Home / Current Issue / Paper 1722534
Multi-Source Remote Sensing and Explainable Machine Learning for Cassava Yield Monitoring Across West Africa
Subject area: Science,Engineering and Technology · Area of research: Machine Learning, Precision Agriculture
Abstract
Accurate monitoring of cassava (Manihot esculenta Crantz) yields in West Africa continues to be considerable issue faced by those working in food security also agricultural policy planners. Conventional methods of estimating yields suffer from poor scalability, high operational costs, and significant temporal delays that limit their practical usefulness. This 25-year investigation (2000–2024) presents an explainable Remote Sensing-Ensemble Machine Learning (RS-EML) framework that brings together multi-source geospatial observations—FAOSTAT production records, MODIS vegetation indices (NDVI, EVI), CHIRPS rainfall measurements, and ERA5 reanalysis temperature fields—into a unified prediction system covering ten West African nations. The proposed pipeline has three stages that operate. The first one is that a multi-source data must be harmonized and then standardized even though the data can have different time and space scales. Second selection occurs for the best ensemble configuration using three algorithms (Merit-Based Ranking, Pareto-Optimal Frontier Selection and an Adaptive Threshold Optimization) which have multi-objective optimization at their core; and third, applying Shapley Additive exPlanations (SHAP) to quantify each input feature's contribution to yield forecasts. Validation was done in three different agroecological areas. These were Guinea Savanna (R² counted as 0.8506), Transitional/Humid zone (R² is 0.8394) and Sudano-Sahelian region (R² ended at 0.8123).An independent ground-truthing using information from a 600 cassava farms showed solid ability to predict yield. Random Forest and XGBoost algorithms were put together for stacking ensemble, where the metrics were R² 0.8456 and RMSE estimated 1662.70 kg ha⁻¹ while MAPE was 18.2 percent against farmer-reported yields. SHAP result found vegetation changes (38 percent of total) and rainfall (about 35 percent) are main yield causes, temperature has 18 percent. The system does not need ground-truthing data to run. This framework makes crop monitoring tools transparent, repeatable and scalable for food security checking in the Sub-Saharan Africa.
Keywords
cassava yield prediction; ensemble machine learning; explainable ai; remote sensing; west africa
How to cite this paper
@article{1722534,
author = {Edith Ugochi. Omede, Ufuoma Kazeem Okpeki, Chika Lilian Onyagu, Blessing Ewoma Oyovwe, Veronica Uchechukwu Ikenga},
title = {Multi-Source Remote Sensing and Explainable Machine Learning for Cassava Yield Monitoring Across West Africa},
journal = {Iconic Research And Engineering Journals},
year = {2026},
volume = {10},
number = {2},
pages = {3410-3422},
issn = {2456-8880},
url = {https://www.irejournals.com/formatedpaper/1722534.pdf},
abstract = {Accurate monitoring of cassava (Manihot esculenta Crantz) yields in West Africa continues to be considerable issue faced by those working in food security also agricultural policy planners. Conventional methods of estimating yields suffer from poor scalability, high operational costs, and significant temporal delays that limit their practical usefulness. This 25-year investigation (2000–2024) presents an explainable Remote Sensing-Ensemble Machine Learning (RS-EML) framework that brings together multi-source geospatial observations—FAOSTAT production records, MODIS vegetation indices (NDVI, EVI), CHIRPS rainfall measurements, and ERA5 reanalysis temperature fields—into a unified prediction system covering ten West African nations. The proposed pipeline has three stages that operate. The first one is that a multi-source data must be harmonized and then standardized even though the data can have different time and space scales. Second selection occurs for the best ensemble configuration using three algorithms (Merit-Based Ranking, Pareto-Optimal Frontier Selection and an Adaptive Threshold Optimization) which have multi-objective optimization at their core; and third, applying Shapley Additive exPlanations (SHAP) to quantify each input feature's contribution to yield forecasts. Validation was done in three different agroecological areas. These were Guinea Savanna (R² counted as 0.8506), Transitional/Humid zone (R² is 0.8394) and Sudano-Sahelian region (R² ended at 0.8123).An independent ground-truthing using information from a 600 cassava farms showed solid ability to predict yield. Random Forest and XGBoost algorithms were put together for stacking ensemble, where the metrics were R² 0.8456 and RMSE estimated 1662.70 kg ha⁻¹ while MAPE was 18.2 percent against farmer-reported yields. SHAP result found vegetation changes (38 percent of total) and rainfall (about 35 percent) are main yield causes, temperature has 18 percent. The system does not need ground-truthing data to run. This framework makes crop monitoring tools transparent, repeatable and scalable for food security checking in the Sub-Saharan Africa.},
keywords = {cassava yield prediction; ensemble machine learning; explainable ai; remote sensing; west africa},
month = {August},
}