Search

Crop field boundaries aid in mapping crop types, predicting yields, and delivering field-scale analytics to farmers. Recent years have seen the successful application of deep learning to delineating field boundaries in industrial agricultural systems, but field boundary datasets remain missing in smallholder systems due to (1) small fields that require high resolution satellite imagery to delineate and (2) a lack of ground labels for model training and validation. In this work, we combine transfer learning and weak supervision to overcome these challenges, and we demonstrate the methods' success in India where we efficiently generated 10,000 new field labels. Our best model uses 1.5m resolution Airbus SPOT imagery as input, pre-trains a state-of-the-art neural network on France field boundaries, and fine-tunes on India labels to achieve a median Intersection over Union (IoU) of 0.86 in India. If using 4.8m resolution PlanetScope imagery instead, the best model achieves a median IoU of 0.72. Experiments also show that pre-training in France reduces the number of India field labels needed to achieve a given performance level by as much as $20\times$ when datasets are small. These findings suggest our method is a scalable approach for delineating crop fields in regions of the world that currently lack field boundary datasets. We publicly release the 10,000 labels and delineation model to facilitate the creation of field boundary maps and new methods by the community., Comment: Under submission

Forecasting crop yields is important for food security, in particular to predict where crop production is likely to drop. Climate records and remotely-sensed data have become instrumental sources of data for crop yield forecasting systems. Similarly, machine learning methods are increasingly used to process big Earth observation data. However, access to data necessary to train such algorithms is often limited in food-insecure countries. Here, we evaluate the performance of machine learning algorithms and small data to forecast yield on a monthly basis between the start and the end of the growing season. To do so, we developed a robust and automated machine-learning pipeline which selects the best features and model for prediction. Taking Algeria as case study, we predicted national yields for barley, soft wheat and durum wheat with an accuracy of 0.16-0.2 t/ha (13-14 % of mean yield) within the season. The best machine-learning models always outperformed simple benchmark models. This was confirmed in low-yielding years, which is particularly relevant for early warning. Nonetheless, the differences in accuracy between machine learning and benchmark models were not always of practical significance. Besides, the benchmark models outperformed up to 60% of the machine learning models that were tested, which stresses the importance of proper model calibration and selection. For crop yield forecasting, like for many application domains, machine learning has delivered significant improvement in predictive power. Nonetheless, superiority over simple benchmarks is often fully achieved after extensive calibration, especially when dealing with small data., Comment: 17 pages, 7 figures

Change detection, i.e. identification per pixel of changes for some classes of interest from a set of bi-temporal co-registered images, is a fundamental task in the field of remote sensing. It remains challenging due to unrelated forms of change that appear at different times in input images. Here, we propose a reliable deep learning framework for the task of semantic change detection in very high-resolution aerial images. Our framework consists of a new loss function, new attention modules, new feature extraction building blocks, and a new backbone architecture that is tailored for the task of semantic change detection. Specifically, we define a new form of set similarity, that is based on an iterative evaluation of a variant of the Dice coefficient. We use this similarity metric to define a new loss function as well as a new spatial and channel convolution Attention layer (the FracTAL). The new attention layer, designed specifically for vision tasks, is memory efficient, thus suitable for use in all levels of deep convolutional networks. Based on these, we introduce two new efficient self-contained feature extraction convolution units. We validate the performance of these feature extraction building blocks on the CIFAR10 reference data and compare the results with standard ResNet modules. Further, we introduce a new encoder/decoder scheme, a network macro-topology, that is tailored for the task of change detection. Our network moves away from any notion of subtraction of feature layers for identifying change. We validate our approach by showing excellent performance and achieving state of the art score (F1 and Intersection over Union-hereafter IoU) on two building change detection datasets, namely, the LEVIRCD (F1: 0.918, IoU: 0.848) and the WHU (F1: 0.938, IoU: 0.882) datasets., Comment: 28 pages, under review in ISPRS P&RS, 1st revision. Figures of low quality due to compression for arxiv. Reduced abstract in arxiv due to character limitations

Applications of digital agricultural services often require either farmers or their advisers to provide digital records of their field boundaries. Automatic extraction of field boundaries from satellite imagery would reduce the reliance on manual input of these records which is time consuming and error-prone, and would underpin the provision of remote products and services. The lack of current field boundary data sets seems to indicate low uptake of existing methods,presumably because of expensive image preprocessing requirements and local, often arbitrary, tuning. In this paper, we address the problem of field boundary extraction from satellite images as a multitask semantic segmentation problem. We used ResUNet-a, a deep convolutional neural network with a fully connected UNet backbone that features dilated convolutions and conditioned inference, to assign three labels to each pixel: 1) the probability of belonging to a field; 2) the probability of being part of a boundary; and 3) the distance to the closest boundary. These labels can then be combined to obtain closed field boundaries. Using a single composite image from Sentinel-2, the model was highly accurate in mapping field extent, field boundaries, and, consequently, individual fields. Replacing the monthly composite with a single-date image close to the compositing period only marginally decreased accuracy. We then showed in a series of experiments that our model generalised well across resolutions, sensors, space and time without recalibration. Building consensus by averaging model predictions from at least four images acquired across the season is the key to coping with the temporal variations of accuracy. By minimising image preprocessing requirements and replacing local arbitrary decisions by data-driven ones, our approach is expected to facilitate the extraction of individual crop fields at scale.

Reference data collected to validate land cover maps are generally considered free of errors. In practice, however, they contain errors despite all efforts to minimise them. These errors then propagate up to the accuracy assessment stage and impact the validation results. For photo-interpreted reference data, the three most widely studied sources of error are systematic incorrect labelling, vigilance drops, and demographic factors. How internal estimation errors, i.e., errors intrinsic to the response design, affect the accuracy of reference data is far less understood. We analysed the impact of estimation errors for two types of legends as well as for point-based and partition-based response designs with a range of sub-sample sizes. We showed that the accuracy of response designs depends on the class proportions within the sampling units, with complex landscapes being more prone to errors. As a result, response designs where the number of sub-samples are fixed are inefficient, and the labels of reference data sets have inconsistent confidence levels. To control estimation errors, to guarantee high accuracy standards of validation data, and to minimise data collection efforts, we proposed to rely on confidence intervals of the photo-interpreted data to define how many sub-samples should be labelled. In practice, sub-samples are iteratively selected and labelled until the estimated class proportions reach the desired level of confidence. As a result, less effort is spent on labelling obvious cases and the spared effort can be allocated to more complex cases. This approach could reduce the labelling effort by 50% to 75% in our study site, with greater gains in homogeneous landscapes. We contend that adopting this optimisation approach will not only increase the efficiency of reference data collection but will also help deliver reliable accuracy estimates to the user community.

Scene understanding of high resolution aerial images is of great importance for the task of automated monitoring in various remote sensing applications. Due to the large within-class and small between-class variance in pixel values of objects of interest, this remains a challenging task. In recent years, deep convolutional neural networks have started being used in remote sensing applications and demonstrate state of the art performance for pixel level classification of objects. \textcolor{black}{Here we propose a reliable framework for performant results for the task of semantic segmentation of monotemporal very high resolution aerial images. Our framework consists of a novel deep learning architecture, ResUNet-a, and a novel loss function based on the Dice loss. ResUNet-a uses a UNet encoder/decoder backbone, in combination with residual connections, atrous convolutions, pyramid scene parsing pooling and multi-tasking inference. ResUNet-a infers sequentially the boundary of the objects, the distance transform of the segmentation mask, the segmentation mask and a colored reconstruction of the input. Each of the tasks is conditioned on the inference of the previous ones, thus establishing a conditioned relationship between the various tasks, as this is described through the architecture's computation graph. We analyse the performance of several flavours of the Generalized Dice loss for semantic segmentation, and we introduce a novel variant loss function for semantic segmentation of objects that has excellent convergence properties and behaves well even under the presence of highly imbalanced classes.} The performance of our modeling framework is evaluated on the ISPRS 2D Potsdam dataset. Results show state-of-the-art performance with an average F1 score of 92.9\% over all classes for our best model., Comment: Accepted for publication to the ISPRS Journal of Photogrammetry and Remote Sensing