. Here we advance past approaches by (i) grouping simulations by Global Agro-Environmental Stratification (GAES) zones across the two countries, in order to account for landscape heterogeneity, (ii) utilizing gridded datasets on soil and sowing and harvest dates to setup model simulations in a scalable way; and (iii) utilizing all available satellite observations during the growing season in a parsimonious way by using harmonic regression fits implemented in GEE. SCYM estimates were able to capture about 50% of the variation in the yields at the district level in Western Kenya as measured by objective ground-based crop cuts.

Finally, we illustrated the utility of our yield maps with two case studies. First, we document the magnitude and interannual variability of spatial heterogeneity of yields in each district, and how it varies for different parts of the region. Second, we combine our estimates with recently released soil databases in the region to investigate the most important soil constraints in the region. Soil factors explain a high fraction (72%) of variation in predicted yields, with the predominant factor being soil nitrogen levels. Overall, this study illustrates the power of combining Sentinel-1 and Sentinel-2 imagery, the GEE platform, and advanced classification and yield mapping algorithms to advance understanding of smallholder agricultural systems.

Low-intensity tillage has become more popular among farmers in the United States and many other regions. However, accurate data on when and where low-intensity tillage methods are being used remain scarce, and this scarcity impedes understanding of the factors affecting the adoption and the agronomic or environmental impacts of these practices. In this study, we used composites of satellite imagery from Landsat 5, 7, and 8, and Sentinel-1 in combination with producer data from about 5900 georeferenced fields to train a random forest classifier and generate annual large-scale maps of tillage intensity from 2005 to 2016. We tested different combinations of hyper-parameters using cross-validation, splitting the training and testing data alternatively by field, year, and state to assess the influence of clustering on validation results and evaluate the generalizability of the classification model. We found that the best model was able to map tillage practices across the entire North Central US region at 30 m-resolution with accuracies spanning between 75% and 79%, depending on the validation approach. We also found that although Sentinel-1 provides an independent measure that should be sensitive to surface moisture and roughness, it currently adds relatively little to classification performance beyond what is possible with Landsat. When aggregated to the state level, the satellite estimates of percentage low- and high-intensity tillage agreed well with a USDA survey on tillage practices in 2006 (R2 = 0.55). The satellite data also revealed clear increases in low-intensity tillage area for most counties in the past decade. Overall, the ability to accurately map spatial and temporal patterns in tillage should facilitate further study of this important practice in the United States, as well as other regions with fewer survey-based estimates.

Fast, accurate and inexpensive estimates of crop yields at the field scale are useful for many applications. Based on the Google Earth Engine (GEE) platform, we recently developed a Scalable satellite-based Crop Yield Mapper (SCYM) that integrates crop simulations with satellite imagery and gridded weather data to generate 30 m resolution yield estimates for multiple crops in different regions. Existing versions of SCYM typically capture one-third to half of the variation in reported county-scale yields. Using rainfed maize in the US Midwest as an example, this study tested multiple approaches for improving SCYM’s accuracy, including (i) calibrating the phenology parameters of the crop model (APSIM) used to generate training samples for SCYM; (ii) using an ensemble of three crop models (APSIM-Maize, CERES-Maize, and Hybrid-Maize) instead of a single model; (iii) using simulated biomass from the crop models instead of simulated yields to train SCYM, with the former assuming a constant harvest index (HI). Results show substantial improvement in performance, as assessed using reported county yields by USDA-NASS, both from calibrating APSIM phenology parameters and from training SCYM on simulated biomass rather than yields. Using a multi-model ensemble further improves SCYM, although the benefit is limited. The proposed preferred version of SCYM on average captures 75% of the yield variation for 2001–2015 in the 3I states (i.e. Illinois, Indiana and Iowa) where SCYM is trained, with RMSE typically less than 1 t/ha, and explains 41% to 83% of multi-year yield variations when tested across nine Midwestern US states for 2008–2015. This level of accuracy is particularly notable given that only data from 2014 were used to calibrate phenology parameters. The yield estimates for multiple years in multiple states utilized 1184 Landsat tiles, but could be completed in about 2 h per year by using the GEE platform. All approaches tested in this study do not require any site-specific measurements, and thus can be readily extended to other regions and crops.