Sen Yang, Quan Feng, Xueze Gao, Wanxia Yang, Guanping Wang
{"title":"Improving Mapping Accuracy of Smallholder Potato Planting Areas by Embedding Prior Knowledge into a Novel Multi-temporal Deep Learning Network","authors":"Sen Yang, Quan Feng, Xueze Gao, Wanxia Yang, Guanping Wang","doi":"10.1007/s11540-024-09769-2","DOIUrl":null,"url":null,"abstract":"<p>Accurate and timely acquisition of potato spatial distribution is crucial for growth monitoring and yield forecasting. Currently, prior knowledge-based methods are very simple and efficient without collecting reference data, but their mapping accuracy in complex cropping planting systems is unsatisfactory. Deep learning approaches have the ability to automatically learn multilevel spatial and spectral features. However, these approaches still face particular challenges in improving potato mapping accuracy due to the limitations of adaptive features and the scarcity of ground samples. This study proposed a potato mapping method integrating a multi-temporal deep learning network and prior knowledge to overcome the shortcomings of the two methods. Specifically, a novel deep learning network, spectral-spatial–temporal ensemble network (SSTEN), was developed for smallholder potato area mapping by embedding unique prior knowledge. To obtain multi-year potato mapping results, we proposed a concise and efficient temporal transfer framework that combines sample generation, SSTEN transfer learning, and agriculture statistics to produce highly accurate potato maps for sample-free years. Independent ground validation data from 2021 to 2022 suggested that the SSTEN achieved an overall accuracy (OA), F1 and Kappa of 91.65%, 92.67% and 0.82, respectively, and its average overall accuracy was superior to other methods. Potato planting areas obtained by SSTEN were highly consistent with the corresponding agricultural statistical area (<i>R</i><sup>2</sup> > 0.87). The results showed that incorporating prior knowledge into SSTEN could improve the accuracy of potato mapping. We also investigated the potential of the proposed temporal transfer method for potato mapping. Our transfer method yielded a high OA of 86.46% and an area error (AE) of 7.94%. The study potentially provides technical references for smallholder potato mapping in similar agricultural regions worldwide.</p>","PeriodicalId":20378,"journal":{"name":"Potato Research","volume":"153 1","pages":""},"PeriodicalIF":2.3000,"publicationDate":"2024-07-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Potato Research","FirstCategoryId":"97","ListUrlMain":"https://doi.org/10.1007/s11540-024-09769-2","RegionNum":3,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"AGRONOMY","Score":null,"Total":0}
引用次数: 0
Abstract
Accurate and timely acquisition of potato spatial distribution is crucial for growth monitoring and yield forecasting. Currently, prior knowledge-based methods are very simple and efficient without collecting reference data, but their mapping accuracy in complex cropping planting systems is unsatisfactory. Deep learning approaches have the ability to automatically learn multilevel spatial and spectral features. However, these approaches still face particular challenges in improving potato mapping accuracy due to the limitations of adaptive features and the scarcity of ground samples. This study proposed a potato mapping method integrating a multi-temporal deep learning network and prior knowledge to overcome the shortcomings of the two methods. Specifically, a novel deep learning network, spectral-spatial–temporal ensemble network (SSTEN), was developed for smallholder potato area mapping by embedding unique prior knowledge. To obtain multi-year potato mapping results, we proposed a concise and efficient temporal transfer framework that combines sample generation, SSTEN transfer learning, and agriculture statistics to produce highly accurate potato maps for sample-free years. Independent ground validation data from 2021 to 2022 suggested that the SSTEN achieved an overall accuracy (OA), F1 and Kappa of 91.65%, 92.67% and 0.82, respectively, and its average overall accuracy was superior to other methods. Potato planting areas obtained by SSTEN were highly consistent with the corresponding agricultural statistical area (R2 > 0.87). The results showed that incorporating prior knowledge into SSTEN could improve the accuracy of potato mapping. We also investigated the potential of the proposed temporal transfer method for potato mapping. Our transfer method yielded a high OA of 86.46% and an area error (AE) of 7.94%. The study potentially provides technical references for smallholder potato mapping in similar agricultural regions worldwide.
期刊介绍:
Potato Research, the journal of the European Association for Potato Research (EAPR), promotes the exchange of information on all aspects of this fast-evolving global industry. It offers the latest developments in innovative research to scientists active in potato research. The journal includes authoritative coverage of new scientific developments, publishing original research and review papers on such topics as:
Molecular sciences;
Breeding;
Physiology;
Pathology;
Nematology;
Virology;
Agronomy;
Engineering and Utilization.