Document Type

Dissertation - Open Access

Award Date

2021

Degree Name

Doctor of Philosophy (PhD)

Department

Geospatial Science and Engineering

First Advisor

Niall Hanan

Keywords

Agricultural land area, Crop type mapping, Food security, Machine Learning Sahel, West Africa

Abstract

In arid and semi-arid West Africa, agricultural production and regional food security depend largely on small-scale subsistence farming and rainfed crops, both of which are vulnerable to climate variability and drought. Efforts made to improve crop monitoring and our ability to estimate crop production (areas planted and yield estimations by crop type) in the major agricultural zones of the region are critical paths for minimizing climate risks and to support food security planning. The main objective of this dissertation research was to contribute to these efforts using remote sensing technologies. In this regard, the first analysis documented the low reliability of existing land cover products for cropland area estimation (Chapter 2). Then two satellite remote sensing-based datasets were developed that 1) accurately map cropland areas in the five countries of Sahelian West Africa (Senegal, Mauritania, Mali, Burkina Faso and Niger; Chapter 3), and 2) focus on the country of Mali to identify the location and prevalence of the major subsistence crops (millet, sorghum, maize and non-irrigated rice; Chapter 4). The regional cropland area product is distributed as the West African Sahel Cropland area at 30 m (WASC30). The development of the new dataset involved high density training data (380,000 samples) developed by USGS in collaboration with CILSS for training about 200 locally optimized random forest (RF) classifiers using Landsat 8 surface reflectances and vegetation indices and the Google Earth Engine platform. WASC30 greatly improves earlier estimates through inclusion of cropland information for both rainfed and irrigated areas mapped with a class-specific accuracy of 79% across the West Africa Sahel. Used as a mask in crop monitoring systems, the new cropland area data could bring critical insights by reducing uncertainties in xv identification of croplands as crop growth condition metrics are extracted. WASC30 allowed us to derive detailed statistics on cultivated areas in the Sahel, at country and agroclimatic scales. Intensive agricultural zones were highlighted as well. The second dataset, mapping crop types for the country of Mali, is meant to separate signals of different crop types for improved crop yield estimation. The crop type map was used to derive detailed agricultural statistics (e.g. acreage by crop types, spatial distribution) at finer administrative scales than has previously been possible. The crop fraction information by crop type extracted from the map, gives additional details on farmers preferences by regions, and the natural adaptability of different crop types. The final analysis of this dissertation explores the use of ensemble machine learning techniques to predict maize yield in Mali (Chapter 5). Climate data (precipitation and temperature), and vegetation indices (Normalized Difference Vegetation Index, NDVI, the Enhanced Vegetation Index, EVI, and the Normalized Difference Water Index, NDWI) are used as predictors, while actual yields collected in 2017 by the Malian Ministry of Agriculture are the reference data. Random forest presented better predictive performance as compared to boosted regression trees (BRT). Results showed that climate variables have more predictive power for maize yield compared to vegetation indices. Among vegetation indices, the NDWI appeared to be the most influential predictor, maybe because of water requirement of maize and the sensitivity of this index to water in semi-arid regions. Tested with two different independent datasets, one constituted by 20% of the reference information, and another including observed yields for year 2018 (a one-year-left analysis), maize yield predictions were promising for year 2017 (RMSE = 362 kg/ha), but showed higher error for 2018 (RMSE = 707 kg/ha). That is, the fitted model may not capture accurately year to year variabilities in predicted maize yield. In this analysis, predictions were limited to field samples (~600 fields) across the country of Mali. It would be valuable in the future to predict maize yield for each pixel of the new developed crop type map. That will lead to a detailed spatial analysis of maize yield, allowing identification of low yielding regions for targeted interventions which could improve food security. Keywords: Agricultural identification of croplands as crop growth condition metrics are extracted. WASC30 allowed us to derive detailed statistics on cultivated areas in the Sahel, at country and agroclimatic scales. Intensive agricultural zones were highlighted as well. The second dataset, mapping crop types for the country of Mali, is meant to separate signals of different crop types for improved crop yield estimation. The crop type map was used to derive detailed agricultural statistics (e.g. acreage by crop types, spatial distribution) at finer administrative scales than has previously been possible. The crop fraction information by crop type extracted from the map, gives additional details on farmers preferences by regions, and the natural adaptability of different crop types. The final analysis of this dissertation explores the use of ensemble machine learning techniques to predict maize yield in Mali (Chapter 5). Climate data (precipitation and temperature), and vegetation indices (Normalized Difference Vegetation Index, NDVI, the Enhanced Vegetation Index, EVI, and the Normalized Difference Water Index, NDWI) are used as predictors, while actual yields collected in 2017 by the Malian Ministry of Agriculture are the reference data. Random forest presented better predictive performance as compared to boosted regression trees (BRT). Results showed that climate variables have more predictive power for maize yield compared to vegetation indices. Among vegetation indices, the NDWI appeared to be the most influential predictor, maybe because of water requirement of maize and the sensitivity of this index to water in semi-arid regions. Tested with two different independent datasets, one constituted by 20% of the reference information, and another including observed yields for year 2018 (a one-year-left analysis), maize yield predictions were promising for year 2017 (RMSE = 362 kg/ha), but showed higher error for 2018 (RMSE = 707 kg/ha). That is, the fitted model may not capture accurately year to year variabilities in predicted maize yield. In this analysis, predictions were limited to field samples (~600 fields) across the country of Mali. It would be valuable in the future to predict maize yield for each pixel of the new developed crop type map. That will lead to a detailed spatial analysis of maize yield, allowing identification of low yielding regions for targeted interventions which could improve food security.

Format

application/pdf

Number of Pages

161

Publisher

South Dakota State University

Rights

Copyright © 2021 the Author

Share

COinS