Untangling intercropping in heterogeneous smallholder maize-cassava farming systems with remote sensing

Akinyemi, Felicia O.; Rufin, Philippe; Ibrahim, Esther Shupel; Hostert, Patrick; Ogunsumi, Lucia O.; Egbetokun, Olugbenga; Ifejika Speranza, Chinwe (2023). Untangling intercropping in heterogeneous smallholder maize-cassava farming systems with remote sensing (EarthArXiv). 10.31223/X57T2M

Preview

Text akinyemi-prediction_intercropping-cropping-systems_rs-ml_preprint.pdf - Published Version Available under License Creative Commons: Attribution-Noncommercial (CC-BY-NC). Download (3MB) | Preview

Earth observation approaches for large-scale crop monocultures are often not transferable to heterogeneous smallholder systems. Key challenges in this regard are intercropping, high intra-field crop type variability, wide sowing windows, presence of non-crop vegetation and small but variable field sizes. Currently, studies on smallholder agriculture mainly focus on specific crops and seldom account for crop mixtures or multiple growing cycles. Moreover, our knowledge about ongoing processes of farm consolidation and effects on intercropping remains limited due to the absence of spatially detailed information on field size. We mapped monocropping and maize-cassava intercropping in 2022/2023 and the relationship with field sizes. We combined Sentinel-1 radar and optical Sentinel-2 time series to classify farming systems across two growing cycles in the Guinea Savannah of southwest Nigeria. We tested spectral-temporal features at monthly and bimonthly intervals for the growing season and off-season. We used deep transfer learning to fine-tune a pre-trained convolutional neural network designed for crop field delineation. Using very high resolution imagery (0.6 m) for a regularly distributed sample across the study region (n=2,333), mean overall accuracy based on k-fold cross-validation was 0.79 (+/-0.02%), whereas User and Producer accuracies were above 0.70 for most classes. Sentinel-1 alone underperformed, while models using only Sentinel-2 had higher overall accuracies but suffered from cloud-induced data gaps. Field size estimation revealed a high spatial agreement with mean intersection over union scores of up to 0.73 in site-level field size estimation. Small and medium-sized fields were dominant. Monocropping was positively related to field sizes as larger monocropping fields of early-planted cassava, late-planted maize, yam and rice clustered in the North of our study region. In contrast, smaller intercropped fields of maize-cassava mainly occurred in fragmented agricultural landscapes with ample natural vegetation. Our approach demonstrates the potential of integrating radar and optical time series in cloud-prone regions for mapping crop mixtures in complex forest-agricultural mosaic landscapes during multiple growing cycles. Our study provides a valuable workflow for producing timely information for the quantification of crop production in heterogeneous smallholder farming systems.

Interest & Impact

Downloads

13 since deposited on 03 May 2024
13 in the past 12 months

Citations

5 Citations in Web of Science ®
6 Citations in Scopus

Search

in Google Scholar™

Services

Actions (login required)

Edit item

Actions (login required)

Edit item