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.
Geospatial analysis lacks methods like the word vector representations and pre-trained networks that significantly boost performance across a wide range of natural language and computer vision tasks. To fill this gap, we introduce Tile2Vec, an unsupervised representation learning algorithm that extends the distributional hypothesis from natural language — words appearing in similar contexts tend to have similar meanings — to spatially distributed data. We demonstrate empirically that Tile2Vec learns semantically meaningful representations for both image and non-image datasets. Our learned representations significantly improve performance in downstream classification tasks and, similarly to word vectors, allow visual analogies to be obtained via simple arithmetic in the latent space.
Accurate measurements of crop production in smallholder farming systems are critical to the understanding of yield constraints and, thus, setting the appropriate agronomic investments and policies for improving food security and reducing poverty. Nevertheless, mapping the yields of smallholder farms is challenging because of factors such as small field sizes and heterogeneous landscapes. Recent advances in fine-resolution satellite sensors offer promise for monitoring and characterizing the production of smallholder farms. In this study, we investigated the utility of different sensors, including the commercial Skysat and RapidEye satellites and the publicly accessible Sentinel-2, for tracking smallholder maize yield variation throughout a ~40,000 km2 western Kenya region. We tested the potential of two types of multiple regression models for predicting yield: (i) a “calibrated model”, which required ground-measured yield and weather data for calibration, and (ii) an “uncalibrated model”, which used a process-based crop model to generate daily vegetation index and end-of-season biomass and/or yield as pseudo training samples. Model performance was evaluated at the field, division, and district scales using a combination of farmer surveys and crop cuts across thousands of smallholder plots in western Kenya. Results show that the “calibrated” approach captured a significant fraction (R2 between 0.3 and 0.6) of yield variations at aggregated administrative units (e.g., districts and divisions), while the “uncalibrated” approach performed only slightly worse. For both approaches, we found that predictions using the MERIS Terrestrial Chlorophyll Index (MTCI), which included the red edge band available in RapidEye and Sentinel-2, were superior to those made using other commonly used vegetation indices. We also found that multiple refinements to the crop simulation procedures led to improvements in the “uncalibrated” approach. We identified the prevalence of small field sizes, intercropping management, and cloudy satellite images as major challenges to improve the model performance. Overall, this study suggested that high-resolution satellite imagery can be used to map yields of smallholder farming systems, and the methodology presented in this study could serve as a good foundation for other smallholder farming systems in the world.