The application of deep learning for remote sensing of soil organic carbon stocks distribution in South Africa = Ukusetshenziswa kokufunda okujulile kokuzwa kude kokusatshalaliswa kwesitokwe sekhabhoni ephilayo enhlabathini eNingizimu Afrikha.

Abstract

Soil organic carbon (SOC) is a vital measure for ecosystem health and offers opportunities to understand carbon fluxes and associated implications. However, unprecedented anthropogenic disturbances have significantly altered SOC distribution across the globe, leading to considerable carbon losses. In addition, reliable SOC estimates, particularly over large spatial extents remain a major challenge due to among others limited sample points, quality of simulation data and suitable algorithms. Remote sensing (RS) approaches have emerged as a suitable alternative to field and laboratory SOC determination, especially at large spatial extent. Nevertheless, reliable determination of SOC distribution using RS data requires robust analytical approaches. Compared to linear and classical machine learning (ML) models, deep learning (DL) models offer a considerable improvement in data analysis due to their ability to extract more representative features and identify complex spatial patterns associated with big data. Hence, advancements in remote sensing, proliferation of big data, and deep learning architecture offer great potential for large-scale SOC mapping. However, there is paucity in literature on the application of DL-based remote sensing approaches for SOC prediction. To this end, this study is aimed at exploring DL-based approaches for the remote sensing of SOC stocks distribution across South Africa. The first objective sought to provide a synopsis of the use of traditional neural network (TNN) and DL-based remote sensing of SOC with emphasis on basic concepts, differences, similarities and limitations, while the second objective provided an in-depth review of the history, utility, challenges, and prospects of DL-based remote sensing approaches for mapping SOC. A quantitative evaluation between the use of TNN and DL frameworks was also conducted. Findings show that majority of published literature were conducted in the Northern Hemisphere while Africa have only four publications. Results also reveal that most studies adopted hyperspectral data, particularly spectrometers as compared to multispectral data. In comparison to DL (10%), TNN (90%) models were more commonly utilized in the literature; yet, DL models produced higher median accuracy (93%) than TNN (85%) models. The review concludes by highlighting future opportunities for retrieving SOC from remotely sensed data using DL frameworks. The third objective compared the accuracy of DL—deep neural network (DNN) model and a TNN—artificial neural network (ANN), as well as other popular classical ML models that include random forest (RF) and support vector machine (SVM), for national scale SOC mapping using Sentinel-3 data. With a root mean square error (RMSE) of 10.35 t/ha, the DNN model produced the best results, followed by RF (11.2 t/ha), ANN (11.6 t/ha), and SVM (13.6 t/ha). The DNN's analytical abilities, combined with its capacity to handle large amounts of data is a key advantage over other classical ML models. Having established the superiority of DL models over TNN and other classical models, the fourth objective focused on investigating SOC stocks distribution across South Africa’s major land uses, using Deep Neural Networks (DNN) and Sentinel-3 satellite data. Findings show that grasslands contributed the most to overall SOC stocks (31.36 %), while urban vegetation contributed the least (0.04%). Results also show that commercial (46.06 t/h) and natural (44.34 t/h) forests had better carbon sequestration capacity than other classes. These findings provide an important guideline for managing SOC stocks in South Africa, useful in climate change mitigation by promoting sustainable land-use practices. The fifth objective sought to determine the distribution of SOC within South Africa’s major biomes using remotely sensed-topo-climatic data and Concrete Autoencoder-Deep Neural Networks (CAE-DNN). Findings show that the CAE-DNN model (built from 26 selected variables) had the best accuracy of the DNNs examined, with an RMSE of 7.91 t/h. Soil organic carbon stock was also shown to be related to biome coverage, with the grassland (32.38%) and savanna (31.28%) biomes contributing the most to the overall SOC pool in South Africa. forests (44.12 t/h) and the Indian ocean coastal belt (43.05 t/h) biomes, despite having smaller footprints, have the highest SOC sequestration capacity. To increase SOC storage, it is recommended that degraded biomes be restored; however, a balance must be maintained between carbon sequestration capability, biodiversity health, and adequate provision of ecosystem services. The sixth objective sought to project the present SOC stocks in South Africa into the future (i.e. 2050). Soil organic carbon variations generated by projected climate change and land cover were mapped and analysed using a digital soil mapping (DSM) technique combined with space-for-time substitution (SFTS) procedures over South Africa through 2050. The potential SOC stocks variations across South Africa's major land uses were also assessed from current (2021) to future (2050). The first part of the study uses a Deep Neural Network (DNN) to estimate current SOC content (2021), while the second phase uses an average of five WorldClim General Circulation Models to project SOC to the future (2050) under four Shared Socio-economic Pathways (SSPs). Results show a general decline in projected future SOC stocks by 2050, ranging from 4.97 to 5.38 Pg, compared to estimated current stocks of 5.64 Pg. The findings are critical for government and policymakers in assessing the efficacy of current management systems in South Africa. Overall, this study provides a cost-effective framework for national scale mapping of SOC stocks, which is the largest terrestrial carbon pool using advanced DL-based remote sensing approach. These findings are valuable for designing appropriate management strategies to promote carbon uptake, soil quality, and measuring terrestrial ecosystem responses and feedbacks to climate change. This study is also the first DL-based remote sensing of SOC stocks distribution in South Africa.

Iqoqa

Ikhabhoni yomhlabathi engaguquliwe, phecelezi i-soil organic carbon (SOC) ibaluleke kakhulu ebudlelwaneni bezinto eziphilayo nendawo eziphila kuyona futhi isiza ukuqonda kabanzi ubudlelwane bamakhabhoni nezinto eziyidingayo, nokuthi lokho kunayiphi imiphumela. Nokho, ukungcola okungajwayelekile okudalwa ngabantu kubonakala kuza noshintsho olukhulu ekuhanjisweni kwe-SOC kuwona wonke umhlaba, okuholela ekutheni kube nokulahleka kwekhabhoni eningi. Ukwengeza, ukuhlawumbisela okuthembekile nge-SOC, ikakhulukazi okuthinta umthamo omkhulu kubonakala kuqhubeka nokuba yinkinga ngenxa yesizathu sokuthi kukhona izibonelo zayo ezimbalwa, kanti nezinga elihle liyagqoza, nendlela eyiyo okumele ilandelwe ukukala. Indlela yokuhlola buqamama, phecelezi i-Remote sensing (RS) iyona ebonakala njengendlela engalandelwa ukuhlola iSOC ensimini noma egunjini lokuhlolela, ikakhulukazi uma kuthinta indawo enkulu. Nokho, ukuhlonza indlela ethembekile yokusatshalaliswa kwe-SOC kusetshenziswa imininingwane ye-RS kudinga izindlela eziseqophelweni eliphezulu zokuhlaziya. Ukuqhathanisa nendleya ye-linear ne-classical machine learning (ML), Indlela ye-deep learning (DL) yona ibonakala iletha ubungcono obukhulu ekuhlaziyeni imininingo ngenxa yokukwazi ukuhlonza izinto ezidingekalayo nokuveza izinto eziyinkinga kuleyo ndawo enemininingo eminingi. Ukuthuthuka kwendlela yokuhlola buqamama, ukuhlaziywa kwemininingo eminingi, nokufunda ngobuciko bokwakha indlela yokwenza, konke kuza namathuba amahle okuphaka uhlelo ngobuningi balo be-SOC. Nakuba kunjalo, kubonakala kunolwazi oluyingcosana emibhalweni yocwaningo mayelana nokusebenza kwezindlela zokuhlola buqamama ngesihlawumbiselo se-SOC. Kuze kube manje, lolu cwaningo beluhlose ukucubungula izindlela zokuhlaziya buqamama ze-DL ngokusatshalaliswa kwe-SOC eNingizimu-Afrikha yonkana. Inhloso yokuqala bekuwukunikeza ulwazi olufushane mayelana nendlela ejwayelekile yokusabalalisa, phecelezi i- traditional neural network (TNN) nendlela yokuhlaziya buqamama ye-DL ye-SOC ngokugcizelela ukubaluleka kwezinto ezijwayelekile, nomahluko. Okufanayo nezingqinamba, kanti inhloso yesibili inikeze ukucubungula okujulile komlando, ukusetshenziswa kwento, izinselelo, kanye namathuba okusebenza kwendlela yokuhlaziya buqamama ye-DL ekusabalaliseni i-SOC. Ukuhlaziya ngokwekhwalithethivu ekusebenziseni i-TNN ne-DL nakho kwenziwa. Imiphumela iveza ukuthi imibhalo eminingi yocwaningo yashicilelwa eNyakatho nomhlaba kanti lapha e-Afrika khona kwashicilelwa emine nje kuphela. Imiphumela iphinde iveze ukuthi ucwaningo oluningi lulandela indlela ebheka imininingo ehlanngene, ikakhulukazi ukubheka imibala eminingi yokukhanya, uma kuqhathaniswa nengxubevange yemininingo.

Ukuqhathanisa nezindlela ze-DL (10%) ne-TNN (90%) yizona ezithandwa kakhulu emibhalweni yocwaningo, kodwa, indlela ye-DL ikhiqize okuneqinso (93%) kunendlela ye-TNN (85%). Ukucubungula kuphetha ngokugqamisa amathuba azayo okuthola ulwazi lwe-SOC kusetshenziswa indlela yokuhlola buqamama nge-DL.

Inhloso yesithathu yona yayiqhathanisa ukuthembeka kwe-DL-deep neural network (DNN) kanye ne-TNN—artificial neural network (ANN), kanye nenye indlela endala edumile ye-ML efaka phakathi i-random forest (RF) kanye ne-support vector machine (SVM), ukhubheka isikalo sikazwelonke se-SOC kusetshenziswa imininingo ye-Sentinel-3. Nge-root mean square error (RMSE) ye-10.35 t/ha, indlela ye-DNN yakhiqiza imiphumela emihle kakhulu, ilandelwa yi-RF (11.2 t/ha), i-ANN (11.6 t/ha), ne-SVM (13.6 t/ha). Amandla okuhlaziya e-DNN, ehlanganiswe namandla okukwazi ukubhekana nemininingo eminingi kakhulu yikhona okwenza ibaluleke ukwedlula indlela ye-ML. Emva kokuhlonza amandla e-DL phezu kwawe-TNN kanye nezinye izindlela ezaziwayo, inhloso yesine yona ibigxile ekuhloleni ukusabalaliswa kwe-SOC ekusetshenzisweni komhlaba ngobuningi bawo lapha eNingizimu-Afrika, kusetshenziswa i-Deep Neural Networks (DNN) ne-Sentinel-3 satellite data. Imiphumela iveza ukuthi indawo enotshani iyona eletha izibalo eziphezulu ze-SOC sekuhlangene yonke into (31.36 %), kanti izimilo zasedolobheni zona zaletha izibalo ezincane (0.04%). Imiphumela iphinde ikhombise ukuthi amahlathi atshalelwe ukudayisa (46.06 t/h) kanye nawemvelo (44.34 t/h) ayekhombisa ukuba nekhabhoni enhle uma kuqhathaniswa namanye. Le miphumela ingumhlahlandlela obalulekile mayelana nokugcinwa kwe-SOC lapha eNingizimu-Afrikha, okuyinto ebaluleke kakhulu ukubhekana nokuguquguquka kwesimo sezulu ngokugqugquzela izindlela eziyizo zokusebenzisa umhlaba.

Inhloso yesihlanu yona ibihlose ukuthola ulwazi ngokusatshalaliswa kwe-SOC ezinhlakeni ezahlukene zomphakathi lapha eNingizimu-Afrikha kusetshenziswa indlela yokuhlaziya buqama kanye ne-Concrete Autoencoder-Deep Neural Networks (CAE-DNN). Imiphumela iveza ukuthi indlela ye-CAE-DNN (eyakhiwe ngezinto ezikhethiweyo ezingama-26) yakhombisa ubuqiniso obukhulu be-DNNs, eyayihloliwe, nge-RMSE of 7.91 t/h. I-SOC yaphinde yakhombisa ukuba nobudlelwane nokusabalala kohlaza olumilile, kukhona indawo enotshani (32.38%) nendawo engenazihlahla yohlaza lotshani (31.28%) okuyiyona ephethe umhlaba omningi uma usuhlanganisiwe we-SOC lapha eNingizimu-Afrikha, amahlathi (44.12 t/h) kanye nomhlaba onohlaza lotshani ukugudla ulwandle i Indian Ocean (43.05 t/h), ngale okubonakala kancane, kodwa inokugcwala okuningi kwe-SOC. Ukukhulisa indawo yokugcina i-SOC, kuphakanyiswa ukuba indawo yohlaza lotshani ebisagciniwe iphinde isetshenziswe; nokho-ke, kumele kube khona ukulinganisa okuyikhona phakathi kokuthathwa kwekhabhoni, impilo yokubhekwayo kanye nokunakekelwa kwezinto eziphilayo nendawo eziphila kuyona. Inhloso yesithupha yayihlose ukuvikela ukukhiqizwa kwe-SOC lapha eNingizimu-Afrikha ngisho eminyakeni eminingi ezayo (okuwunyaka wezi-2050). Izinhlobo ezahlukene ze-SOC ezikhiqizwe ngokuguquguquka okwahlukene kwezimo zezulu nomhlaba kwakahlwa futhi kwahlaziywa kusetshenziswa isu le-digital soil mapping (DSM) lihlanganiswe nezindlela ze-space-for-time substitution (SFTS) eNingizimu-Afrikha kuze kube unyaka wezi-2050. Abasebenzisi bakusasa bomhlaba okhiqiza izinhlobo ezahlukene ze-SOC nabo bahlolwa kusukela kwabamanje (2021) kuye kwabangomuso (2050). Ingxenye yokuqala yalolu cwaningo isebenzisa i-Deep Neural Network (DNN) ukugagula isimo ngqo sengqikithi ye-SOC (2021), kanti ingxenye yesibili yona isebenzisa okungenani izindlela ezinhlanu zokusabalalisa nge-WorldClim General Circulation kumaphrojekthi e-SOC nangeminyaka ezayo (2050) ngaphansi kwe-Shared Socio-economic Pathways (SSPs) emine. Imiphumela iveza ukwehla kwe-SOC ngokwezibalo zangomuso ngeminyaka yezi-2050, kusukela ku-4.97 kuya ku-5.38 Pg, uma kuqhathaniswa ne-SOC ekhona manje engu-5.64 Pg. Le miphumela ibaluleke kakhulu kuhulumeni nakulaba abakhipha izinqubomgomo ukubheka ukusebenza kahle kwezinhlaka zokuphatha eNingizimu-Afrikha. Sekukonke, lolu cwaningo lunikeza indlela yokucabanga nokwenza engabizi neyimpumelelo yokusabalalisa i-SOC, okuyiyona enkulu kakhulu ekukhiqizweni kwekhaboni kusetshenziswa indlela yokuhlola buqamama ye-DL. Lolu cwaningo lubalulekile ukuqhamuka nezindlela ezintsha zokuphatha ukuze kugqugquzeleke ukukhiqizwa kwekhabhoni, umhlaba ovundile, nokukala izinto eziphilayo nendawo eziphila kuyona kanye nokunikeza impendulo mayelana nokuguquguquka kwesimo sezulu. Lolu cwaningo lungolokuqala ngqa lokubheka indlela yokuhlaziya buqamama nge-DL lapha eNingizimu Afrikha.