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.

Accurate prediction of crop yields in developing countries in advance of harvest time is central to preventing famine, improving food security, and sustainable development of agriculture. Existing techniques are expensive and difficult to scale as they require locally collected survey data. Approaches utilizing remotely sensed data, such as satellite imagery, potentially provide a cheap, equally effective alternative. Our work shows promising results in predicting soybean crop yields in Argentina using deep learning techniques. We also achieve satisfactory results with a transfer learning approach to predict Brazil soybean harvests with a smaller amount of data. The motivation for transfer learning is that the success of deep learning models is largely dependent on abundant ground truth training data. Successful crop yield prediction with deep learning in regions with little training data relies on the ability to fine-tune pre-trained models.

Our Report draws attention to a complex but understudied issue: How will climate warming alter losses of major food crops to insect pests? Because empirical evidence on plant-insect-climate interactions is scarce and geographically localized, we developed a physiologically based model that incorporates strong and well-established effects of temperature on metabolic rates and on population growth rates. We acknowledged that other factors are involved, but the ones we analyzed are general, robust, and global (1–3).

Parmesan and colleagues argue that our model is overly simplistic and that any general model is premature. They are concerned that our model does not incorporate admittedly idiosyncratic and geographically localized aspects of plant-insect interactions. Some local effects, such as evidence that warmer winters will harm some insects but not others, were in fact evaluated in our sensitivity analyses and shown to be minor (see the Report's Supplementary Materials). Other phenomena, such as plant defenses that benefit some insects and threaten others, are relevant but are neither global nor directional. Furthermore, because Parmesan et al. present no evidence that such idiosyncratic and localized interactions will outweigh the cardinal and universally strong impacts of temperature on populations and on metabolic rates (1–3), their conclusion is subjective.

We agree with Parmesan and colleagues that the question of future crop losses is important and needs further study, that targeted experimental data are needed (as we wrote in our Report), and that our estimates are likely to be conservative (as we concluded, but for reasons different from theirs). However, we strongly disagree with their recommendation to give research priority to gathering localized experimental data. That strategy will only induce a substantial time lag before future crop losses can be addressed.

We draw a lesson from models projecting future climates. Those models lack the “complexity and idiosyncratic nature” of many climate processes, but by building from a few robust principles, they successfully capture the essence of climate patterns and trends (4). Similarly, we hold that the most expeditious and effective way to anticipate crop losses is to develop well-evidenced ecological models and use them to help guide targeted experimental approaches, which can subsequently guide revised ecological models. Experiments and models should be complementary, not sequential.

Reconciling higher freshwater demands with finite freshwater resources remains one of the great policy dilemmas. Given that crop irrigation constitutes 70% of global water extractions, which contributes up to 40% of globally available calories (1), governments often support increases in irrigation efficiency (IE), promoting advanced technologies to improve the “crop per drop.” This provides private benefits to irrigators and is justified, in part, on the premise that increases in IE “save” water for reallocation to other sectors, including cities and the environment. Yet substantial scientific evidence (2) has long shown that increased IE rarely delivers the presumed public-good benefits of increased water availability. Decision-makers typically have not known or understood the importance of basin-scale water accounting or of the behavioral responses of irrigators to subsidies to increase IE. We show that to mitigate global water scarcity, increases in IE must be accompanied by robust water accounting and measurements, a cap on extractions, an assessment of uncertainties, the valuation of trade-offs, and a better understanding of the incentives and behavior of irrigators.

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.