From local to global: A transfer learning-based approach for mapping poplar plantations at national scale using Sentinel-2

Hamrouni, Yousra; Paillassa, Eric; Chéret, Véronique; Monteil, Claude

doi:10.1016/j.isprsjprs.2020.10.018

Cited by 40 publications

(24 citation statements)

References 72 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Active learning improved overall model performance relative to randomized training site selection, in line with findings from two recent efforts (Debats et al, 2017;Hamrouni et al, 2021). Although the relative performance gains that we observed were smaller (e.g., Debats et al, 2017 found 29% higher model performance after one iteration, and 8% higher on the final iterations), those comparisons were made by starting with a training sample that was <1/10 the size of ours.…”

Section: Error Mitigation Featuressupporting

confidence: 90%

“…We couple the labelling platform with a machine learning model inside an active learning (Cohn et al, 1994;Tuia et al, 2011) framework, in which the model is trained interactively, using the model's prediction uncertainty over unlabelled areas to select new sites for additional labelling (Cohn et al, 1994;Tuia et al, 2011). This approach helps boost the performance of the classifier while reducing the overall number of labels required to achieve a given level of performance (Debats et al, 2017;Hamrouni et al, 2021). An unsupervised segmentation step is then applied to convert pixel-wise cropland predictions into vectorized maps of individual field boundaries.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

High Resolution, Annual Maps of Field Boundaries for Smallholder-Dominated Croplands at National Scales

Estes

Song

et al. 2022

Front. Artif. Intell.

View full text Add to dashboard Cite

Mapping the characteristics of Africa’s smallholder-dominated croplands, including the sizes and numbers of fields, can provide critical insights into food security and a range of other socioeconomic and environmental concerns. However, accurately mapping these systems is difficult because there is 1) a spatial and temporal mismatch between satellite sensors and smallholder fields, and 2) a lack of high-quality labels needed to train and assess machine learning classifiers. We developed an approach designed to address these two problems, and used it to map Ghana’s croplands. To overcome the spatio-temporal mismatch, we converted daily, high resolution imagery into two cloud-free composites (the primary growing season and subsequent dry season) covering the 2018 agricultural year, providing a seasonal contrast that helps to improve classification accuracy. To address the problem of label availability, we created a platform that rigorously assesses and minimizes label error, and used it to iteratively train a Random Forests classifier with active learning, which identifies the most informative training sample based on prediction uncertainty. Minimizing label errors improved model F1 scores by up to 25%. Active learning increased F1 scores by an average of 9.1% between first and last training iterations, and 2.3% more than models trained with randomly selected labels. We used the resulting 3.7 m map of cropland probabilities within a segmentation algorithm to delineate crop field boundaries. Using an independent map reference sample (n = 1,207), we found that the cropland probability and field boundary maps had respective overall accuracies of 88 and 86.7%, user’s accuracies for the cropland class of 61.2 and 78.9%, and producer’s accuracies of 67.3 and 58.2%. An unbiased area estimate calculated from the map reference sample indicates that cropland covers 17.1% (15.4–18.9%) of Ghana. Using the most accurate validation labels to correct for biases in the segmented field boundaries map, we estimated that the average size and total number of field in Ghana are 1.73 ha and 1,662,281, respectively. Our results demonstrate an adaptable and transferable approach for developing annual, country-scale maps of crop field boundaries, with several features that effectively mitigate the errors inherent in remote sensing of smallholder-dominated agriculture.

show abstract

Section: Error Mitigation Featuressupporting

confidence: 90%

Section: Introductionmentioning

confidence: 99%

High Resolution, Annual Maps of Field Boundaries for Smallholder-Dominated Croplands at National Scales

Estes

Song

et al. 2022

Front. Artif. Intell.

View full text Add to dashboard Cite

show abstract

“…The active learning approach helped to improve overall model performance relative to randomized training site selection, in line with findings from two recent efforts (Debats et al 2017, Hamrouni et al 2021. Although the performance gains relative to randomized model training that we observed were smaller (e.g.…”

Section: Error Mitigation Featuressupporting

confidence: 89%

High resolution, annual maps of the characteristics of smallholder-dominated croplands at national scales

Estes¹,

Su²,

Song³

et al. 2021

Preprint

View full text Add to dashboard Cite

Understanding agricultural change requires reliable, frequently updated maps that describe the characteristics of croplands. Such data are often unavailable for regions dominated by smallholder agricultural systems, which are particularly challenging for remote sensing. To overcome these challenges, we designed a system to minimize several sources of error that arise when mapping smallholder croplands. To overcome errors caused by mismatches between image resolution and cropland scales, as well as persistent cloud cover, the system converts daily, 3.7 m PlanetScope imagery into two seasonal composites within a single agricultural year. To reduce errors that occur when training classifiers, we built a labelling platform that rigorously assesses label accuracy, and creates more accurate consensus labels that train a Random Forests model. The labelling platform and model interact within an active learning process that boosts the accuracy of the resulting cropland probability map, which is used in a segmentation process to delineate individual field boundaries. We applied this system to map Ghana's croplands for the year 2018. We divided Ghana into 16 mapping regions (12,160-23,535 km 2), training separate models for each using a total of 6,299 labels, plus 1,600 for validation. Using an independent map reference sample (n=1,207), we found that overall accuracies of the resulting cropland probability and field boundary maps were 88% and 86.7%, respectively, with User's accuracies for the cropland class of 61.2% and 78.9%, and Producer's accuracies of 67.3% and 58.2%. Croplands covered 16.1-23.2% of the mapped area, comprising 1,131,146 total fields with an average size of 3.92 ha. Estimates based on the map reference sample indicate the cropland percentage is 17.1% (15.4-18.9%) or 17.6% (15.6-19.6%), depending on the map used to estimate the standard error. Using the labellers' digitized field boundaries to estimate biases in field boundary statistics, we calculated an adjusted mean field size of 1.73 ha and total field count of 1,662,281. Although the cropland class contained substantial errors, the system was effective in mitigating error and quantifying resulting performance gains. By minimizing training errors, consensus labelling improved the model's F1 scores by up to 25%, while 3 iterations of active learning increased the F1 score by 9.1%, on average, which was 2.3% higher than training models with randomly selected labels. Map accuracy can be improved by replacing Random Forests with a convolutional neural network. These results demonstrate a readily adapted, transferrable framework for developing high resolution, annual, nation-scale maps that provide important details about smallholder-dominated croplands.

show abstract

“…Therefore, the Simpson index (" r 2 = 0.67) is more useful for monitoring and prediction using satellite sensors on a large geographical scale. Heterogeneity indices (i.e., WSLR, growth index and texture diversity) measures also showed an " r 2 > 0.35, which is promising despite the fact that their links to remote sensing signals are not as clear as those for species diversity [9].…”

Section: Large-area Plant Diversity Indices Spatial Distributionmentioning

confidence: 91%

“…Therefore, large-area wall-to-wall maps on plant diversity in forests, as well as spatial changes in its distribution, are frequently required in the forestry operations and conservation of relevant sites [7,8]. Such spatially explicit maps are vital for ecological function stability [5,9,10], exotic species invasion assessments [11], and prevention of plant diseases and insect pests [3,12]-particularly in maintaining the stability of plant community composition [2,6,10,13].…”

Section: Introductionmentioning

confidence: 99%

Mapping Plant Diversity Based on Combined SENTINEL-1/2 Data—Opportunities for Subtropical Mountainous Forests

et al. 2022

View full text Add to dashboard Cite

Plant diversity is an important parameter in maintaining forest ecosystem services, functions and stability. Timely and accurate monitoring and evaluation of large-area wall-to-wall maps on plant diversity and its spatial heterogeneity are crucial for the conservation and management of forest resources. However, traditional botanical field surveys designed to estimate plant diversity are usually limited in their spatiotemporal resolutions. Using Sentinel-1 (S-1) and Sentinel-2 (S-2) data at high spatiotemporal scales, combined with and referenced to botanical field surveys, may be the best choice to provide accurate plant diversity distribution information over a large area. In this paper, we predicted and mapped plant diversity in a subtropical forest using 24 months of freely and openly available S-1 and S-2 images (10 m × 10 m) data over a large study area (15,290 km2). A total of 448 quadrats (10 m × 10 m) of forestry field surveys were captured in a subtropical evergreen-deciduous broad-leaved mixed forest to validate a machine learning algorithm. The objective was to link the fine Sentinel spectral and radar data to several ground-truthing plant diversity indices in the forests. The results showed that: (1) The Simpson and Shannon-Wiener diversity indices were the best predicted indices using random forest regression, with ȓ2 of around 0.65; (2) The use of S-1 radar data can enhance the accuracy of the predicted heterogeneity indices in the forests by approximately 0.2; (3) As for the mapping of Simpson and Shannon-Wiener, the overall accuracy was 67.4% and 64.2% respectively, while the texture diversity’s overall accuracy was merely 56.8%; (4) From the evaluation and prediction map information, the Simpson, Shannon-Wiener and texture diversity values (and its confidence interval values) indicate spatial heterogeneity in pixel level. The large-area forest plant diversity indices maps add spatially explicit information to the ground-truthing data. Based on the results, we conclude that using the time-series of S-1 and S-2 radar and spectral characteristics, when coupled with limited ground-truthing data, can provide reasonable assessments of plant spatial heterogeneity and diversity across wide areas. It could also help promote forest ecosystem and resource conservation activities in the forestry sector.

show abstract

From local to global: A transfer learning-based approach for mapping poplar plantations at national scale using Sentinel-2

Cited by 40 publications

References 72 publications

High Resolution, Annual Maps of Field Boundaries for Smallholder-Dominated Croplands at National Scales

High Resolution, Annual Maps of Field Boundaries for Smallholder-Dominated Croplands at National Scales

High resolution, annual maps of the characteristics of smallholder-dominated croplands at national scales

Mapping Plant Diversity Based on Combined SENTINEL-1/2 Data—Opportunities for Subtropical Mountainous Forests

Contact Info

Product

Resources

About